CLI reference

This page is generated from the qsmxt binary’s argument definitions, so it lists exactly the commands and options the binary accepts. qsmxt <command> --help prints the same text. For what the options do, see Commands, Algorithms and Standalone tools.

Commands

• qsmxt run
• qsmxt init
• qsmxt validate
• qsmxt dicom-convert
• qsmxt slurm
• qsmxt aggregate
• qsmxt mask
• qsmxt combine
• qsmxt unwrap
• qsmxt fieldmap
• qsmxt bgremove
• qsmxt invert
• qsmxt separate
• qsmxt qsmart
• qsmxt swi
• qsmxt segment
• qsmxt r2star
• qsmxt t2star
• qsmxt r2
• qsmxt r2prime
• qsmxt homogeneity
• qsmxt resample
• qsmxt quality-map
• qsmxt example
• qsmxt tui
• qsmxt update
• qsmxt news

qsmxt run

Run the full QSM pipeline on a BIDS dataset

qsmxt run [OPTIONS] <BIDS_DIR> [OUTPUT_DIR]

Argument	Description
<BIDS_DIR>	Input BIDS directory. Required.
<OUTPUT_DIR>	Output directory (defaults to bids_dir; outputs go into <dir>/derivatives/qsmxt/)
--config <CONFIG>	Pipeline configuration file (TOML)
--include <INCLUDE>	Include only runs matching these glob patterns (e.g. “sub-1*” “*ses-pre*”)
--exclude <EXCLUDE>	Exclude runs matching these glob patterns (e.g. “*mygrea*”)
--num-echoes <NUM_ECHOES>	Limit number of echoes to process.
--pipeline <PIPELINE>	Start from a pipeline: qsmxt-defaults, or an original toolbox’s (medi-toolbox, sti-suite-ilsqr, sti-suite-star, qsmart, heidi, chi-sep-ilsqr, chi-sep-medi, chi-sepnet, decompose). The stages it covers (masking, field mapping, background removal, inversion, referencing, χ-separation) go back to QSMxT’s defaults and then take the toolbox’s settings; the other flags given then change any setting. Values: qsmxt-defaults, medi-toolbox, sti-suite-ilsqr, sti-suite-star, qsmart, heidi, chi-sep-ilsqr, chi-sep-medi, chi-sepnet, decompose.
--qsm-algorithm <QSM_ALGORITHM>	QSM dipole inversion algorithm (default: ilsqr) Values: rts, tv, tkd, tsvd, tgv, tikhonov, nltv, medi, tfi, ilsqr, star, lsqr, heidi, qsmart, ndi, fansi, fansi-tgv, l1qsm, whqsm, hdqsm, amp-pe, xqsm, qsmnet, qsmnet-plus, qsmnetme, autoqsm, qsmgan, ir2qsm, lpcnn, modl-qsm, nextqsm, iqsm, iqsm-plus.
--heidi	HEIDI streaking reduction after the inversion: keep the map’s well-conditioned k-space and refill the dipole cone under a homogeneity prior (--heidi-* sets it up). With LSQR this is HEIDI as published (--qsm-algorithm heidi); with two-pass it runs once, on the combined map.
--orientation-group <PATTERN>	Group runs into multi-orientation sets and reconstruct each with COSMOS or STI. Takes an entity name (acq, run, rec, inv), a glob over the run key (*acq-dir*), re: and a regex with one capture group, or runs: and the groups’ run keys (, between runs, ; between groups, each group’s reference first). Orientations are co-registered if they need it.
--orientation-reference <PATTERN>	The orientation each group is registered onto, and whose grid and segmentation the group map uses: a glob over the run key (*acq-Neutral*) or a word in it. Must match one run per group. Default: the first run in name order.
--no-orientation-registration	Do not co-register the orientations in a group — use this when they were already registered externally and a second interpolation is not wanted. A group that still needs registration is then refused instead of reconstructed.
--multi-orientation-algorithm <MULTI_ORIENTATION_ALGORITHM>	Which reconstruction to run over each orientation group. Values: cosmos, sti.
--multi-orientation-lambda <MULTI_ORIENTATION_LAMBDA>	Regularization for the multi-orientation reconstruction (COSMOS k-space L2, or STI ridge)
--multi-orientation-force	Reconstruct an orientation group even when its B0 directions look degenerate.
--cosmos-magnitude-weighting	COSMOS: weight each orientation by its combined magnitude (conjugate-gradient solve)
--cosmos-tol <COSMOS_TOL>	Magnitude-weighted COSMOS: conjugate-gradient tolerance (needs –cosmos-magnitude-weighting)
--cosmos-max-iter <COSMOS_MAX_ITER>	Magnitude-weighted COSMOS: conjugate-gradient iteration cap (needs –cosmos-magnitude-weighting)
--unwrapping-algorithm <UNWRAPPING_ALGORITHM>	Unwrapping algorithm (default: romeo) Values: romeo, laplacian, prelude, medi-region-growing.
--bf-algorithm <BF_ALGORITHM>	Background field removal algorithm (default: vsharp) Values: vsharp, pdf, lbv, ismv, sharp, resharp, harperella, iharperella, bfrnet, iqfm.
--bf-erode-voxels <N>	Erode the brain mask by N voxels after field mapping, before background removal; the field map is not masked again (default: 0, none)
--bf-polish-mask	After background removal, keep the mask’s largest connected component with its holes filled and zero the local field outside it, as STI Suite’s QSM_GUI does (PolishMask)
--no-bf-polish-mask	Do not polish the background-removal mask (the default), after a --pipeline that does.
--masking-input <MASKING_INPUT>	Masking input image (overrides the input of the configured mask sections; combine with –mask-preset, e.g. –mask-preset robust-threshold –masking-input magnitude) Values: magnitude-first, magnitude, magnitude-last, phase-quality.
--phase-offset-removal <PHASE_OFFSET_REMOVAL>	Enable phase offset removal for multi-echo data (default: true)
--phase-offset-sigma-mm <PHASE_OFFSET_SIGMA_MM>	Phase offset removal smoothing sigma (3 values, in mm; default: 7 7 7, the ROMEO command line’s)
--coil-combination-sigma <COIL_COMBINATION_SIGMA>	MCPC-3D-S coil combination: smoothing sigma for the per-coil phase offsets (3 values, in voxels; default 10 10 5, as in MriResearchTools). Only used for uncombined multi-coil runs.
--bipolar-correction	Enable bipolar gradient correction (requires >= 3 echoes)
--b0-estimation <B0_ESTIMATION>	B0 field estimation method (default: weighted-avg) Values: weighted-avg, linear-fit, robust-fit, qsmart-echofit, complex-fit.
--qsmart-fit-threshold-mode <QSMART_FIT_THRESHOLD_MODE>	--b0-estimation qsmart-echofit: how the fit residual at which a voxel is left out of QSMART’s reconstruction is set: a percentile of the blurred residual over the volume (--qsmart-fit-percentile), or a fixed residual (--qsmart-fit-threshold; default, QSMART v1.0’s 40). Without it, giving one of those two selects its mode. Values: percentile, absolute.
--qsmart-fit-percentile <QSMART_FIT_PERCENTILE>	--b0-estimation qsmart-echofit: the percentile of the blurred fit residual, in (0, 100], above which a voxel is left out of QSMART’s reconstruction (default 99)
--qsmart-fit-threshold <QSMART_FIT_THRESHOLD>	--b0-estimation qsmart-echofit: a fixed blurred fit residual above which a voxel is left out of QSMART’s reconstruction, instead of a percentile (40 when the mode is absolute and none is given)
--b0-weight-type <B0_WEIGHT_TYPE>	B0 weighted averaging weight type. assumed-decay gives each echo’s phase a fixed weight TE·exp(−TE/T2*) computed from the echo times and an assumed T2* (--b0-weight-assumed-t2star), the same in every voxel: no T2* map, no magnitude (Wang et al. 2022, Nat Neurosci 25:818). phase-snr, which uses the measured magnitude in each voxel, is generally preferable when magnitude is available. Applies to Laplacian unwrapping as well as ROMEO. Values: phase-snr, phase-var, average, tes, mag, assumed-decay.
--b0-weight-assumed-t2star <MS>	T2* in ms assumed by --b0-weight-type assumed-decay to compute its fixed echo weights (default 40; no T2* map is used)
--laplacian-solver <LAPLACIAN_SOLVER>	Poisson solver for --unwrapping-algorithm laplacian: dct (unweighted least squares, Neumann/DCT solve; default), fft (Schofield & Zhu sin/cos Laplacian, FFT solve on a zero-padded volume) or mirror (the same on the volume mirrored along each axis, as QSMART unwraps) Values: dct, fft, mirror.
--laplacian-fft-pad <VOXELS>	Zero-padding in voxels per side for --laplacian-solver fft: one value for every axis, or three (x y z). Default 64.
--laplacian-unmasked	--unwrapping-algorithm laplacian: unwrap the raw phase over the whole volume, instead of the phase set to zero outside the brain mask, as QSMART does. With --b0-estimation qsmart-echofit the field map is then kept over the whole volume.
--no-laplacian-unmasked	--unwrapping-algorithm laplacian: unwrap the phase set to zero outside the brain mask (the default), after a --pipeline that unwraps the whole volume.
--bet-fractional-intensity <BET_FRACTIONAL_INTENSITY>	BET fractional intensity (0.0-1.0; smaller = larger brain). Applies to every BET step in the masking recipe, overriding a bet:<f> in a –mask spec.
--bet-smoothness <BET_SMOOTHNESS>	BET surface smoothness factor. Applies to every BET step in the masking recipe.
--bet-gradient-threshold <BET_GRADIENT_THRESHOLD>	BET gradient threshold (-1 to 1). Applies to every BET step in the masking recipe.
--bet-iterations <BET_ITERATIONS>	BET surface evolution iterations. Applies to every BET step in the masking recipe.
--bet-subdivisions <BET_SUBDIVISIONS>	BET icosphere subdivision level. Applies to every BET step in the masking recipe.
--bet-voxel-scale <FACTOR>	Factor applied to the voxel sizes before BET runs. BET’s surface model is tuned to a human brain in millimetres, so preclinical data has to be presented at human scale: 10 is the usual setting for mouse. 1 (the default) is the acquisition’s own geometry. Applies to every BET step in the masking recipe.
--bet-orientation <ORIENTATION>	Which way round BET sees the left-right axis: fsl (the default; reversed for a neurologically stored image, as FSL’s bet reads a NIfTI file) or as-stored (the voxels in the order they are stored, as MEDI’s BET.m runs bet2). Applies to every BET step in the masking recipe.
--qsm-reference <SPEC>	What the susceptibility map’s zero is pinned to: mean (the mean over the voxels the map is defined on — the brain mask less the rim background-field removal eroded; the default), none, or a SynthSeg region — a structure with both sides merged (thalamus), one side (left-thalamus), the ventricles composite, a raw FreeSurfer id (4), or a comma-separated list of any of those. A region implies –do-segmentation. An unrecognised name lists the valid ones.
--qsm-per-echo <BOOL>	Reconstruct a QSM from each echo on its own (each echo’s phase unwrapped and divided by its echo time, then background removal, dipole inversion and referencing per echo) and publish each echo’s map (desc-echo<n>_Chimap) and their mean as the QSM (Chimap), instead of one QSM from the echo-combined field map. The echo combination (–b0-estimation), phase-offset removal and bipolar correction do not run. DECOMPOSE takes the echoes’ maps. Multi-echo only; not with TGV, QSMART, iQSM, iQFM, two-pass, or χ-separation methods that fit the local field.
--rts-delta <RTS_DELTA>	RTS delta parameter.
--rts-mu <RTS_MU>	RTS mu parameter.
--rts-tol <RTS_TOL>	RTS tolerance.
--rts-rho <RTS_RHO>	RTS rho (ADMM penalty)
--rts-max-iter <RTS_MAX_ITER>	RTS max iterations.
--rts-lsmr-iter <RTS_LSMR_ITER>	RTS LSMR iterations.
--rts-tau <RTS_TAU>	RTS: rho is raised when the primal residual falls by less than this factor (tau)
--rts-gamma <RTS_GAMMA>	RTS: the factor rho is raised by (gamma)
--rts-pad <X> <Y> <Z>	RTS zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tv-lambda <TV_LAMBDA>	TV lambda parameter.
--tv-rho <RHO>	TV rho (ADMM penalty), or auto for 100 × lambda (default: auto)
--tv-tol <TV_TOL>	TV tolerance.
--tv-max-iter <TV_MAX_ITER>	TV max iterations.
--tv-pad <X> <Y> <Z>	TV zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tkd-threshold <TKD_THRESHOLD>	TKD threshold.
--tkd-pad <X> <Y> <Z>	TKD zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tsvd-threshold <TSVD_THRESHOLD>	TSVD threshold.
--tsvd-pad <X> <Y> <Z>	TSVD zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tgv-iterations <TGV_ITERATIONS>	TGV iterations (default: from the voxel size, as the reference does)
--tgv-erosions <TGV_EROSIONS>	TGV erosions.
--tgv-alpha1 <TGV_ALPHA1>	TGV alpha1 (first-order weight), in radians at the echo time (–tgv-te) and field strength (the reference’s units)
--tgv-alpha0 <TGV_ALPHA0>	TGV alpha0 (second-order weight), in radians at the echo time (–tgv-te) and field strength (the reference’s units)
--tgv-step-size <TGV_STEP_SIZE>	TGV primal step size multiplier.
--tgv-te <TE>	TGV: the echo time a field from several echoes is converted to phase at, and its weights are in radians at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first). A single echo is at its own.
--tikhonov-lambda <TIKHONOV_LAMBDA>	Tikhonov lambda.
--tikhonov-pad <X> <Y> <Z>	Tikhonov zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--nltv-lambda <NLTV_LAMBDA>	NLTV lambda.
--nltv-rho <RHO>	NLTV rho (ADMM penalty), or auto for 100 × lambda (default: auto)
--nltv-mu <NLTV_MU>	NLTV mu (penalty parameter)
--nltv-tol <NLTV_TOL>	NLTV tolerance.
--nltv-max-iter <NLTV_MAX_ITER>	NLTV max iterations.
--nltv-tol-newton <NLTV_TOL_NEWTON>	NLTV convergence tolerance of the Newton iterations of the data-term subproblem.
--nltv-newton-iter <NLTV_NEWTON_ITER>	NLTV maximum Newton iterations of the data-term subproblem.
--nltv-te <TE>	NLTV: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--nltv-pad <X> <Y> <Z>	NLTV zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--medi-lambda <MEDI_LAMBDA>	MEDI lambda: the data-fidelity weight (larger fits the field more closely)
--medi-merit <MEDI_MERIT>	MEDI: enable MERIT weighting.
--medi-csf <MEDI_CSF>	MEDI+0 (Liu et al. 2018): zero-reference χ to the ventricular CSF, found from R2*, and keep it uniform there (default: true). Needs at least 3 echoes; with fewer, MEDI runs without it.
--medi-lambda-csf <MEDI_LAMBDA_CSF>	MEDI+0: weight of the CSF uniformity term (default: 100)
--medi-csf-r2s-threshold <MEDI_CSF_R2S_THRESHOLD>	MEDI+0: R2* in Hz below which a voxel is a CSF candidate (default: 5)
--medi-csf-erode <MEDI_CSF_ERODE>	MEDI+0: erode the mask by a 10 mm sphere before finding the CSF (default: true)
--medi-zeropad <X> <Y> <Z>	MEDI: voxels of zeros appended at the end of x, y and z before the inversion; the result is cropped back (default: 0 0 0)
--medi-smv	MEDI: high-pass the local field with a spherical mean value filter, removing residual background near the brain edge (MEDI-SMV; on by default)
--no-medi-smv	MEDI: no SMV high-pass of the local field.
--medi-smv-radius <MEDI_SMV_RADIUS>	MEDI SMV radius in mm (also mSMV’s)
--medi-smv-shrink-mask	MEDI-SMV: shrink the mask to where the SMV of the mask exceeds 0.999 (on by default)
--no-medi-smv-shrink-mask	MEDI-SMV: keep the full mask, high-passing the field over it.
--medi-msmv	MEDI-mSMV (Roberts et al. 2024): filter the local field by maximum spherical mean value before MEDI, which fits it with the SMV-filtered dipole kernel (it implies the SMV). Needs at least 3 echoes.
--medi-msmv-maxk <MEDI_MSMV_MAXK>	MEDI-mSMV: the boundary correction makes at most this many passes, less one.
--medi-msmv-vessel-radius <MEDI_MSMV_VESSEL_RADIUS>	MEDI-mSMV: largest vessel thickness of the vessel mask, in voxels (default: 15, or 8 × the largest voxel dimension if that is more)
--no-medi-msmv-prefilter	MEDI-mSMV: no SMV prefilter before mSMV. Only consistent for a field already high-passed by the same SMV; after PDF, V-SHARP or LBV it diverges (expert)
--medi-data-weighting <MEDI_DATA_WEIGHTING>	MEDI: data weighting mode (0=uniform, 1=SNR)
--medi-percentage <MEDI_PERCENTAGE>	MEDI edge percentage (0.0-1.0)
--medi-cg-tol <MEDI_CG_TOL>	MEDI CG tolerance.
--medi-cg-max-iter <MEDI_CG_MAX_ITER>	MEDI CG max iterations.
--medi-max-iter <MEDI_MAX_ITER>	MEDI max outer iterations.
--medi-tol <MEDI_TOL>	MEDI outer tolerance.
--medi-te <TE>	MEDI: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)
--tfi-lambda <TFI_LAMBDA>	TFI lambda: the data-fidelity weight (larger fits the field more closely)
--tfi-precond <TFI_PRECOND>	TFI preconditioner value (susceptibility scaling outside the mask)
--tfi-merit <TFI_MERIT>	TFI: enable MERIT weighting (default: true)
--tfi-linear <TFI_LINEAR>	TFI: fit the field linearly, as the TFI paper does (default: true); false uses MEDI’s exponential data term.
--tfi-data-weighting <TFI_DATA_WEIGHTING>	TFI: data weighting mode (0=uniform, 1=SNR)
--tfi-percentage <TFI_PERCENTAGE>	TFI edge percentage (0.0-1.0)
--tfi-cg-tol <TFI_CG_TOL>	TFI CG tolerance.
--tfi-cg-max-iter <TFI_CG_MAX_ITER>	TFI CG max iterations.
--tfi-max-iter <TFI_MAX_ITER>	TFI max outer iterations.
--tfi-tol <TFI_TOL>	TFI outer tolerance.
--tfi-te <TE>	TFI: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)
--ilsqr-tol <ILSQR_TOL>	iLSQR tolerance of the streaking-artefact solve.
--ilsqr-max-iter <ILSQR_MAX_ITER>	iLSQR max iterations of each LSQR solve (default: 100)
--ilsqr-pad-mm <MM>	iLSQR zero-padding in mm per side.
--ilsqr-precision <PRECISION>	iLSQR arithmetic of the two LSQR solves: double, or single as the original implementation (default: double)
--star-pad <X> <Y> <Z>	STAR-QSM zero-padding in voxels per side, X Y Z (default: 8 8 20)
--star-te <TE>	Echo time STAR-QSM converts the field to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: sum)
--lsqr-residual-weighting <LSQR_RESIDUAL_WEIGHTING>	LSQR residual-field weight w — the only thing regularising the solve. Omit to use qsm-core’s field-strength-scaled default; 0 drops the residual term and streaks badly.
--no-lsqr-global-offset	Do not fit a global field offset alongside χ.
--lsqr-tol <LSQR_TOL>	LSQR convergence tolerance.
--lsqr-max-iter <LSQR_MAX_ITER>	LSQR max iterations.
--heidi-cone-threshold <HEIDI_CONE_THRESHOLD>	|D(k)| above which a coefficient is well-conditioned and kept from the seed map.
--heidi-gradient-threshold <HEIDI_GRADIENT_THRESHOLD>	Field-gradient threshold (ppm/mm); below it a direction reads as homogeneous.
--no-heidi-laplacian-correction	Do not open the homogeneity masks where the field’s Laplacian is large.
--heidi-laplacian-threshold <HEIDI_LAPLACIAN_THRESHOLD>	Field-Laplacian threshold (ppm per voxel²)
--heidi-gradient-mask-floor <HEIDI_GRADIENT_MASK_FLOOR>	TV weight kept where every direction reads as an edge (0 reproduces upstream’s shipped binarised masks)
--heidi-continuation-steps <HEIDI_CONTINUATION_STEPS>	NESTA continuation steps on μ.
--heidi-inner-iterations <HEIDI_INNER_ITERATIONS>	Accelerated-gradient iterations per continuation step.
--heidi-mu-min <HEIDI_MU_MIN>	Final (smallest) smoothing parameter μ.
--heidi-tol <HEIDI_TOL>	Relative objective change that ends an inner loop early (the default runs every iteration)
--no-heidi-denoise	Skip the anisotropic-diffusion smoothing of the field before its gradients are taken.
--heidi-denoise-iterations <HEIDI_DENOISE_ITERATIONS>	Anisotropic-diffusion iterations.
--heidi-denoise-time-step <HEIDI_DENOISE_TIME_STEP>	Anisotropic-diffusion time step.
--heidi-denoise-conductance <HEIDI_DENOISE_CONDUCTANCE>	Anisotropic-diffusion conductance.
--heidi-denoise-reference-itk	ITK 3.20’s float arithmetic for the denoising, as the authors’ code runs it (for comparison with it; the default computes its conductance scale in double precision)
--no-heidi-denoise-reference-itk	The denoising in double precision, after a pipeline that turns ITK’s float arithmetic on (the HEIDI pipeline)
--heidi-te <TE>	HEIDI: the echo time the field stands for the phase of, in the conversion to HEIDI’s units: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: last)
--qsmart-ilsqr-max-iter <QSMART_ILSQR_MAX_ITER>	QSMART iLSQR max iterations per stage. The iLSQR tolerance is –ilsqr-tol.
--qsmart-ilsqr-pad-mm <MM>	Zero-padding of QSMART’s inner iLSQR solves, in mm per side (default: 6)
--qsmart-vasc-sphere-radius <QSMART_VASC_SPHERE_RADIUS>	QSMART vasculature detection sphere radius, in –qsmart-vessel-units (default 8)
--qsmart-inversion <QSMART_INVERSION>	QSMART inner dipole inversion algorithm (default: ilsqr) Values: rts, tv, tkd, tsvd, tgv, tikhonov, nltv, medi, tfi, ilsqr, star, lsqr, heidi, qsmart, ndi, fansi, fansi-tgv, l1qsm, whqsm, hdqsm, amp-pe, xqsm, qsmnet, qsmnet-plus, qsmnetme, autoqsm, qsmgan, ir2qsm, lpcnn, modl-qsm, nextqsm, iqsm, iqsm-plus.
--qsmart-sdf-sigma1-stage1 <QSMART_SDF_SIGMA1_STAGE1>	QSMART SDF sigma1, stage 1 (voxels)
--qsmart-sdf-sigma2-stage1 <QSMART_SDF_SIGMA2_STAGE1>	QSMART SDF sigma2, stage 1 (voxels)
--qsmart-sdf-sigma1-stage2 <QSMART_SDF_SIGMA1_STAGE2>	QSMART SDF sigma1, stage 2 (voxels)
--qsmart-sdf-sigma2-stage2 <QSMART_SDF_SIGMA2_STAGE2>	QSMART SDF sigma2, stage 2 (voxels)
--qsmart-sdf-lower-lim <QSMART_SDF_LOWER_LIM>	QSMART SDF proximity lower limit.
--qsmart-sdf-curv-constant <QSMART_SDF_CURV_CONSTANT>	QSMART SDF curvature constant.
--qsmart-frangi-scale-min <QSMART_FRANGI_SCALE_MIN>	QSMART Frangi min vessel radius, in –qsmart-vessel-units (default 0.5)
--qsmart-frangi-scale-max <QSMART_FRANGI_SCALE_MAX>	QSMART Frangi max vessel radius, in –qsmart-vessel-units (default 6)
--qsmart-frangi-scale-ratio <QSMART_FRANGI_SCALE_RATIO>	QSMART Frangi scale step, in –qsmart-vessel-units (default 0.5)
--qsmart-frangi-c <QSMART_FRANGI_C>	QSMART Frangi C noise threshold.
--qsmart-vessel-units <QSMART_VESSEL_UNITS>	Units of QSMART’s vessel sizes (–qsmart-vasc-sphere-radius and the Frangi scales): voxels (default, as QSMART gives them) or mm (converted to voxels with the mean voxel size) Values: voxels, mm.
--qsmart-vessel-magnitude <QSMART_VESSEL_MAGNITUDE>	The magnitude QSMART detects vessels on: mean-n4 (the mean over echoes, N4 bias-corrected, as QSMART prepares it; default) or combined (the root sum of squares over echoes, inhomogeneity-corrected unless –no-inhomogeneity-correction) Values: mean-n4, combined.
--qsmart-offset-grid <QSMART_OFFSET_GRID>	The dipole grid of QSMART’s offset adjustment: fft (default) or qsmart-reference (adjust_offset.m’s own, half a step off the FFT grid on odd matrix sizes, as the QSMART pipeline runs it) Values: fft, qsmart-reference.
--ndi-tau <NDI_TAU>	NDI gradient-descent step size (tau)
--ndi-alpha <NDI_ALPHA>	NDI L2 regularization weight (alpha)
--ndi-max-iter <NDI_MAX_ITER>	NDI max iterations.
--ndi-auto-stop <NDI_AUTO_STOP>	NDI: stop automatically once noise starts to build up along the dipole cone (FANSI ndi_auto); –ndi-max-iter is then the maximum.
--ndi-te <TE>	NDI: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--ndi-weight-norm <NORM>	NDI magnitude-weight normaliser: global-max (the magnitude’s maximum, as FANSI; default) or mask-p999 (its 99.9th percentile inside the mask; more consistent in vivo)
--fansi-alpha1 <FANSI_ALPHA1>	FANSI first-order (TV/TGV) L1 penalty weight (alpha1)
--fansi-mu1 <MU1>	FANSI gradient-consistency ADMM weight (mu1), or auto for 100 × alpha1 (default: auto)
--fansi-mu2 <FANSI_MU2>	FANSI fidelity-consistency ADMM weight (mu2)
--fansi-alpha0 <ALPHA0>	FANSI second-order L1 penalty weight (alpha0, nlTGV only), or auto for 2 × alpha1 (default: auto)
--fansi-mu0 <MU0>	FANSI second-order consistency ADMM weight (mu0, nlTGV only), or auto for 2 × mu1 (default: auto)
--fansi-max-iter <FANSI_MAX_ITER>	FANSI outer ADMM iterations.
--fansi-tol-update <FANSI_TOL_UPDATE>	FANSI percent-update convergence tolerance.
--fansi-tol-delta <FANSI_TOL_DELTA>	FANSI inner Newton convergence tolerance.
--fansi-te <TE>	FANSI (nlTV and nlTGV): the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)
--l1qsm-alpha1 <L1QSM_ALPHA1>	L1-QSM gradient (TV) L1 penalty weight (alpha1)
--l1qsm-mu1 <MU1>	L1-QSM gradient-consistency ADMM weight (mu1), or auto for 100 × alpha1 (default: auto)
--l1qsm-mu2 <L1QSM_MU2>	L1-QSM fidelity-consistency ADMM weight (mu2)
--l1qsm-mu3 <L1QSM_MU3>	L1-QSM L1 proximal ADMM weight (mu3)
--l1qsm-lambda <L1QSM_LAMBDA>	L1-QSM L1 fidelity strength (lambda)
--l1qsm-max-iter <L1QSM_MAX_ITER>	L1-QSM outer ADMM iterations.
--l1qsm-tol-update <L1QSM_TOL_UPDATE>	L1-QSM percent-update convergence tolerance.
--l1qsm-tol-delta <L1QSM_TOL_DELTA>	L1-QSM inner Newton convergence tolerance.
--l1qsm-te <TE>	L1-QSM: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--whqsm-alpha1 <WHQSM_ALPHA1>	WH-QSM TV regularization weight (alpha1)
--whqsm-mu1 <WHQSM_MU1>	WH-QSM ADMM penalty for TV splitting (mu1)
--whqsm-mu2 <WHQSM_MU2>	WH-QSM ADMM penalty for data-fidelity splitting (mu2)
--whqsm-beta <WHQSM_BETA>	WH-QSM weak-harmonic ROI penalty (beta)
--whqsm-muh <WHQSM_MUH>	WH-QSM ADMM penalty for harmonic-field splitting (muh)
--whqsm-max-iter <WHQSM_MAX_ITER>	WH-QSM max outer iterations.
--whqsm-tol-update <WHQSM_TOL_UPDATE>	WH-QSM percent-update convergence tolerance.
--whqsm-tol-delta <WHQSM_TOL_DELTA>	WH-QSM inner Newton convergence tolerance.
--whqsm-weight <WEIGHT>	WH-QSM data weight: flat (1 over the whole volume, WH_nlTV.m’s default), flat-in-mask (1 inside the mask only), max-normalised (the magnitude over its maximum) or in-vivo (min(2 × magnitude / maximum, 1)); the last two need magnitude images (default: flat)
--whqsm-te <TE>	WH-QSM: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--hdqsm-alpha-l2 <HDQSM_ALPHA_L2>	HD-QSM L2-stage TV weight (alpha_l2)
--hdqsm-mu1-l2 <MU1_L2>	HD-QSM L2-stage gradient-consistency ADMM weight (mu1_l2), or auto for 100 × alpha_l2 (default: auto)
--hdqsm-mu2 <HDQSM_MU2>	HD-QSM fidelity consistency weight (mu2)
--hdqsm-max-iter-l1 <HDQSM_MAX_ITER_L1>	HD-QSM stage-1 (L1) iterations (max_iter_l1)
--hdqsm-max-iter-l2 <HDQSM_MAX_ITER_L2>	HD-QSM stage-2 (L2) iterations (max_iter_l2)
--hdqsm-tol-update <HDQSM_TOL_UPDATE>	HD-QSM stage-2 percent-update convergence tolerance.
--amp-pe-wave-order <AMP_PE_WAVE_ORDER>	AMP-PE Daubechies wavelet order (1=db1, 2=db2)
--amp-pe-nlevel <AMP_PE_NLEVEL>	AMP-PE wavelet decomposition levels.
--amp-pe-wave-pec <AMP_PE_WAVE_PEC>	AMP-PE morphology-mask energy retention fraction (0.0-1.0)
--amp-pe-simulated-te <AMP_PE_SIMULATED_TE>	AMP-PE simulated echo time (s) used to turn the field into phase.
--amp-pe-max-linearization-ite <AMP_PE_MAX_LINEARIZATION_ITE>	AMP-PE linearization iterations per stage.
--amp-pe-damp-rate-sig <AMP_PE_DAMP_RATE_SIG>	AMP-PE GAMP signal-update damping rate.
--amp-pe-damp-rate-par <AMP_PE_DAMP_RATE_PAR>	AMP-PE parameter-estimation learning rate (kappa)
--amp-pe-max-pe-spar-ite <AMP_PE_MAX_PE_SPAR_ITE>	AMP-PE inner sparse-reconstruction iterations.
--amp-pe-max-pe-est-ite <AMP_PE_MAX_PE_EST_ITE>	AMP-PE inner parameter-estimation iterations.
--amp-pe-cvg-thd <AMP_PE_CVG_THD>	AMP-PE GAMP inner convergence threshold.
--amp-pe-tikhonov-beta <AMP_PE_TIKHONOV_BETA>	AMP-PE L2-seed Tikhonov weight.
--r2star-qsm-r-const-3t <R2STAR_QSM_R_CONST_3T>	R2*-QSM relaxometric constant at 3T (Hz/ppm)
--decompose-n-inner <DECOMPOSE_N_INNER>	DECOMPOSE inner alternating passes per voxel.
--decompose-chi-bound <BOUND>	DECOMPOSE bound on |χ| in the fit: a number in ppm (default: 0.5), or echo-spacing to compute it from the field strength and echo spacing.
--decompose-unwrap-phase <BOOL>	DECOMPOSE: unwrap the phase along the echoes (default: false, as in the authors’ code)
--decompose-fit <DECOMPOSE_FIT>	DECOMPOSE’s fit: reference-code (Tim Ho’s script; default) or paper (Chen et al. 2021: amplitudes summing to one, per-voxel S(0) = 1 scale), as the DECOMPOSE pipeline runs it. Values: reference-code, paper.
--decompose-max-iterations <DECOMPOSE_MAX_ITERATIONS>	DECOMPOSE: maximum solver iterations of each fit (lsqcurvefit MaxIterations; default: 400)
--decompose-max-function-evaluations <DECOMPOSE_MAX_FUNCTION_EVALUATIONS>	DECOMPOSE: maximum function evaluations of each fit (lsqcurvefit MaxFunctionEvaluations; default: 100 × the fit’s parameters)
--decompose-function-tolerance <DECOMPOSE_FUNCTION_TOLERANCE>	DECOMPOSE: relative cost change each fit stops at (lsqcurvefit FunctionTolerance; default: 1e-6)
--decompose-step-tolerance <DECOMPOSE_STEP_TOLERANCE>	DECOMPOSE: relative step each fit stops at (lsqcurvefit StepTolerance; default: 1e-6)
--decompose-optimality-tolerance <DECOMPOSE_OPTIMALITY_TOLERANCE>	DECOMPOSE: first-order optimality each fit stops at (lsqcurvefit OptimalityTolerance; default: 1e-6)
--decompose-finite-difference-step <DECOMPOSE_FINITE_DIFFERENCE_STEP>	DECOMPOSE: relative finite-difference step of the Jacobian (lsqcurvefit FiniteDifferenceStepSize; default: √eps)
--decompose-typical-x <DECOMPOSE_TYPICAL_X>	DECOMPOSE: typical magnitude of each fitted value (lsqcurvefit TypicalX; default: 1)
--chi-sep-ilsqr-dr-pos <HZ_PER_PPM>	χ-sep iLSQR paramagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-ilsqr-dr-neg <HZ_PER_PPM>	χ-sep iLSQR diamagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-ilsqr-pad <X> <Y> <Z>	χ-sep iLSQR zero-padding at both ends of each axis in voxels, X Y Z (default: 16 16 16)
--chi-sep-ilsqr-cg-max-iter-lsqr <CHI_SEP_ILSQR_CG_MAX_ITER_LSQR>	χ-sep iLSQR: maximum CG iterations of the LSQR step (default: 50)
--chi-sep-ilsqr-cg-tol-lsqr <CHI_SEP_ILSQR_CG_TOL_LSQR>	χ-sep iLSQR: CG relative tolerance of the LSQR step (default: 0.02)
--chi-sep-ilsqr-cg-max-iter-sa <CHI_SEP_ILSQR_CG_MAX_ITER_SA>	χ-sep iLSQR: maximum CG iterations of the streaking-removal step (default: 50)
--chi-sep-ilsqr-cg-tol-sa <CHI_SEP_ILSQR_CG_TOL_SA>	χ-sep iLSQR: CG relative tolerance of the streaking-removal step (default: 0.02)
--chi-sep-ilsqr-d2-thr <CHI_SEP_ILSQR_D2_THR>	χ-sep iLSQR: the streaking cone is where |D| is below this (default: 0.11)
--chi-sep-ilsqr-prc-w-i-min <CHI_SEP_ILSQR_PRC_W_I_MIN>	χ-sep iLSQR: percentile of the field’s Laplacian at which the LSQR weight is 1 (default: 60)
--chi-sep-ilsqr-prc-w-i-max <CHI_SEP_ILSQR_PRC_W_I_MAX>	χ-sep iLSQR: percentile of the field’s Laplacian at which the LSQR weight is 0 (default: 99.8)
--chi-sep-ilsqr-prc-w-g-min <CHI_SEP_ILSQR_PRC_W_G_MIN>	χ-sep iLSQR: percentile of each gradient’s magnitude at which the streaking-removal weight is 1 (default: 50)
--chi-sep-ilsqr-prc-w-g-max <CHI_SEP_ILSQR_PRC_W_G_MAX>	χ-sep iLSQR: percentile of each gradient’s magnitude at which the streaking-removal weight is 0 (default: 70)
--chi-sep-ilsqr-qsm-max-iter <CHI_SEP_ILSQR_QSM_MAX_ITER>	χ-sep iLSQR: max iterations of the iLSQR that reconstructs, from the local field, the QSM χ-separation is initialised with (default: 100; not used with –use-custom-qsm)
--chi-sep-ilsqr-qsm-tol <CHI_SEP_ILSQR_QSM_TOL>	χ-sep iLSQR: tolerance of that QSM’s iLSQR.
--chi-sep-ilsqr-qsm-pad-mm <MM>	χ-sep iLSQR: zero-padding of that QSM’s iLSQR, in mm per side (default: 12)
--chi-sep-ilsqr-qsm-precision <PRECISION>	χ-sep iLSQR: arithmetic of that QSM’s two LSQR solves: double, or single as the original implementation (default: double)
--chi-sep-medi-lambda <CHI_SEP_MEDI_LAMBDA>	χ-sep MEDI weight of the data terms against the L1 terms (default: 1)
--chi-sep-medi-lambda-csf <CHI_SEP_MEDI_LAMBDA_CSF>	χ-sep MEDI weight of the CSF-uniformity term (default: 1)
--chi-sep-medi-dr-pos <HZ_PER_PPM>	χ-sep MEDI paramagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-medi-dr-neg <HZ_PER_PPM>	χ-sep MEDI diamagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-medi-pad <CHI_SEP_MEDI_PAD>	χ-sep MEDI zero-padding at both ends of the third axis, in voxels (default: 5)
--chi-sep-medi-tol <CHI_SEP_MEDI_TOL>	χ-sep MEDI: stop when the relative change falls below this (default: 0.03)
--chi-sep-medi-max-iter <CHI_SEP_MEDI_MAX_ITER>	χ-sep MEDI maximum Gauss-Newton iterations (default: 30)
--chi-sep-medi-cg-tol <CHI_SEP_MEDI_CG_TOL>	χ-sep MEDI inner CG tolerance (default: 0.05)
--chi-sep-medi-cg-max-iter <CHI_SEP_MEDI_CG_MAX_ITER>	χ-sep MEDI inner CG max iterations (default: 100)
--chi-sep-medi-qsm-max-iter <CHI_SEP_MEDI_QSM_MAX_ITER>	χ-sep MEDI: max iterations of the iLSQR that reconstructs, from the local field, the QSM χ-separation is initialised with (default: 100; not used with –use-custom-qsm)
--chi-sep-medi-qsm-tol <CHI_SEP_MEDI_QSM_TOL>	χ-sep MEDI: tolerance of that QSM’s iLSQR.
--chi-sep-medi-qsm-pad-mm <MM>	χ-sep MEDI: zero-padding of that QSM’s iLSQR, in mm per side (default: 12)
--chi-sep-medi-qsm-precision <PRECISION>	χ-sep MEDI: arithmetic of that QSM’s two LSQR solves: double, or single as the original implementation (default: double)
--chi-sep-medi-csf-r2star <METHOD>	χ-sep MEDI: the R2* its CSF mask is extracted from: arlo (MEDI’s ARLO; the default) or nnls (the χ-separation toolbox’s r2star_nnls, which its script runs)
--wavesep-dr-pos <WAVESEP_DR_POS>	WaveSep paramagnetic relaxivity (Hz/ppm)
--wavesep-dr-neg <WAVESEP_DR_NEG>	WaveSep diamagnetic relaxivity (Hz/ppm)
--wavesep-alpha <WAVESEP_ALPHA>	WaveSep proximal-gradient step size.
--wavesep-lambda <WAVESEP_LAMBDA>	WaveSep wavelet L1 soft-threshold weight.
--wavesep-wavelet-order <WAVESEP_WAVELET_ORDER>	WaveSep Daubechies wavelet order.
--wavesep-max-iter <WAVESEP_MAX_ITER>	WaveSep ISTA max iterations.
--wavesep-tol <WAVESEP_TOL>	WaveSep relative-change stop tolerance.
--hc-chisep-dr-pos-3t <HC_CHISEP_DR_POS_3T>	HC-ChiSep paramagnetic relaxivity at 3T (Hz/ppm)
--hc-chisep-bin-hz <HC_CHISEP_BIN_HZ>	HC-ChiSep R2’ bin width for the anchored grid search (Hz)
--chi-sepnet-route <CHI_SEPNET_ROUTE>	χ-sepnet’s input route. toolbox (default), the authors’: the inputs resampled to 1 mm with B0 along z, the susceptibility channel computed by QSMnetME, R2’ measured or predicted from R2* by R2PRIMEnet, the maps resampled back. native: the network alone on the acquired grid, given this run’s QSM and R2’. Values: toolbox, native.
--chi-sepnet-patch <D> <H> <W>	χ-sepnet (--chi-sepnet-route native) sliding-window patch in voxels, D H W, each a multiple of 16 (default: 192 192 128, the authors’)
--chi-sepnet-frame-patch <D> <H> <W>	χ-sepnet (--chi-sepnet-route toolbox): run the networks as a sliding window of patches this size over the 1 mm frame, D H W, each a multiple of 16, so memory follows the patch. Omit to run the whole frame at once, as the authors do.
--susep-net-patch <D> <H> <W>	Run SUSEP-Net as a sliding window of patches this size in voxels, D H W, each a multiple of 8, so its memory follows the patch, not the volume. Omit to run the whole volume at once.
--motion-correction	Rigidly co-register every echo onto a reference echo (estimated on magnitude, applied to phase through the complex signal) before field mapping, and restrict the mask to voxels every echo covers. Off by default. Needs magnitude images.
--motion-reference <first|last|N>	Echo the others are aligned onto: first (default), last, or a 1-based echo number.
--no-motion-chain	Register every echo from the identity instead of seeding it from its neighbour’s solution.
--motion-chain-levels <MOTION_CHAIN_LEVELS>	Pyramid levels for a seeded (chained) registration (default 3)
--motion-min-ncc <MOTION_MIN_NCC>	NCC below which an echo’s registration is reported as suspect (default 0.3, set for echoes; ~0.9 suits repeats)
--registration-levels <REGISTRATION_LEVELS>	Registration pyramid levels: the capture range; do not lower (default 5)
--registration-max-samples <REGISTRATION_MAX_SAMPLES>	Registration: voxels sampled per metric evaluation, per level (default 200000)
--registration-max-iterations <REGISTRATION_MAX_ITERATIONS>	Registration: Powell cycles per level (default 20)
--registration-max-evaluations <REGISTRATION_MAX_EVALUATIONS>	Registration: metric evaluations across all levels, a hard stop (default 5000)
--registration-tolerance <REGISTRATION_TOLERANCE>	Registration: relative cost improvement per cycle that counts as converged (default 1e-5)
--registration-rotation-step-deg <REGISTRATION_ROTATION_STEP_DEG>	Registration: initial rotation step at the finest level, degrees (default 3)
--registration-translation-step-mm <REGISTRATION_TRANSLATION_STEP_MM>	Registration: initial translation step at the finest level, mm (default 3)
--registration-min-overlap <REGISTRATION_MIN_OVERLAP>	Registration: smallest fraction of sampled voxels that must overlap (default 0.25)
--vsharp-deconvolve <BOOL>	Undo V-SHARP’s SMV high-pass with a truncated-SVD deconvolution (default: false)
--vsharp-threshold <VSHARP_THRESHOLD>	V-SHARP deconvolution (TSVD) threshold; only used with –vsharp-deconvolve true.
--vsharp-max-radius <VSHARP_MAX_RADIUS>	V-SHARP max kernel radius in mm.
--vsharp-min-radius <VSHARP_MIN_RADIUS>	V-SHARP min kernel radius in mm.
--vsharp-radius-step <MM>	V-SHARP step between kernel radii in mm (default: 1); the radii run from –vsharp-max-radius down to –vsharp-min-radius.
--vsharp-grid <VSHARP_GRID>	Grid V-SHARP estimates the background on: isotropic (default; spacing from –vsharp-grid-spacing), native (the acquired grid), or finest (isotropic at the smallest voxel dimension, e.g. 0.8 mm for 0.8x0.8x3 mm; no axis is coarsened). The local field keeps native resolution. Values: native, finest, isotropic.
--vsharp-grid-spacing <MM>	V-SHARP isotropic working-grid spacing in mm (default 1); used with –vsharp-grid isotropic.
--pdf-tol <PDF_TOL>	PDF tolerance, relative to the noise level.
--pdf-n-cg <PDF_N_CG>	PDF maximum conjugate-gradient iterations (default: 30)
--pdf-space <SPACE>	PDF dipole kernel: image (the FFT of the dipole sampled on the voxel grid) or kspace (the analytic k-space kernel) (default: image)
--pdf-n-pad <VOXELS>	PDF zero-padding around the mask’s bounding box, in voxels; 0 runs on the full volume (default: 40)
--lbv-tol <LBV_TOL>	LBV tolerance: stop the coarsest grid’s relaxation when its summed absolute update falls below this, in radians over the echo spacing (default: 0.01)
--lbv-depth <LEVEL>	LBV coarsest multigrid level, at most 8 (default: log2 of the smallest dimension minus 3)
--lbv-peel <LAYERS>	LBV boundary layers removed from the mask beyond the first (default: 0)
--lbv-n1 <LBV_N1>	LBV relaxations on each grid before the coarser grid (default: 30)
--lbv-n2 <LBV_N2>	LBV relaxations on each grid after the coarser grid (default: 100)
--lbv-n3 <LBV_N3>	LBV relaxations on the finest grid at the end (default: 100)
--lbv-precision <PRECISION>	LBV arithmetic: double, or single as the original implementation (default: double)
--ismv-variant <VARIANT>	iSMV implementation: wen2014 (the authors’ paper and code) or qsmjl (QSM.jl’s) (default: wen2014)
--ismv-tol <ISMV_TOL>	iSMV tolerance: the relative change between iterates it stops at (qsmjl: the residual relative to the first) (default: 5e-5)
--ismv-max-iter <ISMV_MAX_ITER>	iSMV max iterations (default: no limit)
--ismv-radius <MM>	iSMV kernel radius in mm, or auto for 6 × the smallest voxel dimension (qsmjl: 2 × the largest) (default: auto)
--sharp-threshold <SHARP_THRESHOLD>	SHARP threshold.
--sharp-radius <SHARP_RADIUS>	SHARP kernel radius in mm.
--resharp-radius <RESHARP_RADIUS>	RESHARP SMV kernel radius in mm.
--resharp-tik-reg <RESHARP_TIK_REG>	RESHARP Tikhonov regularization parameter.
--resharp-tol <RESHARP_TOL>	RESHARP conjugate gradients squared (CGS) relative tolerance.
--resharp-max-iter <RESHARP_MAX_ITER>	RESHARP maximum CGS iterations.
--harperella-radius <HARPERELLA_RADIUS>	HARPERELLA SMV kernel radius in mm.
--harperella-max-iter <HARPERELLA_MAX_ITER>	HARPERELLA maximum iterations.
--harperella-tol <HARPERELLA_TOL>	HARPERELLA convergence tolerance.
--harperella-pad <X> <Y> <Z>	HARPERELLA zero-padding in voxels per side, X Y Z (default: 0 0 0)
--harperella-rim <RIM>	HARPERELLA boundary rim: whole (zero every Laplacian value in the rim; default) or sti (zero the 20 % largest, as STI Suite does)
--iharperella-max-iter <IHARPERELLA_MAX_ITER>	iHARPERELLA maximum iterations.
--iharperella-tol <IHARPERELLA_TOL>	iHARPERELLA convergence tolerance.
--iharperella-pad <X> <Y> <Z>	iHARPERELLA zero-padding in voxels per side, X Y Z (default: 0 0 0)
--iharperella-rim <RIM>	iHARPERELLA boundary rim: whole (zero every Laplacian value in the rim) or sti (zero the 20 % largest, as STI Suite does; default)
--no-romeo-phase-coherence	ROMEO: disable phase coherence weights (default: on)
--no-romeo-phase-gradient-coherence	ROMEO: disable phase gradient coherence weights (default: on)
--no-romeo-phase-linearity	ROMEO: disable phase linearity weights (default: on; the romeo command line’s romeo3 weights)
--romeo-phase-linearity	ROMEO: enable phase linearity weights (default: on; for a –pipeline that turns them off)
--no-romeo-mag-coherence	ROMEO: disable magnitude coherence weights (default: on)
--romeo-mag-weight	ROMEO: enable magnitude weighting (default: off)
--no-romeo-mag-weight	ROMEO: disable magnitude weighting (overrides –romeo-mag-weight; kept for compatibility)
--romeo-mag-weight2	ROMEO: enable second magnitude weighting (flow-artifact penalty; default: off)
--romeo-bestpath	ROMEO: use Best-path (Abdul-Rahman) weights instead of ROMEO weights (default: off)
--romeo-template <ROMEO_TEMPLATE>	ROMEO: template echo index (1-indexed, only for template mode)
--romeo-individual	ROMEO: use individual per-echo unwrapping (default)
--no-romeo-individual	ROMEO: use template-based temporal unwrapping (disables individual mode)
--romeo-correct-global	ROMEO: correct global 2π offsets, of each echo and between echoes (default: on)
--no-romeo-correct-global	ROMEO: do not correct global 2π offsets.
--romeo-temporal-uncertain-unwrapping <ROMEO_TEMPORAL_UNCERTAIN_UNWRAPPING>	ROMEO: temporal uncertain-unwrapping quality threshold [0,1] (0 disables; default: 0.5)
--romeo-max-seeds <ROMEO_MAX_SEEDS>	ROMEO: maximum number of seed regions (default: 1)
--romeo-merge-regions	ROMEO: merge neighboring regions after unwrapping (default: off)
--romeo-correct-regions	ROMEO: correct each region’s median to nearest 0 (default: off)
--romeo-wrap-addition <ROMEO_WRAP_ADDITION>	ROMEO: additional phase tolerance beyond π for neighbor differences [0,π] (default: 0.0)
--prelude-labelling <PRELUDE_LABELLING>	PRELUDE: how the initial regions are labelled (default: auto) Values: auto, slices, volume.
--prelude-partitions <N>	PRELUDE: number of phase partitions the initial regions are formed from (default: 8)
--swi-unwrapping <SWI_UNWRAPPING>	SWI phase unwrapping before the echoes are combined [default: laplacian]. Values: laplacian, laplacian-slice, romeo.
--swi-hp-sigma <SWI_HP_SIGMA>	SWI high-pass filter sigma (3 values, in voxels)
--swi-scaling <SWI_SCALING>	SWI phase-mask scaling. Values: tanh, negative-tanh, positive, negative, triangular.
--swi-strength <SWI_STRENGTH>	SWI phase scaling strength.
--swi-mip-window <SWI_MIP_WINDOW>	SWI MIP window size in slices.
--smwi-threshold <SMWI_THRESHOLD>	SMWI susceptibility threshold in ppm — the |χ| at which the weighting mask reaches zero. The 1 ppm default follows SEPIA and targets strong sources; brain tissue wants far less.
--smwi-power <SMWI_POWER>	SMWI mask power (contrast strength)
--smwi-mip-window <SMWI_MIP_WINDOW>	SMWI mIP window size in slices.
--synthseg-version <VERSION>	SynthSeg weights generation: v1 (32 labels, ships with SynthSeg) or v2 (33 labels, adds CSF; weights not published yet, so refused unless synthseg-2.0.onnx is in $QSM_MODEL_DIR)
--synthseg-crop <SYNTHSEG_CROP>	Cubic centre-crop in voxels at 1 mm before inference (rounded up to a multiple of 32). Omit to process the whole volume, as the SynthSeg CLI does.
--synthseg-patch <VOXELS>	Run SynthSeg as a sliding window of half-overlapping cubes this many voxels on a side at 1 mm (at least 64, rounded up to a multiple of 32), so its memory follows the patch, not the volume. Omit to run the whole volume at once, as the SynthSeg CLI does.
--no-synthseg-flip-averaging	Skip the left-right flipped second pass. Halves inference time, as SynthSeg’s –fast does.
--no-synthseg-topology-cleanup	Skip resetting each topological class to its largest connected component.
--synthseg-sigma <SYNTHSEG_SIGMA>	Gaussian blur on the posteriors, in voxels (0 disables it)
--use-custom-dseg <TOOL>	Prefer a bring-your-own segmentation from <bids>/derivatives/<TOOL>/ when present. Bare flag = first matching tool alphabetically. Always falls back to running SynthSeg.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--inhomogeneity-method <INHOMOGENEITY_METHOD>	Inhomogeneity correction method: makehomogeneous (default; –homogeneity-*) or n4 (N4 bias field correction, as ANTs’ N4BiasFieldCorrection; –n4-*) Values: makehomogeneous, n4.
--homogeneity-sigma-mm <HOMOGENEITY_SIGMA_MM>	Inhomogeneity correction smoothing sigma in mm (makehomogeneous)
--homogeneity-nbox <HOMOGENEITY_NBOX>	makehomogeneous: boxes per dimension for the tissue segmentation the receive sensitivity is estimated from.
--n4-shrink-factor <N4_SHRINK_FACTOR>	N4: subsample the image by this factor before fitting the bias field (default 4)
--n4-max-iterations <ITERATIONS>	N4: maximum iterations at each fitting level, one number a level (default 50 50 50 50); the B-spline mesh doubles from one level to the next.
--n4-convergence-threshold <N4_CONVERGENCE_THRESHOLD>	N4: stop a level once the change in the bias field (its coefficient of variation) falls to this (default 0: every iteration runs)
--n4-mesh <X> <Y> <Z>	N4: B-spline mesh elements per axis at the first level, X Y Z (default 1 1 1)
--n4-spline-order <N4_SPLINE_ORDER>	N4: order of the B-spline bias field, 1 to 3 (default 3)
--n4-bias-field-fwhm <N4_BIAS_FIELD_FWHM>	N4: FWHM of the Gaussian modelling the bias field’s blurring of the log-intensity histogram (default 0.15)
--n4-wiener-noise <N4_WIENER_NOISE>	N4: noise term of the Wiener deconvolution of the histogram (default 0.01)
--n4-histogram-bins <N4_HISTOGRAM_BINS>	N4: bins of the log-intensity histogram (default 200)
--n4-geometry <GEOMETRY>	N4: the image geometry it is given: scanner (the scan’s header, as ANTs reads the acquired NIfTI; the default) or make-nii (the axis-aligned header MATLAB’s make_nii writes, as QSMART runs N4). N4’s shrunk grid falls where the header puts it.
--linear-fit-estimate-offset <LINEAR_FIT_ESTIMATE_OFFSET>	Linear and robust fit: estimate a phase offset term (default: true). The robust fit needs 4 echoes with it, 3 without.
--robust-fit-loss <ROBUST_FIT_LOSS>	Motion-robust fit (--b0-estimation robust-fit): loss turning residuals into weights. tukey (default) rejects echoes past the tuning radius; huber caps but never rejects. Values: tukey, huber.
--robust-fit-tuning <ROBUST_FIT_TUNING>	Motion-robust fit: tuning radius in robust-scale units (default: the loss’s conventional value, Tukey 4.685 / Huber 1.345). Smaller is more aggressive.
--robust-fit-iterations <ROBUST_FIT_ITERATIONS>	Motion-robust fit: IRLS passes (default 3; 0 = report only, the field is the linear fit’s)
--robust-fit-downweight-level <ROBUST_FIT_DOWNWEIGHT_LEVEL>	Motion-robust fit (reporting only): a robust weight below this counts as down-weighted (default 0.75; in (0, 1])
--robust-fit-flag-ratio <ROBUST_FIT_FLAG_RATIO>	Motion-robust fit (reporting only): flag an echo whose outlier score exceeds this many times the best echo’s (default 3; at least 1)
--complex-fit-reltol <COMPLEX_FIT_RELTOL>	Complex fit (--b0-estimation complex-fit): stop once the update is this fraction of the first estimate, over the whole volume (default 1e-4)
--complex-fit-max-iter <COMPLEX_FIT_MAX_ITER>	Complex fit: Gauss-Newton iterations at most (default 30)
--no-qsm	Skip QSM processing (only run supplementary outputs like SWI, T2*, R2*)
--do-swi	Also compute SWI.
--swi-filtered-phase	Also write CLEAR-SWI’s high-pass filtered phase, the phase the SWI is weighted by (_desc-swi_phase.nii): Hz for several echoes, radians for one, 0 outside CLEAR-SWI’s head mask. Implies –do-swi.
--no-magnitude	Do not publish the combined magnitude (part-mag_T2starw). It is still computed, and kept with the run’s intermediates, when a step reads it (SMWI, segmentation, a magnitude-weighted inversion)
--do-segmentation	Also segment the brain from the GRE magnitude with SynthSeg, writing a FreeSurfer-labelled _dseg.nii and its BIDS lookup table. Needs no T1w and no registration. Needs a deep-learning build. doi:10.1016/j.media.2023.102789.
--do-analysis	Also summarise every quantitative map (χ, χ+/χ−, T2*, R2*, R2, R2’) per segmented structure into a TSV: mean, SD, median, min, max and the 5th/95th percentiles, one row per map and structure, plus a dataset-level table gathering all runs. Summarises whatever the run made — it implies –do-segmentation, and QSM only if no other map was asked for.
--do-smwi	Also compute SMWI (susceptibility map-weighted imaging): weights the magnitude by a mask built from the susceptibility map rather than from high-pass filtered phase, so the contrast stays where the source is. Needs χ, so it implies QSM. doi:10.1002/mrm.24920.
--do-t2starmap	Compute T2* relaxation map from multi-echo magnitude data.
--do-r2starmap	Compute R2* decay rate map from multi-echo magnitude data.
--do-r2map	Compute R2 map from a multi-echo spin-echo (MESE) acquisition (EPG)
--do-r2primemap	Compute R2’ map (= R2* − R2; needs GRE magnitude + a MESE acquisition)
--do-chisep	Compute chi-separation (paramagnetic/diamagnetic susceptibility maps)
--chisep <CHI_SEPARATION_ALGORITHM>	Chi-separation method (default: r2star-qsm) Values: r2star-qsm, decompose, chi-sep-ilsqr, chi-sep-medi, wavesep, hc-chisep, susep-net, chi-sepnet.
--use-custom-qsm <TOOL>	Use a bring-your-own QSM (Chimap) from <bids>/derivatives/<TOOL>/ for chi-separation.
--use-custom-r2 <TOOL>	Use a bring-your-own R2 map from <bids>/derivatives/<TOOL>/ for chi-separation.
--use-custom-r2prime <TOOL>	Use a bring-your-own R2’ map from <bids>/derivatives/<TOOL>/ for chi-separation.
--r2prime-strategy <R2PRIME_STRATEGY>	How to obtain R2’ when no custom map is supplied. auto (default) measures it as R2* - R2 from a MESE acquisition when one is present and predicts it from R2* with R2PRIMEnet when it is not; mese only ever measures it, producing nothing without spin-echo data; r2primenet always predicts it. A measurement and an estimate are not the same thing — the generated methods text says which route was configured. Values: auto, mese, r2primenet.
--r2primenet-ignore-field-mismatch	Run R2PRIMEnet even when the acquisition is not at 3 T, the only field strength its weights were trained at; the predicted R2’ is then uncalibrated. Without it, an explicit --r2prime-strategy r2primenet refuses a non-3 T run and auto skips R2’ with a warning.
--r2primenet-patch <D> <H> <W>	R2PRIMEnet sliding-window patch in voxels, D H W, each a multiple of 16 (default: 192 192 128, the authors’)
--export-dicom	Also export final maps as DICOM series into each subject’s extra_files/ folder.
--inhomogeneity-correction	Apply inhomogeneity correction to magnitude before masking.
--no-inhomogeneity-correction	Disable inhomogeneity correction.
--axial-resampling	Resample oblique acquisitions onto an axial grid when their obliquity exceeds –obliquity-threshold. Off by default (methods that only work on axial data resample regardless)
--no-axial-resampling	Do not resample oblique acquisitions to axial (overrides a config’s axial_resampling)
--obliquity-threshold <OBLIQUITY_THRESHOLD>	Obliquity in degrees above which –axial-resampling resamples (default: 10)
--crop-to-mask	Reconstruct inside a box around the mask instead of the whole field of view. Cuts the FFT-bound stages roughly in proportion to the voxels removed, but those stages are a small share of a full run, so the end-to-end gain is usually modest. Off by default. Outputs are written on the full grid either way; this only changes where the arithmetic happens. Padding to an FFT-friendly size happens regardless.
--ignore-slice-gap	Reconstruct even when a 2D acquisition’s sidecar SliceThickness is thinner than the slice pitch. Without it, QSM refuses such a run: the gap between slices is unsampled tissue, and the FFT kernels assume a contiguous grid, so the map would be wrong.
--fft-padding	Pad the grid outward to an FFT-friendly size before the FFT-bound stages. Much faster on an awkward grid, and discards nothing, but it changes where k-space is sampled and so shifts the reconstruction slightly. Off by default.
--tukey-window <TUKEY_WINDOW>	Apply a k-space Tukey window with this taper ratio (0 to 1) to each echo’s complex image before every other step, as the χ-separation toolbox does: 0.4, and 0 for Philips data. Default 0 (none). Needs magnitude images.
--crop-margin-mm <CROP_MARGIN_MM>	Margin in mm around the mask when –crop-to-mask is set (default 32). Below ~32 mm the reconstruction starts to differ from the uncropped one; it must also clear the largest SMV kernel in use, and V-SHARP defaults to 12 mm.
--output-space <OUTPUT_SPACE>	Grid to write derivatives on when the run was resampled to axial. acquired (default) puts them back where the data came from, so transforms you already hold still apply; working leaves them on the resampled grid and avoids a second interpolation. Values: acquired, working.
--mask-preset <MASK_PRESET>	Mask preset: robust-threshold, bet, hd-bet (HD-BET deep-learning brain extraction + signal-gated erosion; needs a deep-learning build), bet-and-phase (BET on the magnitude AND a thresholded phase-quality map, then hole-filled and eroded), rs2-net (RS2-Net deep-learning rodent brain extraction), or mindgrab (MindGrab deep-learning brain extraction for any contrast; needs a deep-learning build) Values: robust-threshold, bet, hd-bet, bet-and-phase, rs2-net, mindgrab.
--use-custom-masks <USE_CUSTOM_MASKS>	Prefer a bring-your-own mask from BIDS derivatives when present (else compute it). Bare flag = first matching derivatives tool (alphabetical); give a name (e.g. bet) to restrict to derivatives/<tool>/sub-*/anat/*_mask.nii*. Always falls back.
--mask <MASK_SECTIONS_CLI>	Define a mask section (repeatable; multiple sections are combined with –mask-combine). Format: <input>,<generator>,<refinement1>,<refinement2>,… Example: phase-quality,threshold:otsu,dilate:2,fill-holes:0,erode:2 Example: magnitude,bet:0.5,erode:2 Example: magnitude-first,bet:0.2,min-intensity:100,majority,close:2 (QSMART’s mask) Example: magnitude,hd-bet,signal-erode (HD-BET: hd-bet[:full|window|XxYxZ|low-memory][:tta][:step=F]; signal-erode: signal-erode[:threshold[:depth_cap[:global_erosions[:bias_sigma[:min_component]]]]])
--mask-combine <MASK_COMBINE>	How to combine multiple –mask sections: or (union, the default) or and (intersection — a voxel survives only where every section keeps it) Values: or, and.
--mask-refine <MASK_REFINEMENTS_CLI>	Refinement applied to the combined mask, after –mask-combine (repeatable). Takes the same refinement ops as –mask, e.g. erode:2 or fill-holes:0; generators (threshold/bet/hd-bet) belong in a –mask section instead. Example: –mask-combine and –mask-refine fill-holes:0 –mask-refine erode:1.
--two-pass	Two-pass artefact reduction: reconstruct a second time against a mask whose holes are left unfilled, and keep that reconstruction wherever it is defined. Strong susceptibility sources (air-tissue interfaces, haemorrhage, implants) make the phase around them unrecoverable; leaving those voxels out keeps their streaking from spreading, while the ordinary mask still supplies values inside the holes. Roughly doubles background removal and inversion time, and is not free of downsides — where the holes were not artefactual it can be worse than a single pass. Based on doi:10.1002/mrm.29048 (Stewart et al.). Off by default. The conventional single-pass map is kept alongside as _desc-singlepass_Chimap.nii.
--two-pass-mask <TWO_PASS_SECTIONS_CLI>	Define the reliable-pass mask used by –two-pass (repeatable; same format as –mask). Defaults to phase-quality,threshold:otsu — an unfilled threshold on the phase-quality map. Its holes are the point, so refinements that close them (fill-holes, close, gaussian) defeat the method. Folded with –two-pass-mask-combine, then –mask-refine like the main mask, and always intersected with it.
--two-pass-mask-combine <TWO_PASS_MASK_COMBINE>	How to combine multiple –two-pass-mask sections: or (union) or and (intersection). Defaults to the main mask’s –mask-combine. Turns on –two-pass. Values: or, and.
--dismiss-recommendation <RULE=VALUE>	Stop offering a recommendation for the setting’s current value: RULE=VALUE, the rule’s id and the value it was dismissed for, as qsmxt run --dry lists them (e.g. field-mapping.phase-offsets=false). Repeatable. A change to the setting brings the recommendation back. A required one cannot be dismissed, and a reference-pipeline difference has nothing to dismiss.
--n-procs <N_PROCS>	Number of parallel threads.
--source-dicom <SOURCE_DICOM>	Optional source DICOM directory; inherit patient/study identity from the original DICOMs when exporting (used with –export-dicom)
--dicom-outputs <DICOM_OUTPUTS>	Restrict DICOM export to these maps (default: all produced); repeat the flag or comma-separate them (used with –export-dicom) Values: magnitude, chimap, swi, minip, desc-swi_phase, desc-paramagnetic_smwi, desc-diamagnetic_smwi, desc-paramagneticsmwi_minip, desc-diamagneticsmwi_minip, t2starmap, r2starmap, r2map, r2primemap, desc-paramagnetic_chimap, desc-diamagnetic_chimap, desc-total_chimap, dseg.
--dry	Print the processing plan, the pipeline graph and the methods text without running anything or writing any file. Also lists what the configuration departs from in the chosen methods’ reference pipelines, for every run’s scan, and each covered setting’s reference value; with –accept-recommendations or –accept-reference-settings, the settings that would change.
--accept-recommendations	Apply the recommended value of every reasoned recommendation that is not dismissed (required, recommended and discouraged findings, each with its reason) before running. The original authors’ settings are not applied; see –accept-reference-settings. Two equally strong recommendations that disagree on a setting are left for you to choose. Without it, no setting is ever changed.
--accept-reference-settings	Apply the original authors’ settings of the chosen methods (what their code, paper or README does, where no reason is given that it is better) before running, so the run follows the methods’ reference pipelines. Two methods whose authors’ settings disagree on a setting are left for you to choose; a reasoned recommendation on the same setting wins when –accept-recommendations is also given.
--debug	Enable debug logging.
--mem-limit-gb <MEM_LIMIT_GB>	Memory limit in GB for concurrent run scheduling (if not specified: the memory available, within a Slurm job’s or container’s memory limit when there is one)
--no-mem-limit	Disable memory-aware concurrency limiting.
--force	Force re-run, ignoring cached pipeline state.
--clean-intermediates	Once every run (and any multi-orientation combination) is done, delete the intermediate images under workflow/. Published outputs, per-step provenance and the cache state are kept: a re-run with the same configuration recomputes nothing, one with a changed configuration recomputes the intermediates it needs.
--no-dataset-steps	Process each run and stop: skip the dataset-level steps (multi-orientation reconstruction, summary tables, cohort figures, provenance, dataset_description.json, pipeline_config.toml, methods.md, the pipeline graph, changes/) and log to logs/<run>.log instead of qsmxt.log. qsmxt aggregate does those steps once every run is done; qsmxt slurm uses the pair.

qsmxt init

Generate a pipeline configuration file

qsmxt init [OPTIONS]

Argument	Description
-o, --output <OUTPUT>	Output file path (prints to stdout if not specified)

qsmxt validate

Validate BIDS dataset structure for QSM processing

qsmxt validate [OPTIONS] <BIDS_DIR>

Argument	Description
<BIDS_DIR>	Input BIDS directory. Required.
--include <INCLUDE>	Include only runs matching these glob patterns.
--exclude <EXCLUDE>	Exclude runs matching these glob patterns.

qsmxt dicom-convert

Convert a DICOM directory to BIDS using automatic series classification

qsmxt dicom-convert [OPTIONS] <DICOM_DIR> <OUTPUT_DIR>

Argument	Description
<DICOM_DIR>	Input DICOM directory (searched recursively) Required.
<OUTPUT_DIR>	Output BIDS directory. Required.
--dry-run	Print the auto-detected series classification and exit without converting.
--keep-dicom-uids	Keep the study’s DICOM UIDs (Study, Series and Frame of Reference) and dates in the sidecars, so qsmxt run --export-dicom files its series into the original study. Patient identity is still removed, but the UIDs link the dataset to the original study: don’t share such a dataset outside your site.

qsmxt slurm

Generate SLURM job scripts for HPC execution: one job per run, then an aggregation job

qsmxt slurm [OPTIONS] --account <ACCOUNT> <BIDS_DIR> [OUTPUT_DIR]

Argument	Description
<BIDS_DIR>	Input BIDS directory. Required.
<OUTPUT_DIR>	Output directory (defaults to bids_dir; outputs go into <dir>/derivatives/qsmxt/)
--account <ACCOUNT>	SLURM account name. Required.
--partition <PARTITION>	SLURM partition.
--config <CONFIG>	Pipeline configuration file (TOML)
--time <TIME>	Wall time limit (e.g., 02:00:00) Default: 02:00:00.
--mem <MEM>	Memory per job in GB [default: estimated per run from its size and the pipeline].
--cpus-per-task <CPUS_PER_TASK>	CPUs per task. Default: 4.
--submit	Auto-submit scripts via sbatch.
--include <INCLUDE>	Include only runs matching these glob patterns (e.g. “sub-1*” “*ses-pre*”)
--exclude <EXCLUDE>	Exclude runs matching these glob patterns (e.g. “*mygrea*”)
--num-echoes <NUM_ECHOES>	Limit number of echoes to process.
--force	Force re-run, ignoring cached pipeline state (passed to every job)
--clean-intermediates	Once every run and the dataset-level steps are done, delete the intermediate images under workflow/ (done by the aggregation job)
--debug	Enable debug logging (in every job)
--source-dicom <SOURCE_DICOM>	Optional source DICOM directory; inherit patient/study identity from the original DICOMs when exporting (used with –export-dicom; each run’s job exports its maps)
--dicom-outputs <DICOM_OUTPUTS>	Restrict DICOM export to these maps (default: all produced); repeat the flag or comma-separate them (used with –export-dicom) Values: magnitude, chimap, swi, minip, desc-swi_phase, desc-paramagnetic_smwi, desc-diamagnetic_smwi, desc-paramagneticsmwi_minip, desc-diamagneticsmwi_minip, t2starmap, r2starmap, r2map, r2primemap, desc-paramagnetic_chimap, desc-diamagnetic_chimap, desc-total_chimap, dseg.
--accept-recommendations	Apply the recommended value of every reasoned recommendation that is not dismissed before writing the jobs (as qsmxt run --accept-recommendations)
--accept-reference-settings	Apply the original authors’ settings of the chosen methods before writing the jobs (as qsmxt run --accept-reference-settings)
--pipeline <PIPELINE>	Start from a pipeline: qsmxt-defaults, or an original toolbox’s (medi-toolbox, sti-suite-ilsqr, sti-suite-star, qsmart, heidi, chi-sep-ilsqr, chi-sep-medi, chi-sepnet, decompose). The stages it covers (masking, field mapping, background removal, inversion, referencing, χ-separation) go back to QSMxT’s defaults and then take the toolbox’s settings; the other flags given then change any setting. Values: qsmxt-defaults, medi-toolbox, sti-suite-ilsqr, sti-suite-star, qsmart, heidi, chi-sep-ilsqr, chi-sep-medi, chi-sepnet, decompose.
--qsm-algorithm <QSM_ALGORITHM>	QSM dipole inversion algorithm (default: ilsqr) Values: rts, tv, tkd, tsvd, tgv, tikhonov, nltv, medi, tfi, ilsqr, star, lsqr, heidi, qsmart, ndi, fansi, fansi-tgv, l1qsm, whqsm, hdqsm, amp-pe, xqsm, qsmnet, qsmnet-plus, qsmnetme, autoqsm, qsmgan, ir2qsm, lpcnn, modl-qsm, nextqsm, iqsm, iqsm-plus.
--heidi	HEIDI streaking reduction after the inversion: keep the map’s well-conditioned k-space and refill the dipole cone under a homogeneity prior (--heidi-* sets it up). With LSQR this is HEIDI as published (--qsm-algorithm heidi); with two-pass it runs once, on the combined map.
--orientation-group <PATTERN>	Group runs into multi-orientation sets and reconstruct each with COSMOS or STI. Takes an entity name (acq, run, rec, inv), a glob over the run key (*acq-dir*), re: and a regex with one capture group, or runs: and the groups’ run keys (, between runs, ; between groups, each group’s reference first). Orientations are co-registered if they need it.
--orientation-reference <PATTERN>	The orientation each group is registered onto, and whose grid and segmentation the group map uses: a glob over the run key (*acq-Neutral*) or a word in it. Must match one run per group. Default: the first run in name order.
--no-orientation-registration	Do not co-register the orientations in a group — use this when they were already registered externally and a second interpolation is not wanted. A group that still needs registration is then refused instead of reconstructed.
--multi-orientation-algorithm <MULTI_ORIENTATION_ALGORITHM>	Which reconstruction to run over each orientation group. Values: cosmos, sti.
--multi-orientation-lambda <MULTI_ORIENTATION_LAMBDA>	Regularization for the multi-orientation reconstruction (COSMOS k-space L2, or STI ridge)
--multi-orientation-force	Reconstruct an orientation group even when its B0 directions look degenerate.
--cosmos-magnitude-weighting	COSMOS: weight each orientation by its combined magnitude (conjugate-gradient solve)
--cosmos-tol <COSMOS_TOL>	Magnitude-weighted COSMOS: conjugate-gradient tolerance (needs –cosmos-magnitude-weighting)
--cosmos-max-iter <COSMOS_MAX_ITER>	Magnitude-weighted COSMOS: conjugate-gradient iteration cap (needs –cosmos-magnitude-weighting)
--unwrapping-algorithm <UNWRAPPING_ALGORITHM>	Unwrapping algorithm (default: romeo) Values: romeo, laplacian, prelude, medi-region-growing.
--bf-algorithm <BF_ALGORITHM>	Background field removal algorithm (default: vsharp) Values: vsharp, pdf, lbv, ismv, sharp, resharp, harperella, iharperella, bfrnet, iqfm.
--bf-erode-voxels <N>	Erode the brain mask by N voxels after field mapping, before background removal; the field map is not masked again (default: 0, none)
--bf-polish-mask	After background removal, keep the mask’s largest connected component with its holes filled and zero the local field outside it, as STI Suite’s QSM_GUI does (PolishMask)
--no-bf-polish-mask	Do not polish the background-removal mask (the default), after a --pipeline that does.
--masking-input <MASKING_INPUT>	Masking input image (overrides the input of the configured mask sections; combine with –mask-preset, e.g. –mask-preset robust-threshold –masking-input magnitude) Values: magnitude-first, magnitude, magnitude-last, phase-quality.
--phase-offset-removal <PHASE_OFFSET_REMOVAL>	Enable phase offset removal for multi-echo data (default: true)
--phase-offset-sigma-mm <PHASE_OFFSET_SIGMA_MM>	Phase offset removal smoothing sigma (3 values, in mm; default: 7 7 7, the ROMEO command line’s)
--coil-combination-sigma <COIL_COMBINATION_SIGMA>	MCPC-3D-S coil combination: smoothing sigma for the per-coil phase offsets (3 values, in voxels; default 10 10 5, as in MriResearchTools). Only used for uncombined multi-coil runs.
--bipolar-correction	Enable bipolar gradient correction (requires >= 3 echoes)
--b0-estimation <B0_ESTIMATION>	B0 field estimation method (default: weighted-avg) Values: weighted-avg, linear-fit, robust-fit, qsmart-echofit, complex-fit.
--qsmart-fit-threshold-mode <QSMART_FIT_THRESHOLD_MODE>	--b0-estimation qsmart-echofit: how the fit residual at which a voxel is left out of QSMART’s reconstruction is set: a percentile of the blurred residual over the volume (--qsmart-fit-percentile), or a fixed residual (--qsmart-fit-threshold; default, QSMART v1.0’s 40). Without it, giving one of those two selects its mode. Values: percentile, absolute.
--qsmart-fit-percentile <QSMART_FIT_PERCENTILE>	--b0-estimation qsmart-echofit: the percentile of the blurred fit residual, in (0, 100], above which a voxel is left out of QSMART’s reconstruction (default 99)
--qsmart-fit-threshold <QSMART_FIT_THRESHOLD>	--b0-estimation qsmart-echofit: a fixed blurred fit residual above which a voxel is left out of QSMART’s reconstruction, instead of a percentile (40 when the mode is absolute and none is given)
--b0-weight-type <B0_WEIGHT_TYPE>	B0 weighted averaging weight type. assumed-decay gives each echo’s phase a fixed weight TE·exp(−TE/T2*) computed from the echo times and an assumed T2* (--b0-weight-assumed-t2star), the same in every voxel: no T2* map, no magnitude (Wang et al. 2022, Nat Neurosci 25:818). phase-snr, which uses the measured magnitude in each voxel, is generally preferable when magnitude is available. Applies to Laplacian unwrapping as well as ROMEO. Values: phase-snr, phase-var, average, tes, mag, assumed-decay.
--b0-weight-assumed-t2star <MS>	T2* in ms assumed by --b0-weight-type assumed-decay to compute its fixed echo weights (default 40; no T2* map is used)
--laplacian-solver <LAPLACIAN_SOLVER>	Poisson solver for --unwrapping-algorithm laplacian: dct (unweighted least squares, Neumann/DCT solve; default), fft (Schofield & Zhu sin/cos Laplacian, FFT solve on a zero-padded volume) or mirror (the same on the volume mirrored along each axis, as QSMART unwraps) Values: dct, fft, mirror.
--laplacian-fft-pad <VOXELS>	Zero-padding in voxels per side for --laplacian-solver fft: one value for every axis, or three (x y z). Default 64.
--laplacian-unmasked	--unwrapping-algorithm laplacian: unwrap the raw phase over the whole volume, instead of the phase set to zero outside the brain mask, as QSMART does. With --b0-estimation qsmart-echofit the field map is then kept over the whole volume.
--no-laplacian-unmasked	--unwrapping-algorithm laplacian: unwrap the phase set to zero outside the brain mask (the default), after a --pipeline that unwraps the whole volume.
--bet-fractional-intensity <BET_FRACTIONAL_INTENSITY>	BET fractional intensity (0.0-1.0; smaller = larger brain). Applies to every BET step in the masking recipe, overriding a bet:<f> in a –mask spec.
--bet-smoothness <BET_SMOOTHNESS>	BET surface smoothness factor. Applies to every BET step in the masking recipe.
--bet-gradient-threshold <BET_GRADIENT_THRESHOLD>	BET gradient threshold (-1 to 1). Applies to every BET step in the masking recipe.
--bet-iterations <BET_ITERATIONS>	BET surface evolution iterations. Applies to every BET step in the masking recipe.
--bet-subdivisions <BET_SUBDIVISIONS>	BET icosphere subdivision level. Applies to every BET step in the masking recipe.
--bet-voxel-scale <FACTOR>	Factor applied to the voxel sizes before BET runs. BET’s surface model is tuned to a human brain in millimetres, so preclinical data has to be presented at human scale: 10 is the usual setting for mouse. 1 (the default) is the acquisition’s own geometry. Applies to every BET step in the masking recipe.
--bet-orientation <ORIENTATION>	Which way round BET sees the left-right axis: fsl (the default; reversed for a neurologically stored image, as FSL’s bet reads a NIfTI file) or as-stored (the voxels in the order they are stored, as MEDI’s BET.m runs bet2). Applies to every BET step in the masking recipe.
--qsm-reference <SPEC>	What the susceptibility map’s zero is pinned to: mean (the mean over the voxels the map is defined on — the brain mask less the rim background-field removal eroded; the default), none, or a SynthSeg region — a structure with both sides merged (thalamus), one side (left-thalamus), the ventricles composite, a raw FreeSurfer id (4), or a comma-separated list of any of those. A region implies –do-segmentation. An unrecognised name lists the valid ones.
--qsm-per-echo <BOOL>	Reconstruct a QSM from each echo on its own (each echo’s phase unwrapped and divided by its echo time, then background removal, dipole inversion and referencing per echo) and publish each echo’s map (desc-echo<n>_Chimap) and their mean as the QSM (Chimap), instead of one QSM from the echo-combined field map. The echo combination (–b0-estimation), phase-offset removal and bipolar correction do not run. DECOMPOSE takes the echoes’ maps. Multi-echo only; not with TGV, QSMART, iQSM, iQFM, two-pass, or χ-separation methods that fit the local field.
--rts-delta <RTS_DELTA>	RTS delta parameter.
--rts-mu <RTS_MU>	RTS mu parameter.
--rts-tol <RTS_TOL>	RTS tolerance.
--rts-rho <RTS_RHO>	RTS rho (ADMM penalty)
--rts-max-iter <RTS_MAX_ITER>	RTS max iterations.
--rts-lsmr-iter <RTS_LSMR_ITER>	RTS LSMR iterations.
--rts-tau <RTS_TAU>	RTS: rho is raised when the primal residual falls by less than this factor (tau)
--rts-gamma <RTS_GAMMA>	RTS: the factor rho is raised by (gamma)
--rts-pad <X> <Y> <Z>	RTS zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tv-lambda <TV_LAMBDA>	TV lambda parameter.
--tv-rho <RHO>	TV rho (ADMM penalty), or auto for 100 × lambda (default: auto)
--tv-tol <TV_TOL>	TV tolerance.
--tv-max-iter <TV_MAX_ITER>	TV max iterations.
--tv-pad <X> <Y> <Z>	TV zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tkd-threshold <TKD_THRESHOLD>	TKD threshold.
--tkd-pad <X> <Y> <Z>	TKD zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tsvd-threshold <TSVD_THRESHOLD>	TSVD threshold.
--tsvd-pad <X> <Y> <Z>	TSVD zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tgv-iterations <TGV_ITERATIONS>	TGV iterations (default: from the voxel size, as the reference does)
--tgv-erosions <TGV_EROSIONS>	TGV erosions.
--tgv-alpha1 <TGV_ALPHA1>	TGV alpha1 (first-order weight), in radians at the echo time (–tgv-te) and field strength (the reference’s units)
--tgv-alpha0 <TGV_ALPHA0>	TGV alpha0 (second-order weight), in radians at the echo time (–tgv-te) and field strength (the reference’s units)
--tgv-step-size <TGV_STEP_SIZE>	TGV primal step size multiplier.
--tgv-te <TE>	TGV: the echo time a field from several echoes is converted to phase at, and its weights are in radians at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first). A single echo is at its own.
--tikhonov-lambda <TIKHONOV_LAMBDA>	Tikhonov lambda.
--tikhonov-pad <X> <Y> <Z>	Tikhonov zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--nltv-lambda <NLTV_LAMBDA>	NLTV lambda.
--nltv-rho <RHO>	NLTV rho (ADMM penalty), or auto for 100 × lambda (default: auto)
--nltv-mu <NLTV_MU>	NLTV mu (penalty parameter)
--nltv-tol <NLTV_TOL>	NLTV tolerance.
--nltv-max-iter <NLTV_MAX_ITER>	NLTV max iterations.
--nltv-tol-newton <NLTV_TOL_NEWTON>	NLTV convergence tolerance of the Newton iterations of the data-term subproblem.
--nltv-newton-iter <NLTV_NEWTON_ITER>	NLTV maximum Newton iterations of the data-term subproblem.
--nltv-te <TE>	NLTV: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--nltv-pad <X> <Y> <Z>	NLTV zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--medi-lambda <MEDI_LAMBDA>	MEDI lambda: the data-fidelity weight (larger fits the field more closely)
--medi-merit <MEDI_MERIT>	MEDI: enable MERIT weighting.
--medi-csf <MEDI_CSF>	MEDI+0 (Liu et al. 2018): zero-reference χ to the ventricular CSF, found from R2*, and keep it uniform there (default: true). Needs at least 3 echoes; with fewer, MEDI runs without it.
--medi-lambda-csf <MEDI_LAMBDA_CSF>	MEDI+0: weight of the CSF uniformity term (default: 100)
--medi-csf-r2s-threshold <MEDI_CSF_R2S_THRESHOLD>	MEDI+0: R2* in Hz below which a voxel is a CSF candidate (default: 5)
--medi-csf-erode <MEDI_CSF_ERODE>	MEDI+0: erode the mask by a 10 mm sphere before finding the CSF (default: true)
--medi-zeropad <X> <Y> <Z>	MEDI: voxels of zeros appended at the end of x, y and z before the inversion; the result is cropped back (default: 0 0 0)
--medi-smv	MEDI: high-pass the local field with a spherical mean value filter, removing residual background near the brain edge (MEDI-SMV; on by default)
--no-medi-smv	MEDI: no SMV high-pass of the local field.
--medi-smv-radius <MEDI_SMV_RADIUS>	MEDI SMV radius in mm (also mSMV’s)
--medi-smv-shrink-mask	MEDI-SMV: shrink the mask to where the SMV of the mask exceeds 0.999 (on by default)
--no-medi-smv-shrink-mask	MEDI-SMV: keep the full mask, high-passing the field over it.
--medi-msmv	MEDI-mSMV (Roberts et al. 2024): filter the local field by maximum spherical mean value before MEDI, which fits it with the SMV-filtered dipole kernel (it implies the SMV). Needs at least 3 echoes.
--medi-msmv-maxk <MEDI_MSMV_MAXK>	MEDI-mSMV: the boundary correction makes at most this many passes, less one.
--medi-msmv-vessel-radius <MEDI_MSMV_VESSEL_RADIUS>	MEDI-mSMV: largest vessel thickness of the vessel mask, in voxels (default: 15, or 8 × the largest voxel dimension if that is more)
--no-medi-msmv-prefilter	MEDI-mSMV: no SMV prefilter before mSMV. Only consistent for a field already high-passed by the same SMV; after PDF, V-SHARP or LBV it diverges (expert)
--medi-data-weighting <MEDI_DATA_WEIGHTING>	MEDI: data weighting mode (0=uniform, 1=SNR)
--medi-percentage <MEDI_PERCENTAGE>	MEDI edge percentage (0.0-1.0)
--medi-cg-tol <MEDI_CG_TOL>	MEDI CG tolerance.
--medi-cg-max-iter <MEDI_CG_MAX_ITER>	MEDI CG max iterations.
--medi-max-iter <MEDI_MAX_ITER>	MEDI max outer iterations.
--medi-tol <MEDI_TOL>	MEDI outer tolerance.
--medi-te <TE>	MEDI: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)
--tfi-lambda <TFI_LAMBDA>	TFI lambda: the data-fidelity weight (larger fits the field more closely)
--tfi-precond <TFI_PRECOND>	TFI preconditioner value (susceptibility scaling outside the mask)
--tfi-merit <TFI_MERIT>	TFI: enable MERIT weighting (default: true)
--tfi-linear <TFI_LINEAR>	TFI: fit the field linearly, as the TFI paper does (default: true); false uses MEDI’s exponential data term.
--tfi-data-weighting <TFI_DATA_WEIGHTING>	TFI: data weighting mode (0=uniform, 1=SNR)
--tfi-percentage <TFI_PERCENTAGE>	TFI edge percentage (0.0-1.0)
--tfi-cg-tol <TFI_CG_TOL>	TFI CG tolerance.
--tfi-cg-max-iter <TFI_CG_MAX_ITER>	TFI CG max iterations.
--tfi-max-iter <TFI_MAX_ITER>	TFI max outer iterations.
--tfi-tol <TFI_TOL>	TFI outer tolerance.
--tfi-te <TE>	TFI: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)
--ilsqr-tol <ILSQR_TOL>	iLSQR tolerance of the streaking-artefact solve.
--ilsqr-max-iter <ILSQR_MAX_ITER>	iLSQR max iterations of each LSQR solve (default: 100)
--ilsqr-pad-mm <MM>	iLSQR zero-padding in mm per side.
--ilsqr-precision <PRECISION>	iLSQR arithmetic of the two LSQR solves: double, or single as the original implementation (default: double)
--star-pad <X> <Y> <Z>	STAR-QSM zero-padding in voxels per side, X Y Z (default: 8 8 20)
--star-te <TE>	Echo time STAR-QSM converts the field to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: sum)
--lsqr-residual-weighting <LSQR_RESIDUAL_WEIGHTING>	LSQR residual-field weight w — the only thing regularising the solve. Omit to use qsm-core’s field-strength-scaled default; 0 drops the residual term and streaks badly.
--no-lsqr-global-offset	Do not fit a global field offset alongside χ.
--lsqr-tol <LSQR_TOL>	LSQR convergence tolerance.
--lsqr-max-iter <LSQR_MAX_ITER>	LSQR max iterations.
--heidi-cone-threshold <HEIDI_CONE_THRESHOLD>	|D(k)| above which a coefficient is well-conditioned and kept from the seed map.
--heidi-gradient-threshold <HEIDI_GRADIENT_THRESHOLD>	Field-gradient threshold (ppm/mm); below it a direction reads as homogeneous.
--no-heidi-laplacian-correction	Do not open the homogeneity masks where the field’s Laplacian is large.
--heidi-laplacian-threshold <HEIDI_LAPLACIAN_THRESHOLD>	Field-Laplacian threshold (ppm per voxel²)
--heidi-gradient-mask-floor <HEIDI_GRADIENT_MASK_FLOOR>	TV weight kept where every direction reads as an edge (0 reproduces upstream’s shipped binarised masks)
--heidi-continuation-steps <HEIDI_CONTINUATION_STEPS>	NESTA continuation steps on μ.
--heidi-inner-iterations <HEIDI_INNER_ITERATIONS>	Accelerated-gradient iterations per continuation step.
--heidi-mu-min <HEIDI_MU_MIN>	Final (smallest) smoothing parameter μ.
--heidi-tol <HEIDI_TOL>	Relative objective change that ends an inner loop early (the default runs every iteration)
--no-heidi-denoise	Skip the anisotropic-diffusion smoothing of the field before its gradients are taken.
--heidi-denoise-iterations <HEIDI_DENOISE_ITERATIONS>	Anisotropic-diffusion iterations.
--heidi-denoise-time-step <HEIDI_DENOISE_TIME_STEP>	Anisotropic-diffusion time step.
--heidi-denoise-conductance <HEIDI_DENOISE_CONDUCTANCE>	Anisotropic-diffusion conductance.
--heidi-denoise-reference-itk	ITK 3.20’s float arithmetic for the denoising, as the authors’ code runs it (for comparison with it; the default computes its conductance scale in double precision)
--no-heidi-denoise-reference-itk	The denoising in double precision, after a pipeline that turns ITK’s float arithmetic on (the HEIDI pipeline)
--heidi-te <TE>	HEIDI: the echo time the field stands for the phase of, in the conversion to HEIDI’s units: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: last)
--qsmart-ilsqr-max-iter <QSMART_ILSQR_MAX_ITER>	QSMART iLSQR max iterations per stage. The iLSQR tolerance is –ilsqr-tol.
--qsmart-ilsqr-pad-mm <MM>	Zero-padding of QSMART’s inner iLSQR solves, in mm per side (default: 6)
--qsmart-vasc-sphere-radius <QSMART_VASC_SPHERE_RADIUS>	QSMART vasculature detection sphere radius, in –qsmart-vessel-units (default 8)
--qsmart-inversion <QSMART_INVERSION>	QSMART inner dipole inversion algorithm (default: ilsqr) Values: rts, tv, tkd, tsvd, tgv, tikhonov, nltv, medi, tfi, ilsqr, star, lsqr, heidi, qsmart, ndi, fansi, fansi-tgv, l1qsm, whqsm, hdqsm, amp-pe, xqsm, qsmnet, qsmnet-plus, qsmnetme, autoqsm, qsmgan, ir2qsm, lpcnn, modl-qsm, nextqsm, iqsm, iqsm-plus.
--qsmart-sdf-sigma1-stage1 <QSMART_SDF_SIGMA1_STAGE1>	QSMART SDF sigma1, stage 1 (voxels)
--qsmart-sdf-sigma2-stage1 <QSMART_SDF_SIGMA2_STAGE1>	QSMART SDF sigma2, stage 1 (voxels)
--qsmart-sdf-sigma1-stage2 <QSMART_SDF_SIGMA1_STAGE2>	QSMART SDF sigma1, stage 2 (voxels)
--qsmart-sdf-sigma2-stage2 <QSMART_SDF_SIGMA2_STAGE2>	QSMART SDF sigma2, stage 2 (voxels)
--qsmart-sdf-lower-lim <QSMART_SDF_LOWER_LIM>	QSMART SDF proximity lower limit.
--qsmart-sdf-curv-constant <QSMART_SDF_CURV_CONSTANT>	QSMART SDF curvature constant.
--qsmart-frangi-scale-min <QSMART_FRANGI_SCALE_MIN>	QSMART Frangi min vessel radius, in –qsmart-vessel-units (default 0.5)
--qsmart-frangi-scale-max <QSMART_FRANGI_SCALE_MAX>	QSMART Frangi max vessel radius, in –qsmart-vessel-units (default 6)
--qsmart-frangi-scale-ratio <QSMART_FRANGI_SCALE_RATIO>	QSMART Frangi scale step, in –qsmart-vessel-units (default 0.5)
--qsmart-frangi-c <QSMART_FRANGI_C>	QSMART Frangi C noise threshold.
--qsmart-vessel-units <QSMART_VESSEL_UNITS>	Units of QSMART’s vessel sizes (–qsmart-vasc-sphere-radius and the Frangi scales): voxels (default, as QSMART gives them) or mm (converted to voxels with the mean voxel size) Values: voxels, mm.
--qsmart-vessel-magnitude <QSMART_VESSEL_MAGNITUDE>	The magnitude QSMART detects vessels on: mean-n4 (the mean over echoes, N4 bias-corrected, as QSMART prepares it; default) or combined (the root sum of squares over echoes, inhomogeneity-corrected unless –no-inhomogeneity-correction) Values: mean-n4, combined.
--qsmart-offset-grid <QSMART_OFFSET_GRID>	The dipole grid of QSMART’s offset adjustment: fft (default) or qsmart-reference (adjust_offset.m’s own, half a step off the FFT grid on odd matrix sizes, as the QSMART pipeline runs it) Values: fft, qsmart-reference.
--ndi-tau <NDI_TAU>	NDI gradient-descent step size (tau)
--ndi-alpha <NDI_ALPHA>	NDI L2 regularization weight (alpha)
--ndi-max-iter <NDI_MAX_ITER>	NDI max iterations.
--ndi-auto-stop <NDI_AUTO_STOP>	NDI: stop automatically once noise starts to build up along the dipole cone (FANSI ndi_auto); –ndi-max-iter is then the maximum.
--ndi-te <TE>	NDI: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--ndi-weight-norm <NORM>	NDI magnitude-weight normaliser: global-max (the magnitude’s maximum, as FANSI; default) or mask-p999 (its 99.9th percentile inside the mask; more consistent in vivo)
--fansi-alpha1 <FANSI_ALPHA1>	FANSI first-order (TV/TGV) L1 penalty weight (alpha1)
--fansi-mu1 <MU1>	FANSI gradient-consistency ADMM weight (mu1), or auto for 100 × alpha1 (default: auto)
--fansi-mu2 <FANSI_MU2>	FANSI fidelity-consistency ADMM weight (mu2)
--fansi-alpha0 <ALPHA0>	FANSI second-order L1 penalty weight (alpha0, nlTGV only), or auto for 2 × alpha1 (default: auto)
--fansi-mu0 <MU0>	FANSI second-order consistency ADMM weight (mu0, nlTGV only), or auto for 2 × mu1 (default: auto)
--fansi-max-iter <FANSI_MAX_ITER>	FANSI outer ADMM iterations.
--fansi-tol-update <FANSI_TOL_UPDATE>	FANSI percent-update convergence tolerance.
--fansi-tol-delta <FANSI_TOL_DELTA>	FANSI inner Newton convergence tolerance.
--fansi-te <TE>	FANSI (nlTV and nlTGV): the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)
--l1qsm-alpha1 <L1QSM_ALPHA1>	L1-QSM gradient (TV) L1 penalty weight (alpha1)
--l1qsm-mu1 <MU1>	L1-QSM gradient-consistency ADMM weight (mu1), or auto for 100 × alpha1 (default: auto)
--l1qsm-mu2 <L1QSM_MU2>	L1-QSM fidelity-consistency ADMM weight (mu2)
--l1qsm-mu3 <L1QSM_MU3>	L1-QSM L1 proximal ADMM weight (mu3)
--l1qsm-lambda <L1QSM_LAMBDA>	L1-QSM L1 fidelity strength (lambda)
--l1qsm-max-iter <L1QSM_MAX_ITER>	L1-QSM outer ADMM iterations.
--l1qsm-tol-update <L1QSM_TOL_UPDATE>	L1-QSM percent-update convergence tolerance.
--l1qsm-tol-delta <L1QSM_TOL_DELTA>	L1-QSM inner Newton convergence tolerance.
--l1qsm-te <TE>	L1-QSM: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--whqsm-alpha1 <WHQSM_ALPHA1>	WH-QSM TV regularization weight (alpha1)
--whqsm-mu1 <WHQSM_MU1>	WH-QSM ADMM penalty for TV splitting (mu1)
--whqsm-mu2 <WHQSM_MU2>	WH-QSM ADMM penalty for data-fidelity splitting (mu2)
--whqsm-beta <WHQSM_BETA>	WH-QSM weak-harmonic ROI penalty (beta)
--whqsm-muh <WHQSM_MUH>	WH-QSM ADMM penalty for harmonic-field splitting (muh)
--whqsm-max-iter <WHQSM_MAX_ITER>	WH-QSM max outer iterations.
--whqsm-tol-update <WHQSM_TOL_UPDATE>	WH-QSM percent-update convergence tolerance.
--whqsm-tol-delta <WHQSM_TOL_DELTA>	WH-QSM inner Newton convergence tolerance.
--whqsm-weight <WEIGHT>	WH-QSM data weight: flat (1 over the whole volume, WH_nlTV.m’s default), flat-in-mask (1 inside the mask only), max-normalised (the magnitude over its maximum) or in-vivo (min(2 × magnitude / maximum, 1)); the last two need magnitude images (default: flat)
--whqsm-te <TE>	WH-QSM: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--hdqsm-alpha-l2 <HDQSM_ALPHA_L2>	HD-QSM L2-stage TV weight (alpha_l2)
--hdqsm-mu1-l2 <MU1_L2>	HD-QSM L2-stage gradient-consistency ADMM weight (mu1_l2), or auto for 100 × alpha_l2 (default: auto)
--hdqsm-mu2 <HDQSM_MU2>	HD-QSM fidelity consistency weight (mu2)
--hdqsm-max-iter-l1 <HDQSM_MAX_ITER_L1>	HD-QSM stage-1 (L1) iterations (max_iter_l1)
--hdqsm-max-iter-l2 <HDQSM_MAX_ITER_L2>	HD-QSM stage-2 (L2) iterations (max_iter_l2)
--hdqsm-tol-update <HDQSM_TOL_UPDATE>	HD-QSM stage-2 percent-update convergence tolerance.
--amp-pe-wave-order <AMP_PE_WAVE_ORDER>	AMP-PE Daubechies wavelet order (1=db1, 2=db2)
--amp-pe-nlevel <AMP_PE_NLEVEL>	AMP-PE wavelet decomposition levels.
--amp-pe-wave-pec <AMP_PE_WAVE_PEC>	AMP-PE morphology-mask energy retention fraction (0.0-1.0)
--amp-pe-simulated-te <AMP_PE_SIMULATED_TE>	AMP-PE simulated echo time (s) used to turn the field into phase.
--amp-pe-max-linearization-ite <AMP_PE_MAX_LINEARIZATION_ITE>	AMP-PE linearization iterations per stage.
--amp-pe-damp-rate-sig <AMP_PE_DAMP_RATE_SIG>	AMP-PE GAMP signal-update damping rate.
--amp-pe-damp-rate-par <AMP_PE_DAMP_RATE_PAR>	AMP-PE parameter-estimation learning rate (kappa)
--amp-pe-max-pe-spar-ite <AMP_PE_MAX_PE_SPAR_ITE>	AMP-PE inner sparse-reconstruction iterations.
--amp-pe-max-pe-est-ite <AMP_PE_MAX_PE_EST_ITE>	AMP-PE inner parameter-estimation iterations.
--amp-pe-cvg-thd <AMP_PE_CVG_THD>	AMP-PE GAMP inner convergence threshold.
--amp-pe-tikhonov-beta <AMP_PE_TIKHONOV_BETA>	AMP-PE L2-seed Tikhonov weight.
--r2star-qsm-r-const-3t <R2STAR_QSM_R_CONST_3T>	R2*-QSM relaxometric constant at 3T (Hz/ppm)
--decompose-n-inner <DECOMPOSE_N_INNER>	DECOMPOSE inner alternating passes per voxel.
--decompose-chi-bound <BOUND>	DECOMPOSE bound on |χ| in the fit: a number in ppm (default: 0.5), or echo-spacing to compute it from the field strength and echo spacing.
--decompose-unwrap-phase <BOOL>	DECOMPOSE: unwrap the phase along the echoes (default: false, as in the authors’ code)
--decompose-fit <DECOMPOSE_FIT>	DECOMPOSE’s fit: reference-code (Tim Ho’s script; default) or paper (Chen et al. 2021: amplitudes summing to one, per-voxel S(0) = 1 scale), as the DECOMPOSE pipeline runs it. Values: reference-code, paper.
--decompose-max-iterations <DECOMPOSE_MAX_ITERATIONS>	DECOMPOSE: maximum solver iterations of each fit (lsqcurvefit MaxIterations; default: 400)
--decompose-max-function-evaluations <DECOMPOSE_MAX_FUNCTION_EVALUATIONS>	DECOMPOSE: maximum function evaluations of each fit (lsqcurvefit MaxFunctionEvaluations; default: 100 × the fit’s parameters)
--decompose-function-tolerance <DECOMPOSE_FUNCTION_TOLERANCE>	DECOMPOSE: relative cost change each fit stops at (lsqcurvefit FunctionTolerance; default: 1e-6)
--decompose-step-tolerance <DECOMPOSE_STEP_TOLERANCE>	DECOMPOSE: relative step each fit stops at (lsqcurvefit StepTolerance; default: 1e-6)
--decompose-optimality-tolerance <DECOMPOSE_OPTIMALITY_TOLERANCE>	DECOMPOSE: first-order optimality each fit stops at (lsqcurvefit OptimalityTolerance; default: 1e-6)
--decompose-finite-difference-step <DECOMPOSE_FINITE_DIFFERENCE_STEP>	DECOMPOSE: relative finite-difference step of the Jacobian (lsqcurvefit FiniteDifferenceStepSize; default: √eps)
--decompose-typical-x <DECOMPOSE_TYPICAL_X>	DECOMPOSE: typical magnitude of each fitted value (lsqcurvefit TypicalX; default: 1)
--chi-sep-ilsqr-dr-pos <HZ_PER_PPM>	χ-sep iLSQR paramagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-ilsqr-dr-neg <HZ_PER_PPM>	χ-sep iLSQR diamagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-ilsqr-pad <X> <Y> <Z>	χ-sep iLSQR zero-padding at both ends of each axis in voxels, X Y Z (default: 16 16 16)
--chi-sep-ilsqr-cg-max-iter-lsqr <CHI_SEP_ILSQR_CG_MAX_ITER_LSQR>	χ-sep iLSQR: maximum CG iterations of the LSQR step (default: 50)
--chi-sep-ilsqr-cg-tol-lsqr <CHI_SEP_ILSQR_CG_TOL_LSQR>	χ-sep iLSQR: CG relative tolerance of the LSQR step (default: 0.02)
--chi-sep-ilsqr-cg-max-iter-sa <CHI_SEP_ILSQR_CG_MAX_ITER_SA>	χ-sep iLSQR: maximum CG iterations of the streaking-removal step (default: 50)
--chi-sep-ilsqr-cg-tol-sa <CHI_SEP_ILSQR_CG_TOL_SA>	χ-sep iLSQR: CG relative tolerance of the streaking-removal step (default: 0.02)
--chi-sep-ilsqr-d2-thr <CHI_SEP_ILSQR_D2_THR>	χ-sep iLSQR: the streaking cone is where |D| is below this (default: 0.11)
--chi-sep-ilsqr-prc-w-i-min <CHI_SEP_ILSQR_PRC_W_I_MIN>	χ-sep iLSQR: percentile of the field’s Laplacian at which the LSQR weight is 1 (default: 60)
--chi-sep-ilsqr-prc-w-i-max <CHI_SEP_ILSQR_PRC_W_I_MAX>	χ-sep iLSQR: percentile of the field’s Laplacian at which the LSQR weight is 0 (default: 99.8)
--chi-sep-ilsqr-prc-w-g-min <CHI_SEP_ILSQR_PRC_W_G_MIN>	χ-sep iLSQR: percentile of each gradient’s magnitude at which the streaking-removal weight is 1 (default: 50)
--chi-sep-ilsqr-prc-w-g-max <CHI_SEP_ILSQR_PRC_W_G_MAX>	χ-sep iLSQR: percentile of each gradient’s magnitude at which the streaking-removal weight is 0 (default: 70)
--chi-sep-ilsqr-qsm-max-iter <CHI_SEP_ILSQR_QSM_MAX_ITER>	χ-sep iLSQR: max iterations of the iLSQR that reconstructs, from the local field, the QSM χ-separation is initialised with (default: 100; not used with –use-custom-qsm)
--chi-sep-ilsqr-qsm-tol <CHI_SEP_ILSQR_QSM_TOL>	χ-sep iLSQR: tolerance of that QSM’s iLSQR.
--chi-sep-ilsqr-qsm-pad-mm <MM>	χ-sep iLSQR: zero-padding of that QSM’s iLSQR, in mm per side (default: 12)
--chi-sep-ilsqr-qsm-precision <PRECISION>	χ-sep iLSQR: arithmetic of that QSM’s two LSQR solves: double, or single as the original implementation (default: double)
--chi-sep-medi-lambda <CHI_SEP_MEDI_LAMBDA>	χ-sep MEDI weight of the data terms against the L1 terms (default: 1)
--chi-sep-medi-lambda-csf <CHI_SEP_MEDI_LAMBDA_CSF>	χ-sep MEDI weight of the CSF-uniformity term (default: 1)
--chi-sep-medi-dr-pos <HZ_PER_PPM>	χ-sep MEDI paramagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-medi-dr-neg <HZ_PER_PPM>	χ-sep MEDI diamagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-medi-pad <CHI_SEP_MEDI_PAD>	χ-sep MEDI zero-padding at both ends of the third axis, in voxels (default: 5)
--chi-sep-medi-tol <CHI_SEP_MEDI_TOL>	χ-sep MEDI: stop when the relative change falls below this (default: 0.03)
--chi-sep-medi-max-iter <CHI_SEP_MEDI_MAX_ITER>	χ-sep MEDI maximum Gauss-Newton iterations (default: 30)
--chi-sep-medi-cg-tol <CHI_SEP_MEDI_CG_TOL>	χ-sep MEDI inner CG tolerance (default: 0.05)
--chi-sep-medi-cg-max-iter <CHI_SEP_MEDI_CG_MAX_ITER>	χ-sep MEDI inner CG max iterations (default: 100)
--chi-sep-medi-qsm-max-iter <CHI_SEP_MEDI_QSM_MAX_ITER>	χ-sep MEDI: max iterations of the iLSQR that reconstructs, from the local field, the QSM χ-separation is initialised with (default: 100; not used with –use-custom-qsm)
--chi-sep-medi-qsm-tol <CHI_SEP_MEDI_QSM_TOL>	χ-sep MEDI: tolerance of that QSM’s iLSQR.
--chi-sep-medi-qsm-pad-mm <MM>	χ-sep MEDI: zero-padding of that QSM’s iLSQR, in mm per side (default: 12)
--chi-sep-medi-qsm-precision <PRECISION>	χ-sep MEDI: arithmetic of that QSM’s two LSQR solves: double, or single as the original implementation (default: double)
--chi-sep-medi-csf-r2star <METHOD>	χ-sep MEDI: the R2* its CSF mask is extracted from: arlo (MEDI’s ARLO; the default) or nnls (the χ-separation toolbox’s r2star_nnls, which its script runs)
--wavesep-dr-pos <WAVESEP_DR_POS>	WaveSep paramagnetic relaxivity (Hz/ppm)
--wavesep-dr-neg <WAVESEP_DR_NEG>	WaveSep diamagnetic relaxivity (Hz/ppm)
--wavesep-alpha <WAVESEP_ALPHA>	WaveSep proximal-gradient step size.
--wavesep-lambda <WAVESEP_LAMBDA>	WaveSep wavelet L1 soft-threshold weight.
--wavesep-wavelet-order <WAVESEP_WAVELET_ORDER>	WaveSep Daubechies wavelet order.
--wavesep-max-iter <WAVESEP_MAX_ITER>	WaveSep ISTA max iterations.
--wavesep-tol <WAVESEP_TOL>	WaveSep relative-change stop tolerance.
--hc-chisep-dr-pos-3t <HC_CHISEP_DR_POS_3T>	HC-ChiSep paramagnetic relaxivity at 3T (Hz/ppm)
--hc-chisep-bin-hz <HC_CHISEP_BIN_HZ>	HC-ChiSep R2’ bin width for the anchored grid search (Hz)
--chi-sepnet-route <CHI_SEPNET_ROUTE>	χ-sepnet’s input route. toolbox (default), the authors’: the inputs resampled to 1 mm with B0 along z, the susceptibility channel computed by QSMnetME, R2’ measured or predicted from R2* by R2PRIMEnet, the maps resampled back. native: the network alone on the acquired grid, given this run’s QSM and R2’. Values: toolbox, native.
--chi-sepnet-patch <D> <H> <W>	χ-sepnet (--chi-sepnet-route native) sliding-window patch in voxels, D H W, each a multiple of 16 (default: 192 192 128, the authors’)
--chi-sepnet-frame-patch <D> <H> <W>	χ-sepnet (--chi-sepnet-route toolbox): run the networks as a sliding window of patches this size over the 1 mm frame, D H W, each a multiple of 16, so memory follows the patch. Omit to run the whole frame at once, as the authors do.
--susep-net-patch <D> <H> <W>	Run SUSEP-Net as a sliding window of patches this size in voxels, D H W, each a multiple of 8, so its memory follows the patch, not the volume. Omit to run the whole volume at once.
--motion-correction	Rigidly co-register every echo onto a reference echo (estimated on magnitude, applied to phase through the complex signal) before field mapping, and restrict the mask to voxels every echo covers. Off by default. Needs magnitude images.
--motion-reference <first|last|N>	Echo the others are aligned onto: first (default), last, or a 1-based echo number.
--no-motion-chain	Register every echo from the identity instead of seeding it from its neighbour’s solution.
--motion-chain-levels <MOTION_CHAIN_LEVELS>	Pyramid levels for a seeded (chained) registration (default 3)
--motion-min-ncc <MOTION_MIN_NCC>	NCC below which an echo’s registration is reported as suspect (default 0.3, set for echoes; ~0.9 suits repeats)
--registration-levels <REGISTRATION_LEVELS>	Registration pyramid levels: the capture range; do not lower (default 5)
--registration-max-samples <REGISTRATION_MAX_SAMPLES>	Registration: voxels sampled per metric evaluation, per level (default 200000)
--registration-max-iterations <REGISTRATION_MAX_ITERATIONS>	Registration: Powell cycles per level (default 20)
--registration-max-evaluations <REGISTRATION_MAX_EVALUATIONS>	Registration: metric evaluations across all levels, a hard stop (default 5000)
--registration-tolerance <REGISTRATION_TOLERANCE>	Registration: relative cost improvement per cycle that counts as converged (default 1e-5)
--registration-rotation-step-deg <REGISTRATION_ROTATION_STEP_DEG>	Registration: initial rotation step at the finest level, degrees (default 3)
--registration-translation-step-mm <REGISTRATION_TRANSLATION_STEP_MM>	Registration: initial translation step at the finest level, mm (default 3)
--registration-min-overlap <REGISTRATION_MIN_OVERLAP>	Registration: smallest fraction of sampled voxels that must overlap (default 0.25)
--vsharp-deconvolve <BOOL>	Undo V-SHARP’s SMV high-pass with a truncated-SVD deconvolution (default: false)
--vsharp-threshold <VSHARP_THRESHOLD>	V-SHARP deconvolution (TSVD) threshold; only used with –vsharp-deconvolve true.
--vsharp-max-radius <VSHARP_MAX_RADIUS>	V-SHARP max kernel radius in mm.
--vsharp-min-radius <VSHARP_MIN_RADIUS>	V-SHARP min kernel radius in mm.
--vsharp-radius-step <MM>	V-SHARP step between kernel radii in mm (default: 1); the radii run from –vsharp-max-radius down to –vsharp-min-radius.
--vsharp-grid <VSHARP_GRID>	Grid V-SHARP estimates the background on: isotropic (default; spacing from –vsharp-grid-spacing), native (the acquired grid), or finest (isotropic at the smallest voxel dimension, e.g. 0.8 mm for 0.8x0.8x3 mm; no axis is coarsened). The local field keeps native resolution. Values: native, finest, isotropic.
--vsharp-grid-spacing <MM>	V-SHARP isotropic working-grid spacing in mm (default 1); used with –vsharp-grid isotropic.
--pdf-tol <PDF_TOL>	PDF tolerance, relative to the noise level.
--pdf-n-cg <PDF_N_CG>	PDF maximum conjugate-gradient iterations (default: 30)
--pdf-space <SPACE>	PDF dipole kernel: image (the FFT of the dipole sampled on the voxel grid) or kspace (the analytic k-space kernel) (default: image)
--pdf-n-pad <VOXELS>	PDF zero-padding around the mask’s bounding box, in voxels; 0 runs on the full volume (default: 40)
--lbv-tol <LBV_TOL>	LBV tolerance: stop the coarsest grid’s relaxation when its summed absolute update falls below this, in radians over the echo spacing (default: 0.01)
--lbv-depth <LEVEL>	LBV coarsest multigrid level, at most 8 (default: log2 of the smallest dimension minus 3)
--lbv-peel <LAYERS>	LBV boundary layers removed from the mask beyond the first (default: 0)
--lbv-n1 <LBV_N1>	LBV relaxations on each grid before the coarser grid (default: 30)
--lbv-n2 <LBV_N2>	LBV relaxations on each grid after the coarser grid (default: 100)
--lbv-n3 <LBV_N3>	LBV relaxations on the finest grid at the end (default: 100)
--lbv-precision <PRECISION>	LBV arithmetic: double, or single as the original implementation (default: double)
--ismv-variant <VARIANT>	iSMV implementation: wen2014 (the authors’ paper and code) or qsmjl (QSM.jl’s) (default: wen2014)
--ismv-tol <ISMV_TOL>	iSMV tolerance: the relative change between iterates it stops at (qsmjl: the residual relative to the first) (default: 5e-5)
--ismv-max-iter <ISMV_MAX_ITER>	iSMV max iterations (default: no limit)
--ismv-radius <MM>	iSMV kernel radius in mm, or auto for 6 × the smallest voxel dimension (qsmjl: 2 × the largest) (default: auto)
--sharp-threshold <SHARP_THRESHOLD>	SHARP threshold.
--sharp-radius <SHARP_RADIUS>	SHARP kernel radius in mm.
--resharp-radius <RESHARP_RADIUS>	RESHARP SMV kernel radius in mm.
--resharp-tik-reg <RESHARP_TIK_REG>	RESHARP Tikhonov regularization parameter.
--resharp-tol <RESHARP_TOL>	RESHARP conjugate gradients squared (CGS) relative tolerance.
--resharp-max-iter <RESHARP_MAX_ITER>	RESHARP maximum CGS iterations.
--harperella-radius <HARPERELLA_RADIUS>	HARPERELLA SMV kernel radius in mm.
--harperella-max-iter <HARPERELLA_MAX_ITER>	HARPERELLA maximum iterations.
--harperella-tol <HARPERELLA_TOL>	HARPERELLA convergence tolerance.
--harperella-pad <X> <Y> <Z>	HARPERELLA zero-padding in voxels per side, X Y Z (default: 0 0 0)
--harperella-rim <RIM>	HARPERELLA boundary rim: whole (zero every Laplacian value in the rim; default) or sti (zero the 20 % largest, as STI Suite does)
--iharperella-max-iter <IHARPERELLA_MAX_ITER>	iHARPERELLA maximum iterations.
--iharperella-tol <IHARPERELLA_TOL>	iHARPERELLA convergence tolerance.
--iharperella-pad <X> <Y> <Z>	iHARPERELLA zero-padding in voxels per side, X Y Z (default: 0 0 0)
--iharperella-rim <RIM>	iHARPERELLA boundary rim: whole (zero every Laplacian value in the rim) or sti (zero the 20 % largest, as STI Suite does; default)
--no-romeo-phase-coherence	ROMEO: disable phase coherence weights (default: on)
--no-romeo-phase-gradient-coherence	ROMEO: disable phase gradient coherence weights (default: on)
--no-romeo-phase-linearity	ROMEO: disable phase linearity weights (default: on; the romeo command line’s romeo3 weights)
--romeo-phase-linearity	ROMEO: enable phase linearity weights (default: on; for a –pipeline that turns them off)
--no-romeo-mag-coherence	ROMEO: disable magnitude coherence weights (default: on)
--romeo-mag-weight	ROMEO: enable magnitude weighting (default: off)
--no-romeo-mag-weight	ROMEO: disable magnitude weighting (overrides –romeo-mag-weight; kept for compatibility)
--romeo-mag-weight2	ROMEO: enable second magnitude weighting (flow-artifact penalty; default: off)
--romeo-bestpath	ROMEO: use Best-path (Abdul-Rahman) weights instead of ROMEO weights (default: off)
--romeo-template <ROMEO_TEMPLATE>	ROMEO: template echo index (1-indexed, only for template mode)
--romeo-individual	ROMEO: use individual per-echo unwrapping (default)
--no-romeo-individual	ROMEO: use template-based temporal unwrapping (disables individual mode)
--romeo-correct-global	ROMEO: correct global 2π offsets, of each echo and between echoes (default: on)
--no-romeo-correct-global	ROMEO: do not correct global 2π offsets.
--romeo-temporal-uncertain-unwrapping <ROMEO_TEMPORAL_UNCERTAIN_UNWRAPPING>	ROMEO: temporal uncertain-unwrapping quality threshold [0,1] (0 disables; default: 0.5)
--romeo-max-seeds <ROMEO_MAX_SEEDS>	ROMEO: maximum number of seed regions (default: 1)
--romeo-merge-regions	ROMEO: merge neighboring regions after unwrapping (default: off)
--romeo-correct-regions	ROMEO: correct each region’s median to nearest 0 (default: off)
--romeo-wrap-addition <ROMEO_WRAP_ADDITION>	ROMEO: additional phase tolerance beyond π for neighbor differences [0,π] (default: 0.0)
--prelude-labelling <PRELUDE_LABELLING>	PRELUDE: how the initial regions are labelled (default: auto) Values: auto, slices, volume.
--prelude-partitions <N>	PRELUDE: number of phase partitions the initial regions are formed from (default: 8)
--swi-unwrapping <SWI_UNWRAPPING>	SWI phase unwrapping before the echoes are combined [default: laplacian]. Values: laplacian, laplacian-slice, romeo.
--swi-hp-sigma <SWI_HP_SIGMA>	SWI high-pass filter sigma (3 values, in voxels)
--swi-scaling <SWI_SCALING>	SWI phase-mask scaling. Values: tanh, negative-tanh, positive, negative, triangular.
--swi-strength <SWI_STRENGTH>	SWI phase scaling strength.
--swi-mip-window <SWI_MIP_WINDOW>	SWI MIP window size in slices.
--smwi-threshold <SMWI_THRESHOLD>	SMWI susceptibility threshold in ppm — the |χ| at which the weighting mask reaches zero. The 1 ppm default follows SEPIA and targets strong sources; brain tissue wants far less.
--smwi-power <SMWI_POWER>	SMWI mask power (contrast strength)
--smwi-mip-window <SMWI_MIP_WINDOW>	SMWI mIP window size in slices.
--synthseg-version <VERSION>	SynthSeg weights generation: v1 (32 labels, ships with SynthSeg) or v2 (33 labels, adds CSF; weights not published yet, so refused unless synthseg-2.0.onnx is in $QSM_MODEL_DIR)
--synthseg-crop <SYNTHSEG_CROP>	Cubic centre-crop in voxels at 1 mm before inference (rounded up to a multiple of 32). Omit to process the whole volume, as the SynthSeg CLI does.
--synthseg-patch <VOXELS>	Run SynthSeg as a sliding window of half-overlapping cubes this many voxels on a side at 1 mm (at least 64, rounded up to a multiple of 32), so its memory follows the patch, not the volume. Omit to run the whole volume at once, as the SynthSeg CLI does.
--no-synthseg-flip-averaging	Skip the left-right flipped second pass. Halves inference time, as SynthSeg’s –fast does.
--no-synthseg-topology-cleanup	Skip resetting each topological class to its largest connected component.
--synthseg-sigma <SYNTHSEG_SIGMA>	Gaussian blur on the posteriors, in voxels (0 disables it)
--use-custom-dseg <TOOL>	Prefer a bring-your-own segmentation from <bids>/derivatives/<TOOL>/ when present. Bare flag = first matching tool alphabetically. Always falls back to running SynthSeg.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--inhomogeneity-method <INHOMOGENEITY_METHOD>	Inhomogeneity correction method: makehomogeneous (default; –homogeneity-*) or n4 (N4 bias field correction, as ANTs’ N4BiasFieldCorrection; –n4-*) Values: makehomogeneous, n4.
--homogeneity-sigma-mm <HOMOGENEITY_SIGMA_MM>	Inhomogeneity correction smoothing sigma in mm (makehomogeneous)
--homogeneity-nbox <HOMOGENEITY_NBOX>	makehomogeneous: boxes per dimension for the tissue segmentation the receive sensitivity is estimated from.
--n4-shrink-factor <N4_SHRINK_FACTOR>	N4: subsample the image by this factor before fitting the bias field (default 4)
--n4-max-iterations <ITERATIONS>	N4: maximum iterations at each fitting level, one number a level (default 50 50 50 50); the B-spline mesh doubles from one level to the next.
--n4-convergence-threshold <N4_CONVERGENCE_THRESHOLD>	N4: stop a level once the change in the bias field (its coefficient of variation) falls to this (default 0: every iteration runs)
--n4-mesh <X> <Y> <Z>	N4: B-spline mesh elements per axis at the first level, X Y Z (default 1 1 1)
--n4-spline-order <N4_SPLINE_ORDER>	N4: order of the B-spline bias field, 1 to 3 (default 3)
--n4-bias-field-fwhm <N4_BIAS_FIELD_FWHM>	N4: FWHM of the Gaussian modelling the bias field’s blurring of the log-intensity histogram (default 0.15)
--n4-wiener-noise <N4_WIENER_NOISE>	N4: noise term of the Wiener deconvolution of the histogram (default 0.01)
--n4-histogram-bins <N4_HISTOGRAM_BINS>	N4: bins of the log-intensity histogram (default 200)
--n4-geometry <GEOMETRY>	N4: the image geometry it is given: scanner (the scan’s header, as ANTs reads the acquired NIfTI; the default) or make-nii (the axis-aligned header MATLAB’s make_nii writes, as QSMART runs N4). N4’s shrunk grid falls where the header puts it.
--linear-fit-estimate-offset <LINEAR_FIT_ESTIMATE_OFFSET>	Linear and robust fit: estimate a phase offset term (default: true). The robust fit needs 4 echoes with it, 3 without.
--robust-fit-loss <ROBUST_FIT_LOSS>	Motion-robust fit (--b0-estimation robust-fit): loss turning residuals into weights. tukey (default) rejects echoes past the tuning radius; huber caps but never rejects. Values: tukey, huber.
--robust-fit-tuning <ROBUST_FIT_TUNING>	Motion-robust fit: tuning radius in robust-scale units (default: the loss’s conventional value, Tukey 4.685 / Huber 1.345). Smaller is more aggressive.
--robust-fit-iterations <ROBUST_FIT_ITERATIONS>	Motion-robust fit: IRLS passes (default 3; 0 = report only, the field is the linear fit’s)
--robust-fit-downweight-level <ROBUST_FIT_DOWNWEIGHT_LEVEL>	Motion-robust fit (reporting only): a robust weight below this counts as down-weighted (default 0.75; in (0, 1])
--robust-fit-flag-ratio <ROBUST_FIT_FLAG_RATIO>	Motion-robust fit (reporting only): flag an echo whose outlier score exceeds this many times the best echo’s (default 3; at least 1)
--complex-fit-reltol <COMPLEX_FIT_RELTOL>	Complex fit (--b0-estimation complex-fit): stop once the update is this fraction of the first estimate, over the whole volume (default 1e-4)
--complex-fit-max-iter <COMPLEX_FIT_MAX_ITER>	Complex fit: Gauss-Newton iterations at most (default 30)
--no-qsm	Skip QSM processing (only run supplementary outputs like SWI, T2*, R2*)
--do-swi	Also compute SWI.
--swi-filtered-phase	Also write CLEAR-SWI’s high-pass filtered phase, the phase the SWI is weighted by (_desc-swi_phase.nii): Hz for several echoes, radians for one, 0 outside CLEAR-SWI’s head mask. Implies –do-swi.
--no-magnitude	Do not publish the combined magnitude (part-mag_T2starw). It is still computed, and kept with the run’s intermediates, when a step reads it (SMWI, segmentation, a magnitude-weighted inversion)
--do-segmentation	Also segment the brain from the GRE magnitude with SynthSeg, writing a FreeSurfer-labelled _dseg.nii and its BIDS lookup table. Needs no T1w and no registration. Needs a deep-learning build. doi:10.1016/j.media.2023.102789.
--do-analysis	Also summarise every quantitative map (χ, χ+/χ−, T2*, R2*, R2, R2’) per segmented structure into a TSV: mean, SD, median, min, max and the 5th/95th percentiles, one row per map and structure, plus a dataset-level table gathering all runs. Summarises whatever the run made — it implies –do-segmentation, and QSM only if no other map was asked for.
--do-smwi	Also compute SMWI (susceptibility map-weighted imaging): weights the magnitude by a mask built from the susceptibility map rather than from high-pass filtered phase, so the contrast stays where the source is. Needs χ, so it implies QSM. doi:10.1002/mrm.24920.
--do-t2starmap	Compute T2* relaxation map from multi-echo magnitude data.
--do-r2starmap	Compute R2* decay rate map from multi-echo magnitude data.
--do-r2map	Compute R2 map from a multi-echo spin-echo (MESE) acquisition (EPG)
--do-r2primemap	Compute R2’ map (= R2* − R2; needs GRE magnitude + a MESE acquisition)
--do-chisep	Compute chi-separation (paramagnetic/diamagnetic susceptibility maps)
--chisep <CHI_SEPARATION_ALGORITHM>	Chi-separation method (default: r2star-qsm) Values: r2star-qsm, decompose, chi-sep-ilsqr, chi-sep-medi, wavesep, hc-chisep, susep-net, chi-sepnet.
--use-custom-qsm <TOOL>	Use a bring-your-own QSM (Chimap) from <bids>/derivatives/<TOOL>/ for chi-separation.
--use-custom-r2 <TOOL>	Use a bring-your-own R2 map from <bids>/derivatives/<TOOL>/ for chi-separation.
--use-custom-r2prime <TOOL>	Use a bring-your-own R2’ map from <bids>/derivatives/<TOOL>/ for chi-separation.
--r2prime-strategy <R2PRIME_STRATEGY>	How to obtain R2’ when no custom map is supplied. auto (default) measures it as R2* - R2 from a MESE acquisition when one is present and predicts it from R2* with R2PRIMEnet when it is not; mese only ever measures it, producing nothing without spin-echo data; r2primenet always predicts it. A measurement and an estimate are not the same thing — the generated methods text says which route was configured. Values: auto, mese, r2primenet.
--r2primenet-ignore-field-mismatch	Run R2PRIMEnet even when the acquisition is not at 3 T, the only field strength its weights were trained at; the predicted R2’ is then uncalibrated. Without it, an explicit --r2prime-strategy r2primenet refuses a non-3 T run and auto skips R2’ with a warning.
--r2primenet-patch <D> <H> <W>	R2PRIMEnet sliding-window patch in voxels, D H W, each a multiple of 16 (default: 192 192 128, the authors’)
--export-dicom	Also export final maps as DICOM series into each subject’s extra_files/ folder.
--inhomogeneity-correction	Apply inhomogeneity correction to magnitude before masking.
--no-inhomogeneity-correction	Disable inhomogeneity correction.
--axial-resampling	Resample oblique acquisitions onto an axial grid when their obliquity exceeds –obliquity-threshold. Off by default (methods that only work on axial data resample regardless)
--no-axial-resampling	Do not resample oblique acquisitions to axial (overrides a config’s axial_resampling)
--obliquity-threshold <OBLIQUITY_THRESHOLD>	Obliquity in degrees above which –axial-resampling resamples (default: 10)
--crop-to-mask	Reconstruct inside a box around the mask instead of the whole field of view. Cuts the FFT-bound stages roughly in proportion to the voxels removed, but those stages are a small share of a full run, so the end-to-end gain is usually modest. Off by default. Outputs are written on the full grid either way; this only changes where the arithmetic happens. Padding to an FFT-friendly size happens regardless.
--ignore-slice-gap	Reconstruct even when a 2D acquisition’s sidecar SliceThickness is thinner than the slice pitch. Without it, QSM refuses such a run: the gap between slices is unsampled tissue, and the FFT kernels assume a contiguous grid, so the map would be wrong.
--fft-padding	Pad the grid outward to an FFT-friendly size before the FFT-bound stages. Much faster on an awkward grid, and discards nothing, but it changes where k-space is sampled and so shifts the reconstruction slightly. Off by default.
--tukey-window <TUKEY_WINDOW>	Apply a k-space Tukey window with this taper ratio (0 to 1) to each echo’s complex image before every other step, as the χ-separation toolbox does: 0.4, and 0 for Philips data. Default 0 (none). Needs magnitude images.
--crop-margin-mm <CROP_MARGIN_MM>	Margin in mm around the mask when –crop-to-mask is set (default 32). Below ~32 mm the reconstruction starts to differ from the uncropped one; it must also clear the largest SMV kernel in use, and V-SHARP defaults to 12 mm.
--output-space <OUTPUT_SPACE>	Grid to write derivatives on when the run was resampled to axial. acquired (default) puts them back where the data came from, so transforms you already hold still apply; working leaves them on the resampled grid and avoids a second interpolation. Values: acquired, working.
--mask-preset <MASK_PRESET>	Mask preset: robust-threshold, bet, hd-bet (HD-BET deep-learning brain extraction + signal-gated erosion; needs a deep-learning build), bet-and-phase (BET on the magnitude AND a thresholded phase-quality map, then hole-filled and eroded), rs2-net (RS2-Net deep-learning rodent brain extraction), or mindgrab (MindGrab deep-learning brain extraction for any contrast; needs a deep-learning build) Values: robust-threshold, bet, hd-bet, bet-and-phase, rs2-net, mindgrab.
--use-custom-masks <USE_CUSTOM_MASKS>	Prefer a bring-your-own mask from BIDS derivatives when present (else compute it). Bare flag = first matching derivatives tool (alphabetical); give a name (e.g. bet) to restrict to derivatives/<tool>/sub-*/anat/*_mask.nii*. Always falls back.
--mask <MASK_SECTIONS_CLI>	Define a mask section (repeatable; multiple sections are combined with –mask-combine). Format: <input>,<generator>,<refinement1>,<refinement2>,… Example: phase-quality,threshold:otsu,dilate:2,fill-holes:0,erode:2 Example: magnitude,bet:0.5,erode:2 Example: magnitude-first,bet:0.2,min-intensity:100,majority,close:2 (QSMART’s mask) Example: magnitude,hd-bet,signal-erode (HD-BET: hd-bet[:full|window|XxYxZ|low-memory][:tta][:step=F]; signal-erode: signal-erode[:threshold[:depth_cap[:global_erosions[:bias_sigma[:min_component]]]]])
--mask-combine <MASK_COMBINE>	How to combine multiple –mask sections: or (union, the default) or and (intersection — a voxel survives only where every section keeps it) Values: or, and.
--mask-refine <MASK_REFINEMENTS_CLI>	Refinement applied to the combined mask, after –mask-combine (repeatable). Takes the same refinement ops as –mask, e.g. erode:2 or fill-holes:0; generators (threshold/bet/hd-bet) belong in a –mask section instead. Example: –mask-combine and –mask-refine fill-holes:0 –mask-refine erode:1.
--two-pass	Two-pass artefact reduction: reconstruct a second time against a mask whose holes are left unfilled, and keep that reconstruction wherever it is defined. Strong susceptibility sources (air-tissue interfaces, haemorrhage, implants) make the phase around them unrecoverable; leaving those voxels out keeps their streaking from spreading, while the ordinary mask still supplies values inside the holes. Roughly doubles background removal and inversion time, and is not free of downsides — where the holes were not artefactual it can be worse than a single pass. Based on doi:10.1002/mrm.29048 (Stewart et al.). Off by default. The conventional single-pass map is kept alongside as _desc-singlepass_Chimap.nii.
--two-pass-mask <TWO_PASS_SECTIONS_CLI>	Define the reliable-pass mask used by –two-pass (repeatable; same format as –mask). Defaults to phase-quality,threshold:otsu — an unfilled threshold on the phase-quality map. Its holes are the point, so refinements that close them (fill-holes, close, gaussian) defeat the method. Folded with –two-pass-mask-combine, then –mask-refine like the main mask, and always intersected with it.
--two-pass-mask-combine <TWO_PASS_MASK_COMBINE>	How to combine multiple –two-pass-mask sections: or (union) or and (intersection). Defaults to the main mask’s –mask-combine. Turns on –two-pass. Values: or, and.
--dismiss-recommendation <RULE=VALUE>	Stop offering a recommendation for the setting’s current value: RULE=VALUE, the rule’s id and the value it was dismissed for, as qsmxt run --dry lists them (e.g. field-mapping.phase-offsets=false). Repeatable. A change to the setting brings the recommendation back. A required one cannot be dismissed, and a reference-pipeline difference has nothing to dismiss.

qsmxt aggregate

Run the dataset-level steps over runs already processed with qsmxt run --no-dataset-steps (multi-orientation, summary tables, provenance, clean-up)

qsmxt aggregate [OPTIONS] <BIDS_DIR> [OUTPUT_DIR]

Argument	Description
<BIDS_DIR>	Input BIDS directory. Required.
<OUTPUT_DIR>	Output directory (defaults to bids_dir; outputs go into <dir>/derivatives/qsmxt/)
--config <CONFIG>	Pipeline configuration file (TOML): the one the runs were processed with.
--include <INCLUDE>	Include only runs matching these glob patterns (e.g. “sub-1*” “*ses-pre*”)
--exclude <EXCLUDE>	Exclude runs matching these glob patterns (e.g. “*mygrea*”)
--num-echoes <NUM_ECHOES>	Limit number of echoes to process.
--pipeline <PIPELINE>	Start from a pipeline: qsmxt-defaults, or an original toolbox’s (medi-toolbox, sti-suite-ilsqr, sti-suite-star, qsmart, heidi, chi-sep-ilsqr, chi-sep-medi, chi-sepnet, decompose). The stages it covers (masking, field mapping, background removal, inversion, referencing, χ-separation) go back to QSMxT’s defaults and then take the toolbox’s settings; the other flags given then change any setting. Values: qsmxt-defaults, medi-toolbox, sti-suite-ilsqr, sti-suite-star, qsmart, heidi, chi-sep-ilsqr, chi-sep-medi, chi-sepnet, decompose.
--qsm-algorithm <QSM_ALGORITHM>	QSM dipole inversion algorithm (default: ilsqr) Values: rts, tv, tkd, tsvd, tgv, tikhonov, nltv, medi, tfi, ilsqr, star, lsqr, heidi, qsmart, ndi, fansi, fansi-tgv, l1qsm, whqsm, hdqsm, amp-pe, xqsm, qsmnet, qsmnet-plus, qsmnetme, autoqsm, qsmgan, ir2qsm, lpcnn, modl-qsm, nextqsm, iqsm, iqsm-plus.
--heidi	HEIDI streaking reduction after the inversion: keep the map’s well-conditioned k-space and refill the dipole cone under a homogeneity prior (--heidi-* sets it up). With LSQR this is HEIDI as published (--qsm-algorithm heidi); with two-pass it runs once, on the combined map.
--orientation-group <PATTERN>	Group runs into multi-orientation sets and reconstruct each with COSMOS or STI. Takes an entity name (acq, run, rec, inv), a glob over the run key (*acq-dir*), re: and a regex with one capture group, or runs: and the groups’ run keys (, between runs, ; between groups, each group’s reference first). Orientations are co-registered if they need it.
--orientation-reference <PATTERN>	The orientation each group is registered onto, and whose grid and segmentation the group map uses: a glob over the run key (*acq-Neutral*) or a word in it. Must match one run per group. Default: the first run in name order.
--no-orientation-registration	Do not co-register the orientations in a group — use this when they were already registered externally and a second interpolation is not wanted. A group that still needs registration is then refused instead of reconstructed.
--multi-orientation-algorithm <MULTI_ORIENTATION_ALGORITHM>	Which reconstruction to run over each orientation group. Values: cosmos, sti.
--multi-orientation-lambda <MULTI_ORIENTATION_LAMBDA>	Regularization for the multi-orientation reconstruction (COSMOS k-space L2, or STI ridge)
--multi-orientation-force	Reconstruct an orientation group even when its B0 directions look degenerate.
--cosmos-magnitude-weighting	COSMOS: weight each orientation by its combined magnitude (conjugate-gradient solve)
--cosmos-tol <COSMOS_TOL>	Magnitude-weighted COSMOS: conjugate-gradient tolerance (needs –cosmos-magnitude-weighting)
--cosmos-max-iter <COSMOS_MAX_ITER>	Magnitude-weighted COSMOS: conjugate-gradient iteration cap (needs –cosmos-magnitude-weighting)
--unwrapping-algorithm <UNWRAPPING_ALGORITHM>	Unwrapping algorithm (default: romeo) Values: romeo, laplacian, prelude, medi-region-growing.
--bf-algorithm <BF_ALGORITHM>	Background field removal algorithm (default: vsharp) Values: vsharp, pdf, lbv, ismv, sharp, resharp, harperella, iharperella, bfrnet, iqfm.
--bf-erode-voxels <N>	Erode the brain mask by N voxels after field mapping, before background removal; the field map is not masked again (default: 0, none)
--bf-polish-mask	After background removal, keep the mask’s largest connected component with its holes filled and zero the local field outside it, as STI Suite’s QSM_GUI does (PolishMask)
--no-bf-polish-mask	Do not polish the background-removal mask (the default), after a --pipeline that does.
--masking-input <MASKING_INPUT>	Masking input image (overrides the input of the configured mask sections; combine with –mask-preset, e.g. –mask-preset robust-threshold –masking-input magnitude) Values: magnitude-first, magnitude, magnitude-last, phase-quality.
--phase-offset-removal <PHASE_OFFSET_REMOVAL>	Enable phase offset removal for multi-echo data (default: true)
--phase-offset-sigma-mm <PHASE_OFFSET_SIGMA_MM>	Phase offset removal smoothing sigma (3 values, in mm; default: 7 7 7, the ROMEO command line’s)
--coil-combination-sigma <COIL_COMBINATION_SIGMA>	MCPC-3D-S coil combination: smoothing sigma for the per-coil phase offsets (3 values, in voxels; default 10 10 5, as in MriResearchTools). Only used for uncombined multi-coil runs.
--bipolar-correction	Enable bipolar gradient correction (requires >= 3 echoes)
--b0-estimation <B0_ESTIMATION>	B0 field estimation method (default: weighted-avg) Values: weighted-avg, linear-fit, robust-fit, qsmart-echofit, complex-fit.
--qsmart-fit-threshold-mode <QSMART_FIT_THRESHOLD_MODE>	--b0-estimation qsmart-echofit: how the fit residual at which a voxel is left out of QSMART’s reconstruction is set: a percentile of the blurred residual over the volume (--qsmart-fit-percentile), or a fixed residual (--qsmart-fit-threshold; default, QSMART v1.0’s 40). Without it, giving one of those two selects its mode. Values: percentile, absolute.
--qsmart-fit-percentile <QSMART_FIT_PERCENTILE>	--b0-estimation qsmart-echofit: the percentile of the blurred fit residual, in (0, 100], above which a voxel is left out of QSMART’s reconstruction (default 99)
--qsmart-fit-threshold <QSMART_FIT_THRESHOLD>	--b0-estimation qsmart-echofit: a fixed blurred fit residual above which a voxel is left out of QSMART’s reconstruction, instead of a percentile (40 when the mode is absolute and none is given)
--b0-weight-type <B0_WEIGHT_TYPE>	B0 weighted averaging weight type. assumed-decay gives each echo’s phase a fixed weight TE·exp(−TE/T2*) computed from the echo times and an assumed T2* (--b0-weight-assumed-t2star), the same in every voxel: no T2* map, no magnitude (Wang et al. 2022, Nat Neurosci 25:818). phase-snr, which uses the measured magnitude in each voxel, is generally preferable when magnitude is available. Applies to Laplacian unwrapping as well as ROMEO. Values: phase-snr, phase-var, average, tes, mag, assumed-decay.
--b0-weight-assumed-t2star <MS>	T2* in ms assumed by --b0-weight-type assumed-decay to compute its fixed echo weights (default 40; no T2* map is used)
--laplacian-solver <LAPLACIAN_SOLVER>	Poisson solver for --unwrapping-algorithm laplacian: dct (unweighted least squares, Neumann/DCT solve; default), fft (Schofield & Zhu sin/cos Laplacian, FFT solve on a zero-padded volume) or mirror (the same on the volume mirrored along each axis, as QSMART unwraps) Values: dct, fft, mirror.
--laplacian-fft-pad <VOXELS>	Zero-padding in voxels per side for --laplacian-solver fft: one value for every axis, or three (x y z). Default 64.
--laplacian-unmasked	--unwrapping-algorithm laplacian: unwrap the raw phase over the whole volume, instead of the phase set to zero outside the brain mask, as QSMART does. With --b0-estimation qsmart-echofit the field map is then kept over the whole volume.
--no-laplacian-unmasked	--unwrapping-algorithm laplacian: unwrap the phase set to zero outside the brain mask (the default), after a --pipeline that unwraps the whole volume.
--bet-fractional-intensity <BET_FRACTIONAL_INTENSITY>	BET fractional intensity (0.0-1.0; smaller = larger brain). Applies to every BET step in the masking recipe, overriding a bet:<f> in a –mask spec.
--bet-smoothness <BET_SMOOTHNESS>	BET surface smoothness factor. Applies to every BET step in the masking recipe.
--bet-gradient-threshold <BET_GRADIENT_THRESHOLD>	BET gradient threshold (-1 to 1). Applies to every BET step in the masking recipe.
--bet-iterations <BET_ITERATIONS>	BET surface evolution iterations. Applies to every BET step in the masking recipe.
--bet-subdivisions <BET_SUBDIVISIONS>	BET icosphere subdivision level. Applies to every BET step in the masking recipe.
--bet-voxel-scale <FACTOR>	Factor applied to the voxel sizes before BET runs. BET’s surface model is tuned to a human brain in millimetres, so preclinical data has to be presented at human scale: 10 is the usual setting for mouse. 1 (the default) is the acquisition’s own geometry. Applies to every BET step in the masking recipe.
--bet-orientation <ORIENTATION>	Which way round BET sees the left-right axis: fsl (the default; reversed for a neurologically stored image, as FSL’s bet reads a NIfTI file) or as-stored (the voxels in the order they are stored, as MEDI’s BET.m runs bet2). Applies to every BET step in the masking recipe.
--qsm-reference <SPEC>	What the susceptibility map’s zero is pinned to: mean (the mean over the voxels the map is defined on — the brain mask less the rim background-field removal eroded; the default), none, or a SynthSeg region — a structure with both sides merged (thalamus), one side (left-thalamus), the ventricles composite, a raw FreeSurfer id (4), or a comma-separated list of any of those. A region implies –do-segmentation. An unrecognised name lists the valid ones.
--qsm-per-echo <BOOL>	Reconstruct a QSM from each echo on its own (each echo’s phase unwrapped and divided by its echo time, then background removal, dipole inversion and referencing per echo) and publish each echo’s map (desc-echo<n>_Chimap) and their mean as the QSM (Chimap), instead of one QSM from the echo-combined field map. The echo combination (–b0-estimation), phase-offset removal and bipolar correction do not run. DECOMPOSE takes the echoes’ maps. Multi-echo only; not with TGV, QSMART, iQSM, iQFM, two-pass, or χ-separation methods that fit the local field.
--rts-delta <RTS_DELTA>	RTS delta parameter.
--rts-mu <RTS_MU>	RTS mu parameter.
--rts-tol <RTS_TOL>	RTS tolerance.
--rts-rho <RTS_RHO>	RTS rho (ADMM penalty)
--rts-max-iter <RTS_MAX_ITER>	RTS max iterations.
--rts-lsmr-iter <RTS_LSMR_ITER>	RTS LSMR iterations.
--rts-tau <RTS_TAU>	RTS: rho is raised when the primal residual falls by less than this factor (tau)
--rts-gamma <RTS_GAMMA>	RTS: the factor rho is raised by (gamma)
--rts-pad <X> <Y> <Z>	RTS zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tv-lambda <TV_LAMBDA>	TV lambda parameter.
--tv-rho <RHO>	TV rho (ADMM penalty), or auto for 100 × lambda (default: auto)
--tv-tol <TV_TOL>	TV tolerance.
--tv-max-iter <TV_MAX_ITER>	TV max iterations.
--tv-pad <X> <Y> <Z>	TV zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tkd-threshold <TKD_THRESHOLD>	TKD threshold.
--tkd-pad <X> <Y> <Z>	TKD zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tsvd-threshold <TSVD_THRESHOLD>	TSVD threshold.
--tsvd-pad <X> <Y> <Z>	TSVD zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--tgv-iterations <TGV_ITERATIONS>	TGV iterations (default: from the voxel size, as the reference does)
--tgv-erosions <TGV_EROSIONS>	TGV erosions.
--tgv-alpha1 <TGV_ALPHA1>	TGV alpha1 (first-order weight), in radians at the echo time (–tgv-te) and field strength (the reference’s units)
--tgv-alpha0 <TGV_ALPHA0>	TGV alpha0 (second-order weight), in radians at the echo time (–tgv-te) and field strength (the reference’s units)
--tgv-step-size <TGV_STEP_SIZE>	TGV primal step size multiplier.
--tgv-te <TE>	TGV: the echo time a field from several echoes is converted to phase at, and its weights are in radians at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first). A single echo is at its own.
--tikhonov-lambda <TIKHONOV_LAMBDA>	Tikhonov lambda.
--tikhonov-pad <X> <Y> <Z>	Tikhonov zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--nltv-lambda <NLTV_LAMBDA>	NLTV lambda.
--nltv-rho <RHO>	NLTV rho (ADMM penalty), or auto for 100 × lambda (default: auto)
--nltv-mu <NLTV_MU>	NLTV mu (penalty parameter)
--nltv-tol <NLTV_TOL>	NLTV tolerance.
--nltv-max-iter <NLTV_MAX_ITER>	NLTV max iterations.
--nltv-tol-newton <NLTV_TOL_NEWTON>	NLTV convergence tolerance of the Newton iterations of the data-term subproblem.
--nltv-newton-iter <NLTV_NEWTON_ITER>	NLTV maximum Newton iterations of the data-term subproblem.
--nltv-te <TE>	NLTV: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--nltv-pad <X> <Y> <Z>	NLTV zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)
--medi-lambda <MEDI_LAMBDA>	MEDI lambda: the data-fidelity weight (larger fits the field more closely)
--medi-merit <MEDI_MERIT>	MEDI: enable MERIT weighting.
--medi-csf <MEDI_CSF>	MEDI+0 (Liu et al. 2018): zero-reference χ to the ventricular CSF, found from R2*, and keep it uniform there (default: true). Needs at least 3 echoes; with fewer, MEDI runs without it.
--medi-lambda-csf <MEDI_LAMBDA_CSF>	MEDI+0: weight of the CSF uniformity term (default: 100)
--medi-csf-r2s-threshold <MEDI_CSF_R2S_THRESHOLD>	MEDI+0: R2* in Hz below which a voxel is a CSF candidate (default: 5)
--medi-csf-erode <MEDI_CSF_ERODE>	MEDI+0: erode the mask by a 10 mm sphere before finding the CSF (default: true)
--medi-zeropad <X> <Y> <Z>	MEDI: voxels of zeros appended at the end of x, y and z before the inversion; the result is cropped back (default: 0 0 0)
--medi-smv	MEDI: high-pass the local field with a spherical mean value filter, removing residual background near the brain edge (MEDI-SMV; on by default)
--no-medi-smv	MEDI: no SMV high-pass of the local field.
--medi-smv-radius <MEDI_SMV_RADIUS>	MEDI SMV radius in mm (also mSMV’s)
--medi-smv-shrink-mask	MEDI-SMV: shrink the mask to where the SMV of the mask exceeds 0.999 (on by default)
--no-medi-smv-shrink-mask	MEDI-SMV: keep the full mask, high-passing the field over it.
--medi-msmv	MEDI-mSMV (Roberts et al. 2024): filter the local field by maximum spherical mean value before MEDI, which fits it with the SMV-filtered dipole kernel (it implies the SMV). Needs at least 3 echoes.
--medi-msmv-maxk <MEDI_MSMV_MAXK>	MEDI-mSMV: the boundary correction makes at most this many passes, less one.
--medi-msmv-vessel-radius <MEDI_MSMV_VESSEL_RADIUS>	MEDI-mSMV: largest vessel thickness of the vessel mask, in voxels (default: 15, or 8 × the largest voxel dimension if that is more)
--no-medi-msmv-prefilter	MEDI-mSMV: no SMV prefilter before mSMV. Only consistent for a field already high-passed by the same SMV; after PDF, V-SHARP or LBV it diverges (expert)
--medi-data-weighting <MEDI_DATA_WEIGHTING>	MEDI: data weighting mode (0=uniform, 1=SNR)
--medi-percentage <MEDI_PERCENTAGE>	MEDI edge percentage (0.0-1.0)
--medi-cg-tol <MEDI_CG_TOL>	MEDI CG tolerance.
--medi-cg-max-iter <MEDI_CG_MAX_ITER>	MEDI CG max iterations.
--medi-max-iter <MEDI_MAX_ITER>	MEDI max outer iterations.
--medi-tol <MEDI_TOL>	MEDI outer tolerance.
--medi-te <TE>	MEDI: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)
--tfi-lambda <TFI_LAMBDA>	TFI lambda: the data-fidelity weight (larger fits the field more closely)
--tfi-precond <TFI_PRECOND>	TFI preconditioner value (susceptibility scaling outside the mask)
--tfi-merit <TFI_MERIT>	TFI: enable MERIT weighting (default: true)
--tfi-linear <TFI_LINEAR>	TFI: fit the field linearly, as the TFI paper does (default: true); false uses MEDI’s exponential data term.
--tfi-data-weighting <TFI_DATA_WEIGHTING>	TFI: data weighting mode (0=uniform, 1=SNR)
--tfi-percentage <TFI_PERCENTAGE>	TFI edge percentage (0.0-1.0)
--tfi-cg-tol <TFI_CG_TOL>	TFI CG tolerance.
--tfi-cg-max-iter <TFI_CG_MAX_ITER>	TFI CG max iterations.
--tfi-max-iter <TFI_MAX_ITER>	TFI max outer iterations.
--tfi-tol <TFI_TOL>	TFI outer tolerance.
--tfi-te <TE>	TFI: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)
--ilsqr-tol <ILSQR_TOL>	iLSQR tolerance of the streaking-artefact solve.
--ilsqr-max-iter <ILSQR_MAX_ITER>	iLSQR max iterations of each LSQR solve (default: 100)
--ilsqr-pad-mm <MM>	iLSQR zero-padding in mm per side.
--ilsqr-precision <PRECISION>	iLSQR arithmetic of the two LSQR solves: double, or single as the original implementation (default: double)
--star-pad <X> <Y> <Z>	STAR-QSM zero-padding in voxels per side, X Y Z (default: 8 8 20)
--star-te <TE>	Echo time STAR-QSM converts the field to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: sum)
--lsqr-residual-weighting <LSQR_RESIDUAL_WEIGHTING>	LSQR residual-field weight w — the only thing regularising the solve. Omit to use qsm-core’s field-strength-scaled default; 0 drops the residual term and streaks badly.
--no-lsqr-global-offset	Do not fit a global field offset alongside χ.
--lsqr-tol <LSQR_TOL>	LSQR convergence tolerance.
--lsqr-max-iter <LSQR_MAX_ITER>	LSQR max iterations.
--heidi-cone-threshold <HEIDI_CONE_THRESHOLD>	|D(k)| above which a coefficient is well-conditioned and kept from the seed map.
--heidi-gradient-threshold <HEIDI_GRADIENT_THRESHOLD>	Field-gradient threshold (ppm/mm); below it a direction reads as homogeneous.
--no-heidi-laplacian-correction	Do not open the homogeneity masks where the field’s Laplacian is large.
--heidi-laplacian-threshold <HEIDI_LAPLACIAN_THRESHOLD>	Field-Laplacian threshold (ppm per voxel²)
--heidi-gradient-mask-floor <HEIDI_GRADIENT_MASK_FLOOR>	TV weight kept where every direction reads as an edge (0 reproduces upstream’s shipped binarised masks)
--heidi-continuation-steps <HEIDI_CONTINUATION_STEPS>	NESTA continuation steps on μ.
--heidi-inner-iterations <HEIDI_INNER_ITERATIONS>	Accelerated-gradient iterations per continuation step.
--heidi-mu-min <HEIDI_MU_MIN>	Final (smallest) smoothing parameter μ.
--heidi-tol <HEIDI_TOL>	Relative objective change that ends an inner loop early (the default runs every iteration)
--no-heidi-denoise	Skip the anisotropic-diffusion smoothing of the field before its gradients are taken.
--heidi-denoise-iterations <HEIDI_DENOISE_ITERATIONS>	Anisotropic-diffusion iterations.
--heidi-denoise-time-step <HEIDI_DENOISE_TIME_STEP>	Anisotropic-diffusion time step.
--heidi-denoise-conductance <HEIDI_DENOISE_CONDUCTANCE>	Anisotropic-diffusion conductance.
--heidi-denoise-reference-itk	ITK 3.20’s float arithmetic for the denoising, as the authors’ code runs it (for comparison with it; the default computes its conductance scale in double precision)
--no-heidi-denoise-reference-itk	The denoising in double precision, after a pipeline that turns ITK’s float arithmetic on (the HEIDI pipeline)
--heidi-te <TE>	HEIDI: the echo time the field stands for the phase of, in the conversion to HEIDI’s units: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: last)
--qsmart-ilsqr-max-iter <QSMART_ILSQR_MAX_ITER>	QSMART iLSQR max iterations per stage. The iLSQR tolerance is –ilsqr-tol.
--qsmart-ilsqr-pad-mm <MM>	Zero-padding of QSMART’s inner iLSQR solves, in mm per side (default: 6)
--qsmart-vasc-sphere-radius <QSMART_VASC_SPHERE_RADIUS>	QSMART vasculature detection sphere radius, in –qsmart-vessel-units (default 8)
--qsmart-inversion <QSMART_INVERSION>	QSMART inner dipole inversion algorithm (default: ilsqr) Values: rts, tv, tkd, tsvd, tgv, tikhonov, nltv, medi, tfi, ilsqr, star, lsqr, heidi, qsmart, ndi, fansi, fansi-tgv, l1qsm, whqsm, hdqsm, amp-pe, xqsm, qsmnet, qsmnet-plus, qsmnetme, autoqsm, qsmgan, ir2qsm, lpcnn, modl-qsm, nextqsm, iqsm, iqsm-plus.
--qsmart-sdf-sigma1-stage1 <QSMART_SDF_SIGMA1_STAGE1>	QSMART SDF sigma1, stage 1 (voxels)
--qsmart-sdf-sigma2-stage1 <QSMART_SDF_SIGMA2_STAGE1>	QSMART SDF sigma2, stage 1 (voxels)
--qsmart-sdf-sigma1-stage2 <QSMART_SDF_SIGMA1_STAGE2>	QSMART SDF sigma1, stage 2 (voxels)
--qsmart-sdf-sigma2-stage2 <QSMART_SDF_SIGMA2_STAGE2>	QSMART SDF sigma2, stage 2 (voxels)
--qsmart-sdf-lower-lim <QSMART_SDF_LOWER_LIM>	QSMART SDF proximity lower limit.
--qsmart-sdf-curv-constant <QSMART_SDF_CURV_CONSTANT>	QSMART SDF curvature constant.
--qsmart-frangi-scale-min <QSMART_FRANGI_SCALE_MIN>	QSMART Frangi min vessel radius, in –qsmart-vessel-units (default 0.5)
--qsmart-frangi-scale-max <QSMART_FRANGI_SCALE_MAX>	QSMART Frangi max vessel radius, in –qsmart-vessel-units (default 6)
--qsmart-frangi-scale-ratio <QSMART_FRANGI_SCALE_RATIO>	QSMART Frangi scale step, in –qsmart-vessel-units (default 0.5)
--qsmart-frangi-c <QSMART_FRANGI_C>	QSMART Frangi C noise threshold.
--qsmart-vessel-units <QSMART_VESSEL_UNITS>	Units of QSMART’s vessel sizes (–qsmart-vasc-sphere-radius and the Frangi scales): voxels (default, as QSMART gives them) or mm (converted to voxels with the mean voxel size) Values: voxels, mm.
--qsmart-vessel-magnitude <QSMART_VESSEL_MAGNITUDE>	The magnitude QSMART detects vessels on: mean-n4 (the mean over echoes, N4 bias-corrected, as QSMART prepares it; default) or combined (the root sum of squares over echoes, inhomogeneity-corrected unless –no-inhomogeneity-correction) Values: mean-n4, combined.
--qsmart-offset-grid <QSMART_OFFSET_GRID>	The dipole grid of QSMART’s offset adjustment: fft (default) or qsmart-reference (adjust_offset.m’s own, half a step off the FFT grid on odd matrix sizes, as the QSMART pipeline runs it) Values: fft, qsmart-reference.
--ndi-tau <NDI_TAU>	NDI gradient-descent step size (tau)
--ndi-alpha <NDI_ALPHA>	NDI L2 regularization weight (alpha)
--ndi-max-iter <NDI_MAX_ITER>	NDI max iterations.
--ndi-auto-stop <NDI_AUTO_STOP>	NDI: stop automatically once noise starts to build up along the dipole cone (FANSI ndi_auto); –ndi-max-iter is then the maximum.
--ndi-te <TE>	NDI: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--ndi-weight-norm <NORM>	NDI magnitude-weight normaliser: global-max (the magnitude’s maximum, as FANSI; default) or mask-p999 (its 99.9th percentile inside the mask; more consistent in vivo)
--fansi-alpha1 <FANSI_ALPHA1>	FANSI first-order (TV/TGV) L1 penalty weight (alpha1)
--fansi-mu1 <MU1>	FANSI gradient-consistency ADMM weight (mu1), or auto for 100 × alpha1 (default: auto)
--fansi-mu2 <FANSI_MU2>	FANSI fidelity-consistency ADMM weight (mu2)
--fansi-alpha0 <ALPHA0>	FANSI second-order L1 penalty weight (alpha0, nlTGV only), or auto for 2 × alpha1 (default: auto)
--fansi-mu0 <MU0>	FANSI second-order consistency ADMM weight (mu0, nlTGV only), or auto for 2 × mu1 (default: auto)
--fansi-max-iter <FANSI_MAX_ITER>	FANSI outer ADMM iterations.
--fansi-tol-update <FANSI_TOL_UPDATE>	FANSI percent-update convergence tolerance.
--fansi-tol-delta <FANSI_TOL_DELTA>	FANSI inner Newton convergence tolerance.
--fansi-te <TE>	FANSI (nlTV and nlTGV): the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)
--l1qsm-alpha1 <L1QSM_ALPHA1>	L1-QSM gradient (TV) L1 penalty weight (alpha1)
--l1qsm-mu1 <MU1>	L1-QSM gradient-consistency ADMM weight (mu1), or auto for 100 × alpha1 (default: auto)
--l1qsm-mu2 <L1QSM_MU2>	L1-QSM fidelity-consistency ADMM weight (mu2)
--l1qsm-mu3 <L1QSM_MU3>	L1-QSM L1 proximal ADMM weight (mu3)
--l1qsm-lambda <L1QSM_LAMBDA>	L1-QSM L1 fidelity strength (lambda)
--l1qsm-max-iter <L1QSM_MAX_ITER>	L1-QSM outer ADMM iterations.
--l1qsm-tol-update <L1QSM_TOL_UPDATE>	L1-QSM percent-update convergence tolerance.
--l1qsm-tol-delta <L1QSM_TOL_DELTA>	L1-QSM inner Newton convergence tolerance.
--l1qsm-te <TE>	L1-QSM: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--whqsm-alpha1 <WHQSM_ALPHA1>	WH-QSM TV regularization weight (alpha1)
--whqsm-mu1 <WHQSM_MU1>	WH-QSM ADMM penalty for TV splitting (mu1)
--whqsm-mu2 <WHQSM_MU2>	WH-QSM ADMM penalty for data-fidelity splitting (mu2)
--whqsm-beta <WHQSM_BETA>	WH-QSM weak-harmonic ROI penalty (beta)
--whqsm-muh <WHQSM_MUH>	WH-QSM ADMM penalty for harmonic-field splitting (muh)
--whqsm-max-iter <WHQSM_MAX_ITER>	WH-QSM max outer iterations.
--whqsm-tol-update <WHQSM_TOL_UPDATE>	WH-QSM percent-update convergence tolerance.
--whqsm-tol-delta <WHQSM_TOL_DELTA>	WH-QSM inner Newton convergence tolerance.
--whqsm-weight <WEIGHT>	WH-QSM data weight: flat (1 over the whole volume, WH_nlTV.m’s default), flat-in-mask (1 inside the mask only), max-normalised (the magnitude over its maximum) or in-vivo (min(2 × magnitude / maximum, 1)); the last two need magnitude images (default: flat)
--whqsm-te <TE>	WH-QSM: the echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--hdqsm-alpha-l2 <HDQSM_ALPHA_L2>	HD-QSM L2-stage TV weight (alpha_l2)
--hdqsm-mu1-l2 <MU1_L2>	HD-QSM L2-stage gradient-consistency ADMM weight (mu1_l2), or auto for 100 × alpha_l2 (default: auto)
--hdqsm-mu2 <HDQSM_MU2>	HD-QSM fidelity consistency weight (mu2)
--hdqsm-max-iter-l1 <HDQSM_MAX_ITER_L1>	HD-QSM stage-1 (L1) iterations (max_iter_l1)
--hdqsm-max-iter-l2 <HDQSM_MAX_ITER_L2>	HD-QSM stage-2 (L2) iterations (max_iter_l2)
--hdqsm-tol-update <HDQSM_TOL_UPDATE>	HD-QSM stage-2 percent-update convergence tolerance.
--amp-pe-wave-order <AMP_PE_WAVE_ORDER>	AMP-PE Daubechies wavelet order (1=db1, 2=db2)
--amp-pe-nlevel <AMP_PE_NLEVEL>	AMP-PE wavelet decomposition levels.
--amp-pe-wave-pec <AMP_PE_WAVE_PEC>	AMP-PE morphology-mask energy retention fraction (0.0-1.0)
--amp-pe-simulated-te <AMP_PE_SIMULATED_TE>	AMP-PE simulated echo time (s) used to turn the field into phase.
--amp-pe-max-linearization-ite <AMP_PE_MAX_LINEARIZATION_ITE>	AMP-PE linearization iterations per stage.
--amp-pe-damp-rate-sig <AMP_PE_DAMP_RATE_SIG>	AMP-PE GAMP signal-update damping rate.
--amp-pe-damp-rate-par <AMP_PE_DAMP_RATE_PAR>	AMP-PE parameter-estimation learning rate (kappa)
--amp-pe-max-pe-spar-ite <AMP_PE_MAX_PE_SPAR_ITE>	AMP-PE inner sparse-reconstruction iterations.
--amp-pe-max-pe-est-ite <AMP_PE_MAX_PE_EST_ITE>	AMP-PE inner parameter-estimation iterations.
--amp-pe-cvg-thd <AMP_PE_CVG_THD>	AMP-PE GAMP inner convergence threshold.
--amp-pe-tikhonov-beta <AMP_PE_TIKHONOV_BETA>	AMP-PE L2-seed Tikhonov weight.
--r2star-qsm-r-const-3t <R2STAR_QSM_R_CONST_3T>	R2*-QSM relaxometric constant at 3T (Hz/ppm)
--decompose-n-inner <DECOMPOSE_N_INNER>	DECOMPOSE inner alternating passes per voxel.
--decompose-chi-bound <BOUND>	DECOMPOSE bound on |χ| in the fit: a number in ppm (default: 0.5), or echo-spacing to compute it from the field strength and echo spacing.
--decompose-unwrap-phase <BOOL>	DECOMPOSE: unwrap the phase along the echoes (default: false, as in the authors’ code)
--decompose-fit <DECOMPOSE_FIT>	DECOMPOSE’s fit: reference-code (Tim Ho’s script; default) or paper (Chen et al. 2021: amplitudes summing to one, per-voxel S(0) = 1 scale), as the DECOMPOSE pipeline runs it. Values: reference-code, paper.
--decompose-max-iterations <DECOMPOSE_MAX_ITERATIONS>	DECOMPOSE: maximum solver iterations of each fit (lsqcurvefit MaxIterations; default: 400)
--decompose-max-function-evaluations <DECOMPOSE_MAX_FUNCTION_EVALUATIONS>	DECOMPOSE: maximum function evaluations of each fit (lsqcurvefit MaxFunctionEvaluations; default: 100 × the fit’s parameters)
--decompose-function-tolerance <DECOMPOSE_FUNCTION_TOLERANCE>	DECOMPOSE: relative cost change each fit stops at (lsqcurvefit FunctionTolerance; default: 1e-6)
--decompose-step-tolerance <DECOMPOSE_STEP_TOLERANCE>	DECOMPOSE: relative step each fit stops at (lsqcurvefit StepTolerance; default: 1e-6)
--decompose-optimality-tolerance <DECOMPOSE_OPTIMALITY_TOLERANCE>	DECOMPOSE: first-order optimality each fit stops at (lsqcurvefit OptimalityTolerance; default: 1e-6)
--decompose-finite-difference-step <DECOMPOSE_FINITE_DIFFERENCE_STEP>	DECOMPOSE: relative finite-difference step of the Jacobian (lsqcurvefit FiniteDifferenceStepSize; default: √eps)
--decompose-typical-x <DECOMPOSE_TYPICAL_X>	DECOMPOSE: typical magnitude of each fitted value (lsqcurvefit TypicalX; default: 1)
--chi-sep-ilsqr-dr-pos <HZ_PER_PPM>	χ-sep iLSQR paramagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-ilsqr-dr-neg <HZ_PER_PPM>	χ-sep iLSQR diamagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-ilsqr-pad <X> <Y> <Z>	χ-sep iLSQR zero-padding at both ends of each axis in voxels, X Y Z (default: 16 16 16)
--chi-sep-ilsqr-cg-max-iter-lsqr <CHI_SEP_ILSQR_CG_MAX_ITER_LSQR>	χ-sep iLSQR: maximum CG iterations of the LSQR step (default: 50)
--chi-sep-ilsqr-cg-tol-lsqr <CHI_SEP_ILSQR_CG_TOL_LSQR>	χ-sep iLSQR: CG relative tolerance of the LSQR step (default: 0.02)
--chi-sep-ilsqr-cg-max-iter-sa <CHI_SEP_ILSQR_CG_MAX_ITER_SA>	χ-sep iLSQR: maximum CG iterations of the streaking-removal step (default: 50)
--chi-sep-ilsqr-cg-tol-sa <CHI_SEP_ILSQR_CG_TOL_SA>	χ-sep iLSQR: CG relative tolerance of the streaking-removal step (default: 0.02)
--chi-sep-ilsqr-d2-thr <CHI_SEP_ILSQR_D2_THR>	χ-sep iLSQR: the streaking cone is where |D| is below this (default: 0.11)
--chi-sep-ilsqr-prc-w-i-min <CHI_SEP_ILSQR_PRC_W_I_MIN>	χ-sep iLSQR: percentile of the field’s Laplacian at which the LSQR weight is 1 (default: 60)
--chi-sep-ilsqr-prc-w-i-max <CHI_SEP_ILSQR_PRC_W_I_MAX>	χ-sep iLSQR: percentile of the field’s Laplacian at which the LSQR weight is 0 (default: 99.8)
--chi-sep-ilsqr-prc-w-g-min <CHI_SEP_ILSQR_PRC_W_G_MIN>	χ-sep iLSQR: percentile of each gradient’s magnitude at which the streaking-removal weight is 1 (default: 50)
--chi-sep-ilsqr-prc-w-g-max <CHI_SEP_ILSQR_PRC_W_G_MAX>	χ-sep iLSQR: percentile of each gradient’s magnitude at which the streaking-removal weight is 0 (default: 70)
--chi-sep-ilsqr-qsm-max-iter <CHI_SEP_ILSQR_QSM_MAX_ITER>	χ-sep iLSQR: max iterations of the iLSQR that reconstructs, from the local field, the QSM χ-separation is initialised with (default: 100; not used with –use-custom-qsm)
--chi-sep-ilsqr-qsm-tol <CHI_SEP_ILSQR_QSM_TOL>	χ-sep iLSQR: tolerance of that QSM’s iLSQR.
--chi-sep-ilsqr-qsm-pad-mm <MM>	χ-sep iLSQR: zero-padding of that QSM’s iLSQR, in mm per side (default: 12)
--chi-sep-ilsqr-qsm-precision <PRECISION>	χ-sep iLSQR: arithmetic of that QSM’s two LSQR solves: double, or single as the original implementation (default: double)
--chi-sep-medi-lambda <CHI_SEP_MEDI_LAMBDA>	χ-sep MEDI weight of the data terms against the L1 terms (default: 1)
--chi-sep-medi-lambda-csf <CHI_SEP_MEDI_LAMBDA_CSF>	χ-sep MEDI weight of the CSF-uniformity term (default: 1)
--chi-sep-medi-dr-pos <HZ_PER_PPM>	χ-sep MEDI paramagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-medi-dr-neg <HZ_PER_PPM>	χ-sep MEDI diamagnetic relaxometric constant (Hz/ppm), or auto for 137 × B0 / 2.9 T (default: auto)
--chi-sep-medi-pad <CHI_SEP_MEDI_PAD>	χ-sep MEDI zero-padding at both ends of the third axis, in voxels (default: 5)
--chi-sep-medi-tol <CHI_SEP_MEDI_TOL>	χ-sep MEDI: stop when the relative change falls below this (default: 0.03)
--chi-sep-medi-max-iter <CHI_SEP_MEDI_MAX_ITER>	χ-sep MEDI maximum Gauss-Newton iterations (default: 30)
--chi-sep-medi-cg-tol <CHI_SEP_MEDI_CG_TOL>	χ-sep MEDI inner CG tolerance (default: 0.05)
--chi-sep-medi-cg-max-iter <CHI_SEP_MEDI_CG_MAX_ITER>	χ-sep MEDI inner CG max iterations (default: 100)
--chi-sep-medi-qsm-max-iter <CHI_SEP_MEDI_QSM_MAX_ITER>	χ-sep MEDI: max iterations of the iLSQR that reconstructs, from the local field, the QSM χ-separation is initialised with (default: 100; not used with –use-custom-qsm)
--chi-sep-medi-qsm-tol <CHI_SEP_MEDI_QSM_TOL>	χ-sep MEDI: tolerance of that QSM’s iLSQR.
--chi-sep-medi-qsm-pad-mm <MM>	χ-sep MEDI: zero-padding of that QSM’s iLSQR, in mm per side (default: 12)
--chi-sep-medi-qsm-precision <PRECISION>	χ-sep MEDI: arithmetic of that QSM’s two LSQR solves: double, or single as the original implementation (default: double)
--chi-sep-medi-csf-r2star <METHOD>	χ-sep MEDI: the R2* its CSF mask is extracted from: arlo (MEDI’s ARLO; the default) or nnls (the χ-separation toolbox’s r2star_nnls, which its script runs)
--wavesep-dr-pos <WAVESEP_DR_POS>	WaveSep paramagnetic relaxivity (Hz/ppm)
--wavesep-dr-neg <WAVESEP_DR_NEG>	WaveSep diamagnetic relaxivity (Hz/ppm)
--wavesep-alpha <WAVESEP_ALPHA>	WaveSep proximal-gradient step size.
--wavesep-lambda <WAVESEP_LAMBDA>	WaveSep wavelet L1 soft-threshold weight.
--wavesep-wavelet-order <WAVESEP_WAVELET_ORDER>	WaveSep Daubechies wavelet order.
--wavesep-max-iter <WAVESEP_MAX_ITER>	WaveSep ISTA max iterations.
--wavesep-tol <WAVESEP_TOL>	WaveSep relative-change stop tolerance.
--hc-chisep-dr-pos-3t <HC_CHISEP_DR_POS_3T>	HC-ChiSep paramagnetic relaxivity at 3T (Hz/ppm)
--hc-chisep-bin-hz <HC_CHISEP_BIN_HZ>	HC-ChiSep R2’ bin width for the anchored grid search (Hz)
--chi-sepnet-route <CHI_SEPNET_ROUTE>	χ-sepnet’s input route. toolbox (default), the authors’: the inputs resampled to 1 mm with B0 along z, the susceptibility channel computed by QSMnetME, R2’ measured or predicted from R2* by R2PRIMEnet, the maps resampled back. native: the network alone on the acquired grid, given this run’s QSM and R2’. Values: toolbox, native.
--chi-sepnet-patch <D> <H> <W>	χ-sepnet (--chi-sepnet-route native) sliding-window patch in voxels, D H W, each a multiple of 16 (default: 192 192 128, the authors’)
--chi-sepnet-frame-patch <D> <H> <W>	χ-sepnet (--chi-sepnet-route toolbox): run the networks as a sliding window of patches this size over the 1 mm frame, D H W, each a multiple of 16, so memory follows the patch. Omit to run the whole frame at once, as the authors do.
--susep-net-patch <D> <H> <W>	Run SUSEP-Net as a sliding window of patches this size in voxels, D H W, each a multiple of 8, so its memory follows the patch, not the volume. Omit to run the whole volume at once.
--motion-correction	Rigidly co-register every echo onto a reference echo (estimated on magnitude, applied to phase through the complex signal) before field mapping, and restrict the mask to voxels every echo covers. Off by default. Needs magnitude images.
--motion-reference <first|last|N>	Echo the others are aligned onto: first (default), last, or a 1-based echo number.
--no-motion-chain	Register every echo from the identity instead of seeding it from its neighbour’s solution.
--motion-chain-levels <MOTION_CHAIN_LEVELS>	Pyramid levels for a seeded (chained) registration (default 3)
--motion-min-ncc <MOTION_MIN_NCC>	NCC below which an echo’s registration is reported as suspect (default 0.3, set for echoes; ~0.9 suits repeats)
--registration-levels <REGISTRATION_LEVELS>	Registration pyramid levels: the capture range; do not lower (default 5)
--registration-max-samples <REGISTRATION_MAX_SAMPLES>	Registration: voxels sampled per metric evaluation, per level (default 200000)
--registration-max-iterations <REGISTRATION_MAX_ITERATIONS>	Registration: Powell cycles per level (default 20)
--registration-max-evaluations <REGISTRATION_MAX_EVALUATIONS>	Registration: metric evaluations across all levels, a hard stop (default 5000)
--registration-tolerance <REGISTRATION_TOLERANCE>	Registration: relative cost improvement per cycle that counts as converged (default 1e-5)
--registration-rotation-step-deg <REGISTRATION_ROTATION_STEP_DEG>	Registration: initial rotation step at the finest level, degrees (default 3)
--registration-translation-step-mm <REGISTRATION_TRANSLATION_STEP_MM>	Registration: initial translation step at the finest level, mm (default 3)
--registration-min-overlap <REGISTRATION_MIN_OVERLAP>	Registration: smallest fraction of sampled voxels that must overlap (default 0.25)
--vsharp-deconvolve <BOOL>	Undo V-SHARP’s SMV high-pass with a truncated-SVD deconvolution (default: false)
--vsharp-threshold <VSHARP_THRESHOLD>	V-SHARP deconvolution (TSVD) threshold; only used with –vsharp-deconvolve true.
--vsharp-max-radius <VSHARP_MAX_RADIUS>	V-SHARP max kernel radius in mm.
--vsharp-min-radius <VSHARP_MIN_RADIUS>	V-SHARP min kernel radius in mm.
--vsharp-radius-step <MM>	V-SHARP step between kernel radii in mm (default: 1); the radii run from –vsharp-max-radius down to –vsharp-min-radius.
--vsharp-grid <VSHARP_GRID>	Grid V-SHARP estimates the background on: isotropic (default; spacing from –vsharp-grid-spacing), native (the acquired grid), or finest (isotropic at the smallest voxel dimension, e.g. 0.8 mm for 0.8x0.8x3 mm; no axis is coarsened). The local field keeps native resolution. Values: native, finest, isotropic.
--vsharp-grid-spacing <MM>	V-SHARP isotropic working-grid spacing in mm (default 1); used with –vsharp-grid isotropic.
--pdf-tol <PDF_TOL>	PDF tolerance, relative to the noise level.
--pdf-n-cg <PDF_N_CG>	PDF maximum conjugate-gradient iterations (default: 30)
--pdf-space <SPACE>	PDF dipole kernel: image (the FFT of the dipole sampled on the voxel grid) or kspace (the analytic k-space kernel) (default: image)
--pdf-n-pad <VOXELS>	PDF zero-padding around the mask’s bounding box, in voxels; 0 runs on the full volume (default: 40)
--lbv-tol <LBV_TOL>	LBV tolerance: stop the coarsest grid’s relaxation when its summed absolute update falls below this, in radians over the echo spacing (default: 0.01)
--lbv-depth <LEVEL>	LBV coarsest multigrid level, at most 8 (default: log2 of the smallest dimension minus 3)
--lbv-peel <LAYERS>	LBV boundary layers removed from the mask beyond the first (default: 0)
--lbv-n1 <LBV_N1>	LBV relaxations on each grid before the coarser grid (default: 30)
--lbv-n2 <LBV_N2>	LBV relaxations on each grid after the coarser grid (default: 100)
--lbv-n3 <LBV_N3>	LBV relaxations on the finest grid at the end (default: 100)
--lbv-precision <PRECISION>	LBV arithmetic: double, or single as the original implementation (default: double)
--ismv-variant <VARIANT>	iSMV implementation: wen2014 (the authors’ paper and code) or qsmjl (QSM.jl’s) (default: wen2014)
--ismv-tol <ISMV_TOL>	iSMV tolerance: the relative change between iterates it stops at (qsmjl: the residual relative to the first) (default: 5e-5)
--ismv-max-iter <ISMV_MAX_ITER>	iSMV max iterations (default: no limit)
--ismv-radius <MM>	iSMV kernel radius in mm, or auto for 6 × the smallest voxel dimension (qsmjl: 2 × the largest) (default: auto)
--sharp-threshold <SHARP_THRESHOLD>	SHARP threshold.
--sharp-radius <SHARP_RADIUS>	SHARP kernel radius in mm.
--resharp-radius <RESHARP_RADIUS>	RESHARP SMV kernel radius in mm.
--resharp-tik-reg <RESHARP_TIK_REG>	RESHARP Tikhonov regularization parameter.
--resharp-tol <RESHARP_TOL>	RESHARP conjugate gradients squared (CGS) relative tolerance.
--resharp-max-iter <RESHARP_MAX_ITER>	RESHARP maximum CGS iterations.
--harperella-radius <HARPERELLA_RADIUS>	HARPERELLA SMV kernel radius in mm.
--harperella-max-iter <HARPERELLA_MAX_ITER>	HARPERELLA maximum iterations.
--harperella-tol <HARPERELLA_TOL>	HARPERELLA convergence tolerance.
--harperella-pad <X> <Y> <Z>	HARPERELLA zero-padding in voxels per side, X Y Z (default: 0 0 0)
--harperella-rim <RIM>	HARPERELLA boundary rim: whole (zero every Laplacian value in the rim; default) or sti (zero the 20 % largest, as STI Suite does)
--iharperella-max-iter <IHARPERELLA_MAX_ITER>	iHARPERELLA maximum iterations.
--iharperella-tol <IHARPERELLA_TOL>	iHARPERELLA convergence tolerance.
--iharperella-pad <X> <Y> <Z>	iHARPERELLA zero-padding in voxels per side, X Y Z (default: 0 0 0)
--iharperella-rim <RIM>	iHARPERELLA boundary rim: whole (zero every Laplacian value in the rim) or sti (zero the 20 % largest, as STI Suite does; default)
--no-romeo-phase-coherence	ROMEO: disable phase coherence weights (default: on)
--no-romeo-phase-gradient-coherence	ROMEO: disable phase gradient coherence weights (default: on)
--no-romeo-phase-linearity	ROMEO: disable phase linearity weights (default: on; the romeo command line’s romeo3 weights)
--romeo-phase-linearity	ROMEO: enable phase linearity weights (default: on; for a –pipeline that turns them off)
--no-romeo-mag-coherence	ROMEO: disable magnitude coherence weights (default: on)
--romeo-mag-weight	ROMEO: enable magnitude weighting (default: off)
--no-romeo-mag-weight	ROMEO: disable magnitude weighting (overrides –romeo-mag-weight; kept for compatibility)
--romeo-mag-weight2	ROMEO: enable second magnitude weighting (flow-artifact penalty; default: off)
--romeo-bestpath	ROMEO: use Best-path (Abdul-Rahman) weights instead of ROMEO weights (default: off)
--romeo-template <ROMEO_TEMPLATE>	ROMEO: template echo index (1-indexed, only for template mode)
--romeo-individual	ROMEO: use individual per-echo unwrapping (default)
--no-romeo-individual	ROMEO: use template-based temporal unwrapping (disables individual mode)
--romeo-correct-global	ROMEO: correct global 2π offsets, of each echo and between echoes (default: on)
--no-romeo-correct-global	ROMEO: do not correct global 2π offsets.
--romeo-temporal-uncertain-unwrapping <ROMEO_TEMPORAL_UNCERTAIN_UNWRAPPING>	ROMEO: temporal uncertain-unwrapping quality threshold [0,1] (0 disables; default: 0.5)
--romeo-max-seeds <ROMEO_MAX_SEEDS>	ROMEO: maximum number of seed regions (default: 1)
--romeo-merge-regions	ROMEO: merge neighboring regions after unwrapping (default: off)
--romeo-correct-regions	ROMEO: correct each region’s median to nearest 0 (default: off)
--romeo-wrap-addition <ROMEO_WRAP_ADDITION>	ROMEO: additional phase tolerance beyond π for neighbor differences [0,π] (default: 0.0)
--prelude-labelling <PRELUDE_LABELLING>	PRELUDE: how the initial regions are labelled (default: auto) Values: auto, slices, volume.
--prelude-partitions <N>	PRELUDE: number of phase partitions the initial regions are formed from (default: 8)
--swi-unwrapping <SWI_UNWRAPPING>	SWI phase unwrapping before the echoes are combined [default: laplacian]. Values: laplacian, laplacian-slice, romeo.
--swi-hp-sigma <SWI_HP_SIGMA>	SWI high-pass filter sigma (3 values, in voxels)
--swi-scaling <SWI_SCALING>	SWI phase-mask scaling. Values: tanh, negative-tanh, positive, negative, triangular.
--swi-strength <SWI_STRENGTH>	SWI phase scaling strength.
--swi-mip-window <SWI_MIP_WINDOW>	SWI MIP window size in slices.
--smwi-threshold <SMWI_THRESHOLD>	SMWI susceptibility threshold in ppm — the |χ| at which the weighting mask reaches zero. The 1 ppm default follows SEPIA and targets strong sources; brain tissue wants far less.
--smwi-power <SMWI_POWER>	SMWI mask power (contrast strength)
--smwi-mip-window <SMWI_MIP_WINDOW>	SMWI mIP window size in slices.
--synthseg-version <VERSION>	SynthSeg weights generation: v1 (32 labels, ships with SynthSeg) or v2 (33 labels, adds CSF; weights not published yet, so refused unless synthseg-2.0.onnx is in $QSM_MODEL_DIR)
--synthseg-crop <SYNTHSEG_CROP>	Cubic centre-crop in voxels at 1 mm before inference (rounded up to a multiple of 32). Omit to process the whole volume, as the SynthSeg CLI does.
--synthseg-patch <VOXELS>	Run SynthSeg as a sliding window of half-overlapping cubes this many voxels on a side at 1 mm (at least 64, rounded up to a multiple of 32), so its memory follows the patch, not the volume. Omit to run the whole volume at once, as the SynthSeg CLI does.
--no-synthseg-flip-averaging	Skip the left-right flipped second pass. Halves inference time, as SynthSeg’s –fast does.
--no-synthseg-topology-cleanup	Skip resetting each topological class to its largest connected component.
--synthseg-sigma <SYNTHSEG_SIGMA>	Gaussian blur on the posteriors, in voxels (0 disables it)
--use-custom-dseg <TOOL>	Prefer a bring-your-own segmentation from <bids>/derivatives/<TOOL>/ when present. Bare flag = first matching tool alphabetically. Always falls back to running SynthSeg.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--inhomogeneity-method <INHOMOGENEITY_METHOD>	Inhomogeneity correction method: makehomogeneous (default; –homogeneity-*) or n4 (N4 bias field correction, as ANTs’ N4BiasFieldCorrection; –n4-*) Values: makehomogeneous, n4.
--homogeneity-sigma-mm <HOMOGENEITY_SIGMA_MM>	Inhomogeneity correction smoothing sigma in mm (makehomogeneous)
--homogeneity-nbox <HOMOGENEITY_NBOX>	makehomogeneous: boxes per dimension for the tissue segmentation the receive sensitivity is estimated from.
--n4-shrink-factor <N4_SHRINK_FACTOR>	N4: subsample the image by this factor before fitting the bias field (default 4)
--n4-max-iterations <ITERATIONS>	N4: maximum iterations at each fitting level, one number a level (default 50 50 50 50); the B-spline mesh doubles from one level to the next.
--n4-convergence-threshold <N4_CONVERGENCE_THRESHOLD>	N4: stop a level once the change in the bias field (its coefficient of variation) falls to this (default 0: every iteration runs)
--n4-mesh <X> <Y> <Z>	N4: B-spline mesh elements per axis at the first level, X Y Z (default 1 1 1)
--n4-spline-order <N4_SPLINE_ORDER>	N4: order of the B-spline bias field, 1 to 3 (default 3)
--n4-bias-field-fwhm <N4_BIAS_FIELD_FWHM>	N4: FWHM of the Gaussian modelling the bias field’s blurring of the log-intensity histogram (default 0.15)
--n4-wiener-noise <N4_WIENER_NOISE>	N4: noise term of the Wiener deconvolution of the histogram (default 0.01)
--n4-histogram-bins <N4_HISTOGRAM_BINS>	N4: bins of the log-intensity histogram (default 200)
--n4-geometry <GEOMETRY>	N4: the image geometry it is given: scanner (the scan’s header, as ANTs reads the acquired NIfTI; the default) or make-nii (the axis-aligned header MATLAB’s make_nii writes, as QSMART runs N4). N4’s shrunk grid falls where the header puts it.
--linear-fit-estimate-offset <LINEAR_FIT_ESTIMATE_OFFSET>	Linear and robust fit: estimate a phase offset term (default: true). The robust fit needs 4 echoes with it, 3 without.
--robust-fit-loss <ROBUST_FIT_LOSS>	Motion-robust fit (--b0-estimation robust-fit): loss turning residuals into weights. tukey (default) rejects echoes past the tuning radius; huber caps but never rejects. Values: tukey, huber.
--robust-fit-tuning <ROBUST_FIT_TUNING>	Motion-robust fit: tuning radius in robust-scale units (default: the loss’s conventional value, Tukey 4.685 / Huber 1.345). Smaller is more aggressive.
--robust-fit-iterations <ROBUST_FIT_ITERATIONS>	Motion-robust fit: IRLS passes (default 3; 0 = report only, the field is the linear fit’s)
--robust-fit-downweight-level <ROBUST_FIT_DOWNWEIGHT_LEVEL>	Motion-robust fit (reporting only): a robust weight below this counts as down-weighted (default 0.75; in (0, 1])
--robust-fit-flag-ratio <ROBUST_FIT_FLAG_RATIO>	Motion-robust fit (reporting only): flag an echo whose outlier score exceeds this many times the best echo’s (default 3; at least 1)
--complex-fit-reltol <COMPLEX_FIT_RELTOL>	Complex fit (--b0-estimation complex-fit): stop once the update is this fraction of the first estimate, over the whole volume (default 1e-4)
--complex-fit-max-iter <COMPLEX_FIT_MAX_ITER>	Complex fit: Gauss-Newton iterations at most (default 30)
--no-qsm	Skip QSM processing (only run supplementary outputs like SWI, T2*, R2*)
--do-swi	Also compute SWI.
--swi-filtered-phase	Also write CLEAR-SWI’s high-pass filtered phase, the phase the SWI is weighted by (_desc-swi_phase.nii): Hz for several echoes, radians for one, 0 outside CLEAR-SWI’s head mask. Implies –do-swi.
--no-magnitude	Do not publish the combined magnitude (part-mag_T2starw). It is still computed, and kept with the run’s intermediates, when a step reads it (SMWI, segmentation, a magnitude-weighted inversion)
--do-segmentation	Also segment the brain from the GRE magnitude with SynthSeg, writing a FreeSurfer-labelled _dseg.nii and its BIDS lookup table. Needs no T1w and no registration. Needs a deep-learning build. doi:10.1016/j.media.2023.102789.
--do-analysis	Also summarise every quantitative map (χ, χ+/χ−, T2*, R2*, R2, R2’) per segmented structure into a TSV: mean, SD, median, min, max and the 5th/95th percentiles, one row per map and structure, plus a dataset-level table gathering all runs. Summarises whatever the run made — it implies –do-segmentation, and QSM only if no other map was asked for.
--do-smwi	Also compute SMWI (susceptibility map-weighted imaging): weights the magnitude by a mask built from the susceptibility map rather than from high-pass filtered phase, so the contrast stays where the source is. Needs χ, so it implies QSM. doi:10.1002/mrm.24920.
--do-t2starmap	Compute T2* relaxation map from multi-echo magnitude data.
--do-r2starmap	Compute R2* decay rate map from multi-echo magnitude data.
--do-r2map	Compute R2 map from a multi-echo spin-echo (MESE) acquisition (EPG)
--do-r2primemap	Compute R2’ map (= R2* − R2; needs GRE magnitude + a MESE acquisition)
--do-chisep	Compute chi-separation (paramagnetic/diamagnetic susceptibility maps)
--chisep <CHI_SEPARATION_ALGORITHM>	Chi-separation method (default: r2star-qsm) Values: r2star-qsm, decompose, chi-sep-ilsqr, chi-sep-medi, wavesep, hc-chisep, susep-net, chi-sepnet.
--use-custom-qsm <TOOL>	Use a bring-your-own QSM (Chimap) from <bids>/derivatives/<TOOL>/ for chi-separation.
--use-custom-r2 <TOOL>	Use a bring-your-own R2 map from <bids>/derivatives/<TOOL>/ for chi-separation.
--use-custom-r2prime <TOOL>	Use a bring-your-own R2’ map from <bids>/derivatives/<TOOL>/ for chi-separation.
--r2prime-strategy <R2PRIME_STRATEGY>	How to obtain R2’ when no custom map is supplied. auto (default) measures it as R2* - R2 from a MESE acquisition when one is present and predicts it from R2* with R2PRIMEnet when it is not; mese only ever measures it, producing nothing without spin-echo data; r2primenet always predicts it. A measurement and an estimate are not the same thing — the generated methods text says which route was configured. Values: auto, mese, r2primenet.
--r2primenet-ignore-field-mismatch	Run R2PRIMEnet even when the acquisition is not at 3 T, the only field strength its weights were trained at; the predicted R2’ is then uncalibrated. Without it, an explicit --r2prime-strategy r2primenet refuses a non-3 T run and auto skips R2’ with a warning.
--r2primenet-patch <D> <H> <W>	R2PRIMEnet sliding-window patch in voxels, D H W, each a multiple of 16 (default: 192 192 128, the authors’)
--export-dicom	Also export final maps as DICOM series into each subject’s extra_files/ folder.
--inhomogeneity-correction	Apply inhomogeneity correction to magnitude before masking.
--no-inhomogeneity-correction	Disable inhomogeneity correction.
--axial-resampling	Resample oblique acquisitions onto an axial grid when their obliquity exceeds –obliquity-threshold. Off by default (methods that only work on axial data resample regardless)
--no-axial-resampling	Do not resample oblique acquisitions to axial (overrides a config’s axial_resampling)
--obliquity-threshold <OBLIQUITY_THRESHOLD>	Obliquity in degrees above which –axial-resampling resamples (default: 10)
--crop-to-mask	Reconstruct inside a box around the mask instead of the whole field of view. Cuts the FFT-bound stages roughly in proportion to the voxels removed, but those stages are a small share of a full run, so the end-to-end gain is usually modest. Off by default. Outputs are written on the full grid either way; this only changes where the arithmetic happens. Padding to an FFT-friendly size happens regardless.
--ignore-slice-gap	Reconstruct even when a 2D acquisition’s sidecar SliceThickness is thinner than the slice pitch. Without it, QSM refuses such a run: the gap between slices is unsampled tissue, and the FFT kernels assume a contiguous grid, so the map would be wrong.
--fft-padding	Pad the grid outward to an FFT-friendly size before the FFT-bound stages. Much faster on an awkward grid, and discards nothing, but it changes where k-space is sampled and so shifts the reconstruction slightly. Off by default.
--tukey-window <TUKEY_WINDOW>	Apply a k-space Tukey window with this taper ratio (0 to 1) to each echo’s complex image before every other step, as the χ-separation toolbox does: 0.4, and 0 for Philips data. Default 0 (none). Needs magnitude images.
--crop-margin-mm <CROP_MARGIN_MM>	Margin in mm around the mask when –crop-to-mask is set (default 32). Below ~32 mm the reconstruction starts to differ from the uncropped one; it must also clear the largest SMV kernel in use, and V-SHARP defaults to 12 mm.
--output-space <OUTPUT_SPACE>	Grid to write derivatives on when the run was resampled to axial. acquired (default) puts them back where the data came from, so transforms you already hold still apply; working leaves them on the resampled grid and avoids a second interpolation. Values: acquired, working.
--mask-preset <MASK_PRESET>	Mask preset: robust-threshold, bet, hd-bet (HD-BET deep-learning brain extraction + signal-gated erosion; needs a deep-learning build), bet-and-phase (BET on the magnitude AND a thresholded phase-quality map, then hole-filled and eroded), rs2-net (RS2-Net deep-learning rodent brain extraction), or mindgrab (MindGrab deep-learning brain extraction for any contrast; needs a deep-learning build) Values: robust-threshold, bet, hd-bet, bet-and-phase, rs2-net, mindgrab.
--use-custom-masks <USE_CUSTOM_MASKS>	Prefer a bring-your-own mask from BIDS derivatives when present (else compute it). Bare flag = first matching derivatives tool (alphabetical); give a name (e.g. bet) to restrict to derivatives/<tool>/sub-*/anat/*_mask.nii*. Always falls back.
--mask <MASK_SECTIONS_CLI>	Define a mask section (repeatable; multiple sections are combined with –mask-combine). Format: <input>,<generator>,<refinement1>,<refinement2>,… Example: phase-quality,threshold:otsu,dilate:2,fill-holes:0,erode:2 Example: magnitude,bet:0.5,erode:2 Example: magnitude-first,bet:0.2,min-intensity:100,majority,close:2 (QSMART’s mask) Example: magnitude,hd-bet,signal-erode (HD-BET: hd-bet[:full|window|XxYxZ|low-memory][:tta][:step=F]; signal-erode: signal-erode[:threshold[:depth_cap[:global_erosions[:bias_sigma[:min_component]]]]])
--mask-combine <MASK_COMBINE>	How to combine multiple –mask sections: or (union, the default) or and (intersection — a voxel survives only where every section keeps it) Values: or, and.
--mask-refine <MASK_REFINEMENTS_CLI>	Refinement applied to the combined mask, after –mask-combine (repeatable). Takes the same refinement ops as –mask, e.g. erode:2 or fill-holes:0; generators (threshold/bet/hd-bet) belong in a –mask section instead. Example: –mask-combine and –mask-refine fill-holes:0 –mask-refine erode:1.
--two-pass	Two-pass artefact reduction: reconstruct a second time against a mask whose holes are left unfilled, and keep that reconstruction wherever it is defined. Strong susceptibility sources (air-tissue interfaces, haemorrhage, implants) make the phase around them unrecoverable; leaving those voxels out keeps their streaking from spreading, while the ordinary mask still supplies values inside the holes. Roughly doubles background removal and inversion time, and is not free of downsides — where the holes were not artefactual it can be worse than a single pass. Based on doi:10.1002/mrm.29048 (Stewart et al.). Off by default. The conventional single-pass map is kept alongside as _desc-singlepass_Chimap.nii.
--two-pass-mask <TWO_PASS_SECTIONS_CLI>	Define the reliable-pass mask used by –two-pass (repeatable; same format as –mask). Defaults to phase-quality,threshold:otsu — an unfilled threshold on the phase-quality map. Its holes are the point, so refinements that close them (fill-holes, close, gaussian) defeat the method. Folded with –two-pass-mask-combine, then –mask-refine like the main mask, and always intersected with it.
--two-pass-mask-combine <TWO_PASS_MASK_COMBINE>	How to combine multiple –two-pass-mask sections: or (union) or and (intersection). Defaults to the main mask’s –mask-combine. Turns on –two-pass. Values: or, and.
--dismiss-recommendation <RULE=VALUE>	Stop offering a recommendation for the setting’s current value: RULE=VALUE, the rule’s id and the value it was dismissed for, as qsmxt run --dry lists them (e.g. field-mapping.phase-offsets=false). Repeatable. A change to the setting brings the recommendation back. A required one cannot be dismissed, and a reference-pipeline difference has nothing to dismiss.
--debug	Enable debug logging.
--force	Recompute the multi-orientation reconstructions, ignoring their cached state.
--clean-intermediates	Once the dataset-level steps are done, delete the intermediate images under workflow/ (as qsmxt run --clean-intermediates)

qsmxt mask

Masking operations

Subcommand	Description
otsu	Otsu automatic thresholding
value	Fixed value thresholding
percentile	Percentile thresholding
sti-gui	STI Suite QSM_GUI’s thresholding: the level its thresh_tool starts at (Otsu’s threshold of the middle slice, scaled to that slice’s maximum)
bet	Brain extraction (BET)
hd-bet	Deep-learning brain extraction (HD-BET; downloads weights on first use)
rs2-net	Deep-learning rodent brain extraction (RS2-Net; downloads weights on first use)
mindgrab	Deep-learning brain extraction for any contrast (MindGrab; downloads weights on first use)
preset	Run a --mask-preset recipe on an image (robust-threshold, bet, hd-bet, bet-and-phase, rs2-net, mindgrab)
robust	Alias for preset robust-threshold
erode	Erode a binary mask
dilate	Dilate a binary mask
close	Morphological closing on a binary mask
majority	3x3x3 majority filter on a binary mask (MATLAB’s bwmorph3 ‘majority’)
largest-component	Keep only the largest 26-connected component of a binary mask (MATLAB’s bwlabeln)
imfill	Fill every hole of a binary mask, whatever its size (MATLAB’s imfill ‘holes’)
fill-holes	Fill holes in a binary mask
smooth	Gaussian smooth a binary mask (re-thresholds at 0.5)
and	Intersect binary masks: keep voxels every input keeps (the run --mask-combine and)
or	Union binary masks: keep voxels any input keeps (the run --mask-combine or)

qsmxt mask otsu

Otsu automatic thresholding

qsmxt mask otsu [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)

qsmxt mask value

Fixed value thresholding

qsmxt mask value [OPTIONS] --output <OUTPUT> --threshold <THRESHOLD> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--threshold <THRESHOLD>	Threshold value. Required.

qsmxt mask percentile

Percentile thresholding

qsmxt mask percentile [OPTIONS] --output <OUTPUT> --percentile <PERCENTILE> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--percentile <PERCENTILE>	Percentile value (0-100) Required.

qsmxt mask sti-gui

STI Suite QSM_GUI’s thresholding: the level its thresh_tool starts at (Otsu’s threshold of the middle slice, scaled to that slice’s maximum)

qsmxt mask sti-gui [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)

qsmxt mask bet

Brain extraction (BET)

qsmxt mask bet [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--fractional-intensity <FRACTIONAL_INTENSITY>	Fractional intensity (0.0-1.0, smaller = larger brain) Default: 0.5.
--voxel-scale <FACTOR>	Factor applied to the voxel sizes before BET runs — 10 for mouse data, 1 for the acquisition’s own geometry (see –bet-voxel-scale). Prefer mask rs2-net, which masks rodent brains without the trick.
--orientation <ORIENTATION>	Which way round BET sees the left-right axis: fsl (default) or as-stored (see –bet-orientation)

qsmxt mask hd-bet

Deep-learning brain extraction (HD-BET; downloads weights on first use)

qsmxt mask hd-bet [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--window	Run in 192x192x96 sliding-window patches (about 3.6–4.5 GB measured) instead of on the whole volume (the default; about 6–8 GB measured, and about twice as fast)
--low-memory	Use 128x128x64 sliding-window patches (about 1.5–1.9 GB measured)
--patch <PATCH>	Sliding-window patch size XxYxZ in voxels at 1 mm (multiples of 32x32x16)
--tta	8-fold mirroring test-time augmentation (about 8x slower)
--tile-step <TILE_STEP>	Sliding-window step as a fraction of the patch, in (0, 1]. Larger steps mean fewer patches and a shorter run: 0.5 (default) is HD-BET’s own, 1.0 abuts the patches. Implies –window.

qsmxt mask rs2-net

Deep-learning rodent brain extraction (RS2-Net; downloads weights on first use)

qsmxt mask rs2-net [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--tta	8-fold mirroring test-time augmentation (about 8x slower)
--tile-step <TILE_STEP>	Sliding-window step as a fraction of the patch, in (0, 1]. Larger steps mean fewer patches and a shorter run: 0.5 (default) is nnU-Net’s own, 1.0 abuts the patches.

qsmxt mask mindgrab

Deep-learning brain extraction for any contrast (MindGrab; downloads weights on first use)

qsmxt mask mindgrab [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--tta	Left-right flip test-time augmentation (about 2x slower)

qsmxt mask preset

Run a --mask-preset recipe on an image (robust-threshold, bet, hd-bet, bet-and-phase, rs2-net, mindgrab)

qsmxt mask preset [OPTIONS] --output <OUTPUT> <PRESET> <INPUT>

Argument	Description
<PRESET>	Which recipe — the same names as run --mask-preset. Values: robust-threshold, bet, hd-bet, bet-and-phase, rs2-net, mindgrab. Required.
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--quality <QUALITY>	Phase-quality map (from qsmxt quality-map) for the sections that read phase quality. Without it those sections read the input image instead, as run --masking-input would.
--phase <PHASE>	Phase image (3-D, or 4-D multi-echo): the mask leaves out voxels where the scanner recorded no phase (a constant value in every echo), as qsmxt run does for every recipe. A phase-quality map rates that flat phase as reliable, so give this with --quality.
--bet-orientation <ORIENTATION>	Which way round the recipe’s BET steps see the left-right axis: fsl (default) or as-stored (see run --bet-orientation)

qsmxt mask robust

Alias for preset robust-threshold

qsmxt mask robust [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)

qsmxt mask erode

Erode a binary mask

qsmxt mask erode [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--iterations <ITERATIONS>	Number of erosion iterations. Default: 1.

qsmxt mask dilate

Dilate a binary mask

qsmxt mask dilate [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--iterations <ITERATIONS>	Number of dilation iterations. Default: 1.

qsmxt mask close

Morphological closing on a binary mask

qsmxt mask close [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--radius <RADIUS>	Closing radius. Default: 1.

qsmxt mask majority

3x3x3 majority filter on a binary mask (MATLAB’s bwmorph3 ‘majority’)

qsmxt mask majority [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)

qsmxt mask largest-component

Keep only the largest 26-connected component of a binary mask (MATLAB’s bwlabeln)

qsmxt mask largest-component [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)

qsmxt mask imfill

Fill every hole of a binary mask, whatever its size (MATLAB’s imfill ‘holes’)

qsmxt mask imfill [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)

qsmxt mask fill-holes

Fill holes in a binary mask

qsmxt mask fill-holes [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--max-size <MAX_SIZE>	Maximum hole size in voxels (0 = automatic, 5% of the volume — what every preset uses) Default: 0.

qsmxt mask smooth

Gaussian smooth a binary mask (re-thresholds at 0.5)

qsmxt mask smooth [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI file (an image for the generating subcommands, a mask for the rest) Required.
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Further refinement, applied in order after the subcommand’s own operation (repeatable). Examples: erode:2, dilate:1, fill-holes:0, close:1, gaussian:4.0, majority (3x3x3 majority filter), largest-component, imfill (every hole), min-intensity:100 (remove voxels below 100 in the input image; needs an image input), signal-erode (signal-gated erosion of low-signal boundary voxels; needs a magnitude input)
--sigma <SIGMA>	Gaussian sigma in mm. Default: 4.

qsmxt mask and

Intersect binary masks: keep voxels every input keeps (the run --mask-combine and)

qsmxt mask and [OPTIONS] --output <OUTPUT> [INPUTS] [INPUTS]...

Argument	Description
<INPUTS>	Input mask NIfTI files (two or more, on the same grid)
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Refinement operation applied to the combined mask (repeatable, applied in order) — the run --mask-refine of this command. Examples: fill-holes:0, erode:1, close:1.
--magnitude <MAGNITUDE>	Magnitude image, needed only by --op signal-erode.

qsmxt mask or

Union binary masks: keep voxels any input keeps (the run --mask-combine or)

qsmxt mask or [OPTIONS] --output <OUTPUT> [INPUTS] [INPUTS]...

Argument	Description
<INPUTS>	Input mask NIfTI files (two or more, on the same grid)
-o, --output <OUTPUT>	Output mask NIfTI file. Required.
--op <OPS>	Refinement operation applied to the combined mask (repeatable, applied in order) — the run --mask-refine of this command. Examples: fill-holes:0, erode:1, close:1.
--magnitude <MAGNITUDE>	Magnitude image, needed only by --op signal-erode.

qsmxt combine

Multi-coil phase combination (uncombined receive-coil channels -> combined echoes)

Subcommand	Description
mcpc3ds	MCPC-3D-S (ASPIRE) coil combination from per-coil multi-echo phase + magnitude

qsmxt combine mcpc3ds

MCPC-3D-S (ASPIRE) coil combination from per-coil multi-echo phase + magnitude

qsmxt combine mcpc3ds [OPTIONS] --phase <PHASE>... --magnitude <MAGNITUDE>... --output <OUTPUT>

Argument	Description
--phase <PHASE>	Per-coil, per-echo wrapped phase NIfTI files (3D each). Files carrying BIDS coil-NN and echo-N entities are ordered by those; otherwise they are taken coil-major in the given order and --num-echoes is required. Required.
--magnitude <MAGNITUDE>	Per-coil, per-echo magnitude NIfTI files, same layout as –phase. Required.
-o, --output <OUTPUT>	Output prefix: writes <prefix>rec-mcpc3ds_echo-N_part-phase<suffix>.nii, <prefix>rec-mcpc3ds_echo-N_part-mag<suffix>.nii and <prefix>_desc-mcpc3ds_mask.nii — the pipeline’s names. <suffix> is the inputs’ BIDS suffix (MEGRE when they carry none) Required.
--tes <TES>	Echo times in seconds, one per echo (read from the phase JSON sidecars when omitted)
--num-echoes <NUM_ECHOES>	Number of echoes when the file names carry no echo-N entity.
--sigma <X> <Y> <Z>	Smoothing sigma (voxels) for the per-coil phase offsets (default 10 10 5)
--hip-unwrapping <HIP_UNWRAPPING>	Unwrapping used on the coil-summed HIP phase. Values: romeo, laplacian, prelude, medi-region-growing. Default: romeo.
--prelude-labelling <PRELUDE_LABELLING>	PRELUDE: how the initial regions are labelled (default: auto) Values: auto, slices, volume.
--prelude-partitions <N>	PRELUDE: number of phase partitions the initial regions are formed from (default: 8)

qsmxt unwrap

Phase unwrapping

Subcommand	Description
romeo	ROMEO region-growing unwrapping
laplacian	Laplacian-based unwrapping
prelude	PRELUDE region-merging unwrapping
medi-region-growing	MEDI’s magnitude-guided region-growing unwrapping (takes no mask)

qsmxt unwrap romeo

ROMEO region-growing unwrapping

qsmxt unwrap romeo [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input wrapped phase NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output unwrapped phase NIfTI file. Required.
--magnitude <MAGNITUDE>	Magnitude image (improves quality)
--no-romeo-phase-coherence	ROMEO: disable phase coherence weights (default: on)
--no-romeo-phase-gradient-coherence	ROMEO: disable phase gradient coherence weights (default: on)
--no-romeo-phase-linearity	ROMEO: disable phase linearity weights (default: on; the romeo command line’s romeo3 weights)
--romeo-phase-linearity	ROMEO: enable phase linearity weights (default: on; for a –pipeline that turns them off)
--no-romeo-mag-coherence	ROMEO: disable magnitude coherence weights (default: on)
--romeo-mag-weight	ROMEO: enable magnitude weighting (default: off)
--no-romeo-mag-weight	ROMEO: disable magnitude weighting (overrides –romeo-mag-weight; kept for compatibility)
--romeo-mag-weight2	ROMEO: enable second magnitude weighting (flow-artifact penalty; default: off)
--romeo-bestpath	ROMEO: use Best-path (Abdul-Rahman) weights instead of ROMEO weights (default: off)
--romeo-template <ROMEO_TEMPLATE>	ROMEO: template echo index (1-indexed, only for template mode)
--romeo-individual	ROMEO: use individual per-echo unwrapping (default)
--no-romeo-individual	ROMEO: use template-based temporal unwrapping (disables individual mode)
--romeo-correct-global	ROMEO: correct global 2π offsets, of each echo and between echoes (default: on)
--no-romeo-correct-global	ROMEO: do not correct global 2π offsets.
--romeo-temporal-uncertain-unwrapping <ROMEO_TEMPORAL_UNCERTAIN_UNWRAPPING>	ROMEO: temporal uncertain-unwrapping quality threshold [0,1] (0 disables; default: 0.5)
--romeo-max-seeds <ROMEO_MAX_SEEDS>	ROMEO: maximum number of seed regions (default: 1)
--romeo-merge-regions	ROMEO: merge neighboring regions after unwrapping (default: off)
--romeo-correct-regions	ROMEO: correct each region’s median to nearest 0 (default: off)
--romeo-wrap-addition <ROMEO_WRAP_ADDITION>	ROMEO: additional phase tolerance beyond π for neighbor differences [0,π] (default: 0.0)

qsmxt unwrap laplacian

Laplacian-based unwrapping

qsmxt unwrap laplacian --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input wrapped phase NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output unwrapped phase NIfTI file. Required.

qsmxt unwrap prelude

PRELUDE region-merging unwrapping

qsmxt unwrap prelude [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input wrapped phase NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output unwrapped phase NIfTI file. Required.
--prelude-labelling <PRELUDE_LABELLING>	PRELUDE: how the initial regions are labelled (default: auto) Values: auto, slices, volume.
--prelude-partitions <N>	PRELUDE: number of phase partitions the initial regions are formed from (default: 8)

qsmxt unwrap medi-region-growing

MEDI’s magnitude-guided region-growing unwrapping (takes no mask)

qsmxt unwrap medi-region-growing --magnitude <MAGNITUDE> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input wrapped phase NIfTI file. Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI file (3D): the order in which the volume is unwrapped. Required. Required.
-o, --output <OUTPUT>	Output unwrapped phase NIfTI file. Required.

qsmxt fieldmap

B0 field mapping from multi-echo phase (NIfTI in, ppm out)

Subcommand	Description
romeo	ROMEO-based multi-echo field mapping (phase offset removal + unwrap + B0 fit)
laplacian	Laplacian-based multi-echo field mapping (per-echo unwrap + echo combination)
prelude	PRELUDE-based multi-echo field mapping (phase offset removal + per-echo unwrap + B0 fit)
medi-region-growing	MEDI region-growing field mapping (phase offset removal + per-echo unwrap + B0 fit; with --b0-estimation complex-fit, MEDI’s field map). Needs --magnitude

qsmxt fieldmap romeo

ROMEO-based multi-echo field mapping (phase offset removal + unwrap + B0 fit)

qsmxt fieldmap romeo [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input multi-echo wrapped phase NIfTI file (4D: x,y,z,echo) Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output B0 field map NIfTI file (in ppm) Required.
--magnitude <MAGNITUDE>	Multi-echo magnitude NIfTI file (4D; improves weighting)
--tes <TES>	Echo times in seconds (space-separated, one per echo)
--b0 <B0>	Field strength in Tesla (alias: –field-strength)
--params <PARAMS>	JSON params file: {“TE”:[..seconds], “B0”:<tesla>, “voxel_size”:[..], “B0_dir”:[..]}. –tes / –b0 override values from this file when both are present.
--b0-estimation <B0_ESTIMATION>	B0 field estimation method. Values: weighted-avg, linear-fit, robust-fit, qsmart-echofit, complex-fit. Default: weighted-avg.
--qsmart-fit-threshold-mode <QSMART_FIT_THRESHOLD_MODE>	--b0-estimation qsmart-echofit: how the fit residual at which a voxel is marked unreliable is set: a percentile of the blurred residual over the volume (--qsmart-fit-percentile), or a fixed residual (--qsmart-fit-threshold; default, QSMART v1.0’s 40). Without it, giving one of those two selects its mode. Values: percentile, absolute.
--qsmart-fit-percentile <QSMART_FIT_PERCENTILE>	--b0-estimation qsmart-echofit: the percentile of the blurred fit residual, in (0, 100], above which a voxel is marked unreliable (default 99)
--qsmart-fit-threshold <QSMART_FIT_THRESHOLD>	--b0-estimation qsmart-echofit: a fixed blurred fit residual above which a voxel is marked unreliable, instead of a percentile (40 when the mode is absolute and none is given)
--reliability-output <RELIABILITY_OUTPUT>	--b0-estimation qsmart-echofit: also save its reliability map (1 = reliable) here, for qsmxt qsmart --reliability.
--field-mask-output <FIELD_MASK_OUTPUT>	Also save the mask the field map is defined in here: the mask, less the voxels where phase offset removal could not estimate the phase offset (the field is 0 there). Pass it to qsmxt bgremove and the inversions in place of the mask.
--b0-weight-type <B0_WEIGHT_TYPE>	B0 weighted-averaging weight type. assumed-decay gives each echo’s phase a fixed weight TE·exp(−TE/T2*) from the echo times and an assumed T2* (no T2* map, no magnitude; Wang et al. 2022). phase-snr is generally preferable with magnitude. Values: phase-snr, phase-var, average, tes, mag, assumed-decay. Default: phase-snr.
--b0-weight-assumed-t2star <MS>	T2* in ms assumed by --b0-weight-type assumed-decay (no T2* map is used) Default: 40.
--linear-fit-estimate-offset <LINEAR_FIT_ESTIMATE_OFFSET>	Linear and robust fit: estimate a phase offset term (default: true). The robust fit needs 4 echoes with it, 3 without.
--robust-fit-loss <ROBUST_FIT_LOSS>	Motion-robust fit (--b0-estimation robust-fit): loss turning residuals into weights. tukey (default) rejects echoes past the tuning radius; huber caps but never rejects. Values: tukey, huber.
--robust-fit-tuning <ROBUST_FIT_TUNING>	Motion-robust fit: tuning radius in robust-scale units (default: the loss’s conventional value, Tukey 4.685 / Huber 1.345). Smaller is more aggressive.
--robust-fit-iterations <ROBUST_FIT_ITERATIONS>	Motion-robust fit: IRLS passes (default 3; 0 = report only, the field is the linear fit’s)
--robust-fit-downweight-level <ROBUST_FIT_DOWNWEIGHT_LEVEL>	Motion-robust fit (reporting only): a robust weight below this counts as down-weighted (default 0.75; in (0, 1])
--robust-fit-flag-ratio <ROBUST_FIT_FLAG_RATIO>	Motion-robust fit (reporting only): flag an echo whose outlier score exceeds this many times the best echo’s (default 3; at least 1)
--complex-fit-reltol <COMPLEX_FIT_RELTOL>	Complex fit (--b0-estimation complex-fit): stop once the update is this fraction of the first estimate, over the whole volume (default 1e-4)
--complex-fit-max-iter <COMPLEX_FIT_MAX_ITER>	Complex fit: Gauss-Newton iterations at most (default 30)
--complex-fit-output <FILE>	--b0-estimation complex-fit: also save the fit itself here, before spatial unwrapping: the phase accrued per echo spacing (radians, wrapped to [-π, π]; MEDI’s iFreq_raw)
--complex-fit-noise-output <FILE>	--b0-estimation complex-fit: also save the fit’s noise map here: the standard deviation of the phase per echo spacing, in radians (MEDI’s N_std)
--phase-offset-removal <PHASE_OFFSET_REMOVAL>	Enable phase offset removal for multi-echo data (default: true)
--phase-offset-sigma-mm <PHASE_OFFSET_SIGMA_MM>	Phase offset removal smoothing sigma (3 values, in mm; default: 7 7 7, the ROMEO command line’s)
--bipolar-correction	Enable bipolar gradient correction (requires >= 3 echoes)
--no-romeo-phase-coherence	ROMEO: disable phase coherence weights (default: on)
--no-romeo-phase-gradient-coherence	ROMEO: disable phase gradient coherence weights (default: on)
--no-romeo-phase-linearity	ROMEO: disable phase linearity weights (default: on; the romeo command line’s romeo3 weights)
--romeo-phase-linearity	ROMEO: enable phase linearity weights (default: on; for a –pipeline that turns them off)
--no-romeo-mag-coherence	ROMEO: disable magnitude coherence weights (default: on)
--romeo-mag-weight	ROMEO: enable magnitude weighting (default: off)
--no-romeo-mag-weight	ROMEO: disable magnitude weighting (overrides –romeo-mag-weight; kept for compatibility)
--romeo-mag-weight2	ROMEO: enable second magnitude weighting (flow-artifact penalty; default: off)
--romeo-bestpath	ROMEO: use Best-path (Abdul-Rahman) weights instead of ROMEO weights (default: off)
--romeo-template <ROMEO_TEMPLATE>	ROMEO: template echo index (1-indexed, only for template mode)
--romeo-individual	ROMEO: use individual per-echo unwrapping (default)
--no-romeo-individual	ROMEO: use template-based temporal unwrapping (disables individual mode)
--romeo-correct-global	ROMEO: correct global 2π offsets, of each echo and between echoes (default: on)
--no-romeo-correct-global	ROMEO: do not correct global 2π offsets.
--romeo-temporal-uncertain-unwrapping <ROMEO_TEMPORAL_UNCERTAIN_UNWRAPPING>	ROMEO: temporal uncertain-unwrapping quality threshold [0,1] (0 disables; default: 0.5)
--romeo-max-seeds <ROMEO_MAX_SEEDS>	ROMEO: maximum number of seed regions (default: 1)
--romeo-merge-regions	ROMEO: merge neighboring regions after unwrapping (default: off)
--romeo-correct-regions	ROMEO: correct each region’s median to nearest 0 (default: off)
--romeo-wrap-addition <ROMEO_WRAP_ADDITION>	ROMEO: additional phase tolerance beyond π for neighbor differences [0,π] (default: 0.0)

qsmxt fieldmap laplacian

Laplacian-based multi-echo field mapping (per-echo unwrap + echo combination)

qsmxt fieldmap laplacian [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input multi-echo wrapped phase NIfTI file (4D: x,y,z,echo) Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output B0 field map NIfTI file (in ppm) Required.
--magnitude <MAGNITUDE>	Multi-echo magnitude NIfTI file (4D; improves weighting)
--tes <TES>	Echo times in seconds (space-separated, one per echo)
--b0 <B0>	Field strength in Tesla (alias: –field-strength)
--params <PARAMS>	JSON params file: {“TE”:[..seconds], “B0”:<tesla>, “voxel_size”:[..], “B0_dir”:[..]}. –tes / –b0 override values from this file when both are present.
--b0-estimation <B0_ESTIMATION>	B0 field estimation method. Values: weighted-avg, linear-fit, robust-fit, qsmart-echofit, complex-fit. Default: weighted-avg.
--qsmart-fit-threshold-mode <QSMART_FIT_THRESHOLD_MODE>	--b0-estimation qsmart-echofit: how the fit residual at which a voxel is marked unreliable is set: a percentile of the blurred residual over the volume (--qsmart-fit-percentile), or a fixed residual (--qsmart-fit-threshold; default, QSMART v1.0’s 40). Without it, giving one of those two selects its mode. Values: percentile, absolute.
--qsmart-fit-percentile <QSMART_FIT_PERCENTILE>	--b0-estimation qsmart-echofit: the percentile of the blurred fit residual, in (0, 100], above which a voxel is marked unreliable (default 99)
--qsmart-fit-threshold <QSMART_FIT_THRESHOLD>	--b0-estimation qsmart-echofit: a fixed blurred fit residual above which a voxel is marked unreliable, instead of a percentile (40 when the mode is absolute and none is given)
--reliability-output <RELIABILITY_OUTPUT>	--b0-estimation qsmart-echofit: also save its reliability map (1 = reliable) here, for qsmxt qsmart --reliability.
--field-mask-output <FIELD_MASK_OUTPUT>	Also save the mask the field map is defined in here: the mask, less the voxels where phase offset removal could not estimate the phase offset (the field is 0 there). Pass it to qsmxt bgremove and the inversions in place of the mask.
--b0-weight-type <B0_WEIGHT_TYPE>	B0 weighted-averaging weight type. assumed-decay gives each echo’s phase a fixed weight TE·exp(−TE/T2*) from the echo times and an assumed T2* (no T2* map, no magnitude; Wang et al. 2022). phase-snr is generally preferable with magnitude. Values: phase-snr, phase-var, average, tes, mag, assumed-decay. Default: phase-snr.
--b0-weight-assumed-t2star <MS>	T2* in ms assumed by --b0-weight-type assumed-decay (no T2* map is used) Default: 40.
--linear-fit-estimate-offset <LINEAR_FIT_ESTIMATE_OFFSET>	Linear and robust fit: estimate a phase offset term (default: true). The robust fit needs 4 echoes with it, 3 without.
--robust-fit-loss <ROBUST_FIT_LOSS>	Motion-robust fit (--b0-estimation robust-fit): loss turning residuals into weights. tukey (default) rejects echoes past the tuning radius; huber caps but never rejects. Values: tukey, huber.
--robust-fit-tuning <ROBUST_FIT_TUNING>	Motion-robust fit: tuning radius in robust-scale units (default: the loss’s conventional value, Tukey 4.685 / Huber 1.345). Smaller is more aggressive.
--robust-fit-iterations <ROBUST_FIT_ITERATIONS>	Motion-robust fit: IRLS passes (default 3; 0 = report only, the field is the linear fit’s)
--robust-fit-downweight-level <ROBUST_FIT_DOWNWEIGHT_LEVEL>	Motion-robust fit (reporting only): a robust weight below this counts as down-weighted (default 0.75; in (0, 1])
--robust-fit-flag-ratio <ROBUST_FIT_FLAG_RATIO>	Motion-robust fit (reporting only): flag an echo whose outlier score exceeds this many times the best echo’s (default 3; at least 1)
--complex-fit-reltol <COMPLEX_FIT_RELTOL>	Complex fit (--b0-estimation complex-fit): stop once the update is this fraction of the first estimate, over the whole volume (default 1e-4)
--complex-fit-max-iter <COMPLEX_FIT_MAX_ITER>	Complex fit: Gauss-Newton iterations at most (default 30)
--complex-fit-output <FILE>	--b0-estimation complex-fit: also save the fit itself here, before spatial unwrapping: the phase accrued per echo spacing (radians, wrapped to [-π, π]; MEDI’s iFreq_raw)
--complex-fit-noise-output <FILE>	--b0-estimation complex-fit: also save the fit’s noise map here: the standard deviation of the phase per echo spacing, in radians (MEDI’s N_std)
--laplacian-solver <LAPLACIAN_SOLVER>	Poisson solver: dct (unweighted least squares, Neumann/DCT solve) or fft (Schofield & Zhu, FFT solve on a zero-padded volume) Values: dct, fft, mirror. Default: dct.
--laplacian-fft-pad <VOXELS>	Zero-padding in voxels per side for --laplacian-solver fft: one value for every axis, or three (x y z) Default: 64.
--laplacian-unmasked	Unwrap the raw phase over the whole volume, instead of the phase set to zero outside the mask, as QSMART does. With --b0-estimation qsmart-echofit the fit then covers the whole volume (the saved field map is still set to zero outside the mask)

qsmxt fieldmap prelude

PRELUDE-based multi-echo field mapping (phase offset removal + per-echo unwrap + B0 fit)

qsmxt fieldmap prelude [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input multi-echo wrapped phase NIfTI file (4D: x,y,z,echo) Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output B0 field map NIfTI file (in ppm) Required.
--magnitude <MAGNITUDE>	Multi-echo magnitude NIfTI file (4D; improves weighting)
--tes <TES>	Echo times in seconds (space-separated, one per echo)
--b0 <B0>	Field strength in Tesla (alias: –field-strength)
--params <PARAMS>	JSON params file: {“TE”:[..seconds], “B0”:<tesla>, “voxel_size”:[..], “B0_dir”:[..]}. –tes / –b0 override values from this file when both are present.
--b0-estimation <B0_ESTIMATION>	B0 field estimation method. Values: weighted-avg, linear-fit, robust-fit, qsmart-echofit, complex-fit. Default: weighted-avg.
--qsmart-fit-threshold-mode <QSMART_FIT_THRESHOLD_MODE>	--b0-estimation qsmart-echofit: how the fit residual at which a voxel is marked unreliable is set: a percentile of the blurred residual over the volume (--qsmart-fit-percentile), or a fixed residual (--qsmart-fit-threshold; default, QSMART v1.0’s 40). Without it, giving one of those two selects its mode. Values: percentile, absolute.
--qsmart-fit-percentile <QSMART_FIT_PERCENTILE>	--b0-estimation qsmart-echofit: the percentile of the blurred fit residual, in (0, 100], above which a voxel is marked unreliable (default 99)
--qsmart-fit-threshold <QSMART_FIT_THRESHOLD>	--b0-estimation qsmart-echofit: a fixed blurred fit residual above which a voxel is marked unreliable, instead of a percentile (40 when the mode is absolute and none is given)
--reliability-output <RELIABILITY_OUTPUT>	--b0-estimation qsmart-echofit: also save its reliability map (1 = reliable) here, for qsmxt qsmart --reliability.
--field-mask-output <FIELD_MASK_OUTPUT>	Also save the mask the field map is defined in here: the mask, less the voxels where phase offset removal could not estimate the phase offset (the field is 0 there). Pass it to qsmxt bgremove and the inversions in place of the mask.
--b0-weight-type <B0_WEIGHT_TYPE>	B0 weighted-averaging weight type. assumed-decay gives each echo’s phase a fixed weight TE·exp(−TE/T2*) from the echo times and an assumed T2* (no T2* map, no magnitude; Wang et al. 2022). phase-snr is generally preferable with magnitude. Values: phase-snr, phase-var, average, tes, mag, assumed-decay. Default: phase-snr.
--b0-weight-assumed-t2star <MS>	T2* in ms assumed by --b0-weight-type assumed-decay (no T2* map is used) Default: 40.
--linear-fit-estimate-offset <LINEAR_FIT_ESTIMATE_OFFSET>	Linear and robust fit: estimate a phase offset term (default: true). The robust fit needs 4 echoes with it, 3 without.
--robust-fit-loss <ROBUST_FIT_LOSS>	Motion-robust fit (--b0-estimation robust-fit): loss turning residuals into weights. tukey (default) rejects echoes past the tuning radius; huber caps but never rejects. Values: tukey, huber.
--robust-fit-tuning <ROBUST_FIT_TUNING>	Motion-robust fit: tuning radius in robust-scale units (default: the loss’s conventional value, Tukey 4.685 / Huber 1.345). Smaller is more aggressive.
--robust-fit-iterations <ROBUST_FIT_ITERATIONS>	Motion-robust fit: IRLS passes (default 3; 0 = report only, the field is the linear fit’s)
--robust-fit-downweight-level <ROBUST_FIT_DOWNWEIGHT_LEVEL>	Motion-robust fit (reporting only): a robust weight below this counts as down-weighted (default 0.75; in (0, 1])
--robust-fit-flag-ratio <ROBUST_FIT_FLAG_RATIO>	Motion-robust fit (reporting only): flag an echo whose outlier score exceeds this many times the best echo’s (default 3; at least 1)
--complex-fit-reltol <COMPLEX_FIT_RELTOL>	Complex fit (--b0-estimation complex-fit): stop once the update is this fraction of the first estimate, over the whole volume (default 1e-4)
--complex-fit-max-iter <COMPLEX_FIT_MAX_ITER>	Complex fit: Gauss-Newton iterations at most (default 30)
--complex-fit-output <FILE>	--b0-estimation complex-fit: also save the fit itself here, before spatial unwrapping: the phase accrued per echo spacing (radians, wrapped to [-π, π]; MEDI’s iFreq_raw)
--complex-fit-noise-output <FILE>	--b0-estimation complex-fit: also save the fit’s noise map here: the standard deviation of the phase per echo spacing, in radians (MEDI’s N_std)
--phase-offset-removal <PHASE_OFFSET_REMOVAL>	Enable phase offset removal for multi-echo data (default: true)
--phase-offset-sigma-mm <PHASE_OFFSET_SIGMA_MM>	Phase offset removal smoothing sigma (3 values, in mm; default: 7 7 7, the ROMEO command line’s)
--bipolar-correction	Enable bipolar gradient correction (requires >= 3 echoes)
--prelude-labelling <PRELUDE_LABELLING>	PRELUDE: how the initial regions are labelled (default: auto) Values: auto, slices, volume.
--prelude-partitions <N>	PRELUDE: number of phase partitions the initial regions are formed from (default: 8)

qsmxt fieldmap medi-region-growing

MEDI region-growing field mapping (phase offset removal + per-echo unwrap + B0 fit; with --b0-estimation complex-fit, MEDI’s field map). Needs --magnitude

qsmxt fieldmap medi-region-growing [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input multi-echo wrapped phase NIfTI file (4D: x,y,z,echo) Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output B0 field map NIfTI file (in ppm) Required.
--magnitude <MAGNITUDE>	Multi-echo magnitude NIfTI file (4D; improves weighting)
--tes <TES>	Echo times in seconds (space-separated, one per echo)
--b0 <B0>	Field strength in Tesla (alias: –field-strength)
--params <PARAMS>	JSON params file: {“TE”:[..seconds], “B0”:<tesla>, “voxel_size”:[..], “B0_dir”:[..]}. –tes / –b0 override values from this file when both are present.
--b0-estimation <B0_ESTIMATION>	B0 field estimation method. Values: weighted-avg, linear-fit, robust-fit, qsmart-echofit, complex-fit. Default: weighted-avg.
--qsmart-fit-threshold-mode <QSMART_FIT_THRESHOLD_MODE>	--b0-estimation qsmart-echofit: how the fit residual at which a voxel is marked unreliable is set: a percentile of the blurred residual over the volume (--qsmart-fit-percentile), or a fixed residual (--qsmart-fit-threshold; default, QSMART v1.0’s 40). Without it, giving one of those two selects its mode. Values: percentile, absolute.
--qsmart-fit-percentile <QSMART_FIT_PERCENTILE>	--b0-estimation qsmart-echofit: the percentile of the blurred fit residual, in (0, 100], above which a voxel is marked unreliable (default 99)
--qsmart-fit-threshold <QSMART_FIT_THRESHOLD>	--b0-estimation qsmart-echofit: a fixed blurred fit residual above which a voxel is marked unreliable, instead of a percentile (40 when the mode is absolute and none is given)
--reliability-output <RELIABILITY_OUTPUT>	--b0-estimation qsmart-echofit: also save its reliability map (1 = reliable) here, for qsmxt qsmart --reliability.
--field-mask-output <FIELD_MASK_OUTPUT>	Also save the mask the field map is defined in here: the mask, less the voxels where phase offset removal could not estimate the phase offset (the field is 0 there). Pass it to qsmxt bgremove and the inversions in place of the mask.
--b0-weight-type <B0_WEIGHT_TYPE>	B0 weighted-averaging weight type. assumed-decay gives each echo’s phase a fixed weight TE·exp(−TE/T2*) from the echo times and an assumed T2* (no T2* map, no magnitude; Wang et al. 2022). phase-snr is generally preferable with magnitude. Values: phase-snr, phase-var, average, tes, mag, assumed-decay. Default: phase-snr.
--b0-weight-assumed-t2star <MS>	T2* in ms assumed by --b0-weight-type assumed-decay (no T2* map is used) Default: 40.
--linear-fit-estimate-offset <LINEAR_FIT_ESTIMATE_OFFSET>	Linear and robust fit: estimate a phase offset term (default: true). The robust fit needs 4 echoes with it, 3 without.
--robust-fit-loss <ROBUST_FIT_LOSS>	Motion-robust fit (--b0-estimation robust-fit): loss turning residuals into weights. tukey (default) rejects echoes past the tuning radius; huber caps but never rejects. Values: tukey, huber.
--robust-fit-tuning <ROBUST_FIT_TUNING>	Motion-robust fit: tuning radius in robust-scale units (default: the loss’s conventional value, Tukey 4.685 / Huber 1.345). Smaller is more aggressive.
--robust-fit-iterations <ROBUST_FIT_ITERATIONS>	Motion-robust fit: IRLS passes (default 3; 0 = report only, the field is the linear fit’s)
--robust-fit-downweight-level <ROBUST_FIT_DOWNWEIGHT_LEVEL>	Motion-robust fit (reporting only): a robust weight below this counts as down-weighted (default 0.75; in (0, 1])
--robust-fit-flag-ratio <ROBUST_FIT_FLAG_RATIO>	Motion-robust fit (reporting only): flag an echo whose outlier score exceeds this many times the best echo’s (default 3; at least 1)
--complex-fit-reltol <COMPLEX_FIT_RELTOL>	Complex fit (--b0-estimation complex-fit): stop once the update is this fraction of the first estimate, over the whole volume (default 1e-4)
--complex-fit-max-iter <COMPLEX_FIT_MAX_ITER>	Complex fit: Gauss-Newton iterations at most (default 30)
--complex-fit-output <FILE>	--b0-estimation complex-fit: also save the fit itself here, before spatial unwrapping: the phase accrued per echo spacing (radians, wrapped to [-π, π]; MEDI’s iFreq_raw)
--complex-fit-noise-output <FILE>	--b0-estimation complex-fit: also save the fit’s noise map here: the standard deviation of the phase per echo spacing, in radians (MEDI’s N_std)
--phase-offset-removal <PHASE_OFFSET_REMOVAL>	Enable phase offset removal for multi-echo data (default: true)
--phase-offset-sigma-mm <PHASE_OFFSET_SIGMA_MM>	Phase offset removal smoothing sigma (3 values, in mm; default: 7 7 7, the ROMEO command line’s)
--bipolar-correction	Enable bipolar gradient correction (requires >= 3 echoes)

qsmxt bgremove

Background field removal

Subcommand	Description
vsharp	Variable radius SHARP (V-SHARP)
pdf	Projection onto Dipole Fields
lbv	Laplacian Boundary Value
ismv	Iterative Spherical Mean Value
sharp	Spherical Harmonic Array Reconstruction Procedure
resharp	Regularized SHARP (RESHARP) with Tikhonov regularization
harperella	HARPERELLA integrated phase unwrapping + background removal
iharperella	Improved HARPERELLA (iHARPERELLA)
bfrnet	BFRnet deep-learning background removal (weights downloaded on first use)
iqfm	iQFM deep-learning joint unwrapping and background removal — input is every echo’s WRAPPED phase (weights downloaded on first use)

qsmxt bgremove vsharp

Variable radius SHARP (V-SHARP)

qsmxt bgremove vsharp [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input total field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output local field NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--output-mask <OUTPUT_MASK>	Output eroded mask (for algorithms that erode)
--deconvolve <BOOL>	Undo the SMV high-pass with a truncated-SVD deconvolution (default: false)
--threshold <THRESHOLD>	Deconvolution (TSVD) threshold; only used with –deconvolve true.
--max-radius <MAX_RADIUS>	Largest SMV kernel radius, in mm.
--min-radius <MIN_RADIUS>	Smallest SMV kernel radius, in mm.
--radius-step <MM>	Step between SMV kernel radii, in mm (default: 1); the radii run from –max-radius down to –min-radius.
--grid <GRID>	Grid the background is estimated on: isotropic (default; spacing from –grid-spacing), native (the acquired grid), or finest (isotropic at the smallest voxel dimension, e.g. 0.8 mm for 0.8x0.8x3 mm; no axis is coarsened). The local field keeps native resolution. Values: native, finest, isotropic.
--grid-spacing <MM>	Isotropic working-grid spacing in mm (default 1); used with –grid isotropic.
--ball <BALL>	SMV ball: default (the voxels within r + 1/6 mm, centre excluded; a ball fits where all of it is in the mask) or qsmjl (QSM.jl’s: within r, centre included) Values: default, qsmjl.

qsmxt bgremove pdf

Projection onto Dipole Fields

qsmxt bgremove pdf [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... --magnitude <MAGNITUDE>... <INPUT>

Argument	Description
<INPUT>	Input total field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output local field NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--output-mask <OUTPUT_MASK>	Output eroded mask (for algorithms that erode)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--magnitude <MAGNITUDE>	Magnitude of every echo: one 4D NIfTI, or one 3D NIfTI per echo in echo order. Required.
--tol <TOL>	Tolerance, relative to the noise level.
--n-cg <N_CG>	Maximum conjugate-gradient iterations (default: 30)
--space <SPACE>	Dipole kernel: image (the FFT of the dipole sampled on the voxel grid) or kspace (the analytic k-space kernel) (default: image)
--n-pad <VOXELS>	Zero-padding around the mask’s bounding box, in voxels; 0 runs on the full volume (default: 40)

qsmxt bgremove lbv

Laplacian Boundary Value

qsmxt bgremove lbv [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... <INPUT>

Argument	Description
<INPUT>	Input total field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output local field NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--output-mask <OUTPUT_MASK>	Output eroded mask (for algorithms that erode)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--tol <TOL>	Tolerance: stop the coarsest grid’s relaxation when its summed absolute update falls below this, in radians over the echo spacing (default: 0.01)
--depth <LEVEL>	Coarsest multigrid level, at most 8 (default: log2 of the smallest dimension minus 3)
--peel <LAYERS>	Boundary layers removed from the mask beyond the first (default: 0)
--n1 <N1>	Relaxations on each grid before the coarser grid (default: 30)
--n2 <N2>	Relaxations on each grid after the coarser grid (default: 100)
--n3 <N3>	Relaxations on the finest grid at the end (default: 100)
--precision <PRECISION>	Arithmetic: double, or single as the original implementation (default: double)

qsmxt bgremove ismv

Iterative Spherical Mean Value

qsmxt bgremove ismv [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input total field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output local field NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--output-mask <OUTPUT_MASK>	Output eroded mask (for algorithms that erode)
--variant <VARIANT>	Implementation: wen2014 (the authors’ paper and code) or qsmjl (QSM.jl’s) (default: wen2014)
--tol <TOL>	Tolerance: the relative change between iterates it stops at (qsmjl: the residual relative to the first) (default: 5e-5)
--max-iter <MAX_ITER>	Max iterations (default: no limit)
--radius <RADIUS>	Kernel radius in mm (default: 6 × the smallest voxel dimension; qsmjl: 2 × the largest)

qsmxt bgremove sharp

Spherical Harmonic Array Reconstruction Procedure

qsmxt bgremove sharp [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input total field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output local field NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--output-mask <OUTPUT_MASK>	Output eroded mask (for algorithms that erode)
--threshold <THRESHOLD>	Deconvolution (TSVD) threshold (default: 0.03; qsmjl: 0.05)
--radius <RADIUS>	Sphere radius in mm (default: 4; qsmjl: 18 × the smallest voxel dimension)
--variant <VARIANT>	Implementation: sepia (SEPIA’s SHARP.m, the default) or qsmjl (QSM.jl’s sharp, with its own defaults) Values: sepia, qsmjl.

qsmxt bgremove resharp

Regularized SHARP (RESHARP) with Tikhonov regularization

qsmxt bgremove resharp [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input total field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output local field NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--output-mask <OUTPUT_MASK>	Output eroded mask (for algorithms that erode)
--radius <RADIUS>	SMV kernel radius in mm.
--tik-reg <TIK_REG>	Tikhonov regularization parameter.
--tol <TOL>	Conjugate gradients squared (CGS) relative tolerance.
--max-iter <MAX_ITER>	Maximum CGS iterations.

qsmxt bgremove harperella

HARPERELLA integrated phase unwrapping + background removal

qsmxt bgremove harperella [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI: a wrapped echo in rad (with –te and –field-strength) or a total field in ppm. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output local field NIfTI file. Required.
--output-mask <OUTPUT_MASK>	Output eroded mask (for algorithms that erode)
--radius <RADIUS>	SMV kernel radius in mm.
--max-iter <MAX_ITER>	Maximum iterations.
--tol <TOL>	Convergence tolerance.
--te <TE>	Echo time (s) of a wrapped-phase input. With –field-strength, the input is one wrapped echo in rad (the published method); without both, it is an unwrapped total field in ppm. The local field is written in ppm either way.
--field-strength <FIELD_STRENGTH>	B0 field strength (T) of a wrapped-phase input — see –te.
--pad <X> <Y> <Z>	Zero-padding in voxels per side, X Y Z (default: 0 0 0)
--rim <RIM>	Boundary rim: whole (zero every Laplacian value in the rim; default) or sti (zero the 20 % largest, as STI Suite does)

qsmxt bgremove iharperella

Improved HARPERELLA (iHARPERELLA)

qsmxt bgremove iharperella [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input NIfTI: a wrapped echo in rad (with –te and –field-strength) or a total field in ppm. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output local field NIfTI file. Required.
--output-mask <OUTPUT_MASK>	Output eroded mask (for algorithms that erode)
--max-iter <MAX_ITER>	Maximum iterations.
--tol <TOL>	Convergence tolerance.
--te <TE>	Echo time (s) of a wrapped-phase input. With –field-strength, the input is one wrapped echo in rad, unwrapped by HARPERELLA’s Laplacian; without both, it is an unwrapped total field in ppm, which iHARPERELLA takes as it is. The local field is written in ppm either way.
--field-strength <FIELD_STRENGTH>	B0 field strength (T) of a wrapped-phase input — see –te.
--pad <X> <Y> <Z>	Zero-padding in voxels per side, X Y Z (default: 0 0 0)
--rim <RIM>	Boundary rim: whole (zero every Laplacian value in the rim) or sti (zero the 20 % largest, as STI Suite does; default)

qsmxt bgremove bfrnet

BFRnet deep-learning background removal (weights downloaded on first use)

qsmxt bgremove bfrnet [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input total field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output local field NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--output-mask <OUTPUT_MASK>	Output eroded mask (for algorithms that erode)

qsmxt bgremove iqfm

iQFM deep-learning joint unwrapping and background removal — input is every echo’s WRAPPED phase (weights downloaded on first use)

qsmxt bgremove iqfm [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... <INPUT>

Argument	Description
<INPUT>	Wrapped phase NIfTI file in radians: 4D (x,y,z,echo), or 3D for one echo. Required.
-m, --mask <MASK>	Binary brain mask NIfTI file. Required.
-o, --output <OUTPUT>	Output NIfTI file in ppm (susceptibility for iqsm/iqsm-plus, the local field for iqfm) Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI file with the same echoes as the phase (4D), to weight the echo combination by magnitude; without it the echoes are weighted by TE² alone.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--echo-weighting <ECHO_WEIGHTING>	Echo weighting of the multi-echo combination: mag2-te2 weights each echo by (magnitude × TE)², the inverse of its noise variance, as in the authors’ Python code; mag-te2 by magnitude × TE², as in the authors’ MATLAB code. Values: mag-te2, mag2-te2. Default: mag2-te2.

qsmxt invert

Dipole inversion

Subcommand	Description
rts	Rapid Two-Step (RTS)
tv	Total Variation (ADMM)
tkd	Truncated K-space Division
tsvd	Truncated Singular Value Decomposition
tgv	Total Generalized Variation
tikhonov	Tikhonov regularization
nltv	Nonlinear Total Variation
medi	Morphology Enabled Dipole Inversion
tfi	Total Field Inversion (preconditioned). Input is the TOTAL field (not a local field)
ilsqr	Iterative Least Squares QR
star	STAR-QSM (Streaking Artifact Reduction): sparse inversion from a k-space initial guess
lsqr	Minimally regularised LSQR
heidi	LSQR+HEIDI (Homogeneity Enabled Incremental Dipole Inversion), or HEIDI on another inversion’s map (--seed)
ndi	Nonlinear Dipole Inversion
fansi	FANSI Nonlinear Total Variation
fansi-tgv	FANSI Nonlinear Total Generalized Variation
l1qsm	L1 Data-Fidelity QSM
whqsm	Weak-Harmonic QSM
hdqsm	Hybrid Data-Fidelity QSM
amp-pe	AMP-PE (Approximate Message Passing with Parameter Estimation)
xqsm	xQSM deep-learning dipole inversion (weights downloaded on first use)
qsmnet	QSMnet deep-learning dipole inversion (weights downloaded on first use)
qsmnet-plus	QSMnet+ deep-learning dipole inversion (weights downloaded on first use)
qsmnetme	QSMnetME deep-learning dipole inversion, the χ-separation toolbox’s QSM network (weights downloaded on first use)
autoqsm	AutoQSM single-step reconstruction — input is the TOTAL field (weights downloaded on first use)
qsmgan	QSMGAN deep-learning dipole inversion (weights downloaded on first use)
ir2qsm	IR2QSM deep-learning dipole inversion (weights downloaded on first use)
lpcnn	LPCNN deep-learning dipole inversion (weights downloaded on first use)
modl-qsm	MoDL-QSM deep-learning dipole inversion → χ33/STI component (weights downloaded on first use)
nextqsm	NeXtQSM single-step reconstruction — input is the TOTAL field (weights downloaded on first use)
iqsm	iQSM single-step reconstruction — input is every echo’s WRAPPED phase (weights downloaded on first use)
iqsm-plus	iQSM+ orientation-adaptive single-step reconstruction — input is every echo’s WRAPPED phase; echoes are combined with (magnitude × TE)² weights (weights downloaded on first use)
cosmos	COSMOS — multi-orientation least squares over 2+ co-registered orientations
sti	STI (susceptibility tensor imaging) — 6+ co-registered orientations to a rank-2 tensor

qsmxt invert rts

Rapid Two-Step (RTS)

qsmxt invert rts [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--delta <DELTA>	Delta parameter.
--mu <MU>	Mu parameter.
--tol <TOL>	Tolerance.
--rho <RHO>	Rho (ADMM penalty)
--max-iter <MAX_ITER>	Max iterations.
--lsmr-iter <LSMR_ITER>	LSMR iterations.
--tau <TAU>	Rho is raised when the primal residual falls by less than this factor (tau)
--gamma <GAMMA>	The factor rho is raised by (gamma)
--pad <X> <Y> <Z>	Zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)

qsmxt invert tv

Total Variation (ADMM)

qsmxt invert tv [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--lambda <LAMBDA>	Lambda parameter.
--rho <RHO>	Rho (ADMM penalty), or auto for 100 × lambda (default: auto)
--tol <TOL>	Tolerance.
--max-iter <MAX_ITER>	Max iterations.
--pad <X> <Y> <Z>	Zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)

qsmxt invert tkd

Truncated K-space Division

qsmxt invert tkd [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--threshold <THRESHOLD>	Threshold.
--pad <X> <Y> <Z>	Zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)

qsmxt invert tsvd

Truncated Singular Value Decomposition

qsmxt invert tsvd [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--threshold <THRESHOLD>	Threshold.
--pad <X> <Y> <Z>	Zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)

qsmxt invert tgv

Total Generalized Variation

qsmxt invert tgv [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--iterations <ITERATIONS>	Iterations (default: from the voxel size, as the reference does)
--erosions <EROSIONS>	Erosions.
--alpha1 <ALPHA1>	Alpha1 (first-order weight), in radians at the echo time (–te) and the field strength.
--alpha0 <ALPHA0>	Alpha0 (second-order weight), in radians at the echo time (–te) and the field strength.
--step-size <STEP_SIZE>	Primal step size multiplier.
--te <TE>	Echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--reference-divergence	Use QuantitativeSusceptibilityMappingTGV.jl’s divergence in the w update rather than the exact adjoint of the symmetric gradient. Changes the result only on anisotropic voxels; with –reference-edge-erosion and –reference-output-box it reproduces that package, for comparison with it.
--reference-edge-erosion	Erode the mask as QuantitativeSusceptibilityMappingTGV.jl does, not from the array’s edge. Changes the result only where the mask reaches an array face.
--reference-output-box	Return χ over QuantitativeSusceptibilityMappingTGV.jl’s whole output box, not only inside the eroded mask: adds the rim of up to two voxels just outside it, where no data constrain χ.

qsmxt invert tikhonov

Tikhonov regularization

qsmxt invert tikhonov [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--lambda <LAMBDA>	Lambda.
--pad <X> <Y> <Z>	Zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)

qsmxt invert nltv

Nonlinear Total Variation

qsmxt invert nltv [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... --magnitude <MAGNITUDE>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI, the data weight (required): one 3D volume, a 4D one, or one per echo; several echoes are RSS-combined. Required.
--lambda <LAMBDA>	Lambda.
--rho <RHO>	Rho (ADMM penalty), or auto for 100 × lambda (default: auto)
--mu <MU>	Mu (penalty parameter)
--tol <TOL>	Tolerance.
--max-iter <MAX_ITER>	Max iterations.
--tol-newton <TOL_NEWTON>	Convergence tolerance of the Newton iterations of the data-term subproblem.
--newton-iter <NEWTON_ITER>	Maximum Newton iterations of the data-term subproblem.
--te <TE>	Echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--pad <X> <Y> <Z>	Zero-padding before the FFTs in voxels, X Y Z: at least this many, rounded up to an FFT-efficient size; -1 leaves an axis unpadded (default: 0 0 0)

qsmxt invert medi

Morphology Enabled Dipole Inversion

qsmxt invert medi [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... --magnitude <MAGNITUDE>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--magnitude <MAGNITUDE>	Magnitude of every echo: one 4D NIfTI, or one 3D NIfTI per echo in echo order. Required.
--lambda <LAMBDA>	Lambda: the data-fidelity weight (larger fits the field more closely)
--merit <MERIT>	Enable MERIT weighting.
--csf <CSF>	MEDI+0 (Liu et al. 2018): zero-reference χ to the ventricular CSF, found from R2* of the echo magnitudes, and keep it uniform there (default: true). Needs at least 3 echoes; with fewer, MEDI runs without it.
--lambda-csf <LAMBDA_CSF>	MEDI+0: weight of the CSF uniformity term (default: 100)
--csf-r2s-threshold <CSF_R2S_THRESHOLD>	MEDI+0: R2* in Hz below which a voxel is a CSF candidate (default: 5)
--csf-erode <CSF_ERODE>	MEDI+0: erode the mask by a 10 mm sphere before finding the CSF (default: true)
--zeropad <X> <Y> <Z>	Voxels of zeros appended at the end of x, y and z before the inversion; the result is cropped back (default: 0 0 0)
--smv <SMV>	High-pass the local field with a spherical mean value filter, removing residual background near the brain edge (MEDI-SMV) [default: on].
--smv-radius <SMV_RADIUS>	SMV radius in mm (also mSMV’s)
--smv-shrink-mask <SMV_SHRINK_MASK>	With SMV, shrink the mask to where the SMV of the mask exceeds 0.999; false keeps the full mask, high-passing the field over it [default: on].
--msmv	MEDI-mSMV (Roberts et al. 2024): filter the local field by maximum spherical mean value first, and fit it with the SMV-filtered dipole kernel. Needs at least 3 echoes.
--msmv-maxk <MSMV_MAXK>	MEDI-mSMV: the boundary correction makes at most this many passes, less one.
--msmv-vessel-radius <MSMV_VESSEL_RADIUS>	MEDI-mSMV: largest vessel thickness of the vessel mask, in voxels (default: 15, or 8 × the largest voxel dimension if that is more)
--no-msmv-prefilter	MEDI-mSMV: no SMV prefilter before mSMV. Only consistent for a field already high-passed by the same SMV; after PDF, V-SHARP or LBV it diverges (expert)
--data-weighting <DATA_WEIGHTING>	Data weighting mode (0=uniform, 1=SNR)
--percentage <PERCENTAGE>	Edge percentage (0.0-1.0)
--cg-tol <CG_TOL>	CG tolerance.
--cg-max-iter <CG_MAX_ITER>	CG max iterations.
--max-iter <MAX_ITER>	Max outer iterations.
--tol <TOL>	Outer tolerance.
--te <TE>	Echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)

qsmxt invert tfi

Total Field Inversion (preconditioned). Input is the TOTAL field (not a local field)

qsmxt invert tfi [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... --magnitude <MAGNITUDE>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--magnitude <MAGNITUDE>	Magnitude of every echo: one 4D NIfTI, or one 3D NIfTI per echo in echo order. Required.
--lambda <LAMBDA>	Lambda: the data-fidelity weight (larger fits the field more closely)
--precond <PRECOND>	Preconditioner value (susceptibility scaling outside the mask)
--merit <MERIT>	Enable MERIT weighting (default: true)
--linear <LINEAR>	Fit the field linearly, as the TFI paper does (default: true); false uses MEDI’s exponential data term.
--data-weighting <DATA_WEIGHTING>	Data weighting mode (0=uniform, 1=SNR)
--percentage <PERCENTAGE>	Edge percentage (0.0-1.0)
--cg-tol <CG_TOL>	CG tolerance.
--cg-max-iter <CG_MAX_ITER>	CG max iterations.
--max-iter <MAX_ITER>	Max outer iterations.
--tol <TOL>	Outer tolerance.
--te <TE>	Echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)
--kernel <KERNEL>	Dipole kernel on a grid with an odd number of voxels along an axis: default samples the dipole on the Fourier grid; original samples it half a frequency step off, as the original implementation does, for comparison with it. Identical on even grids (default: default) Values: default, original.
--precision <PRECISION>	Arithmetic: double, or single as the original implementation runs, for comparison with it (default: double) Values: double, single.

qsmxt invert ilsqr

Iterative Least Squares QR

qsmxt invert ilsqr [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--tol <TOL>	Tolerance.
--max-iter <MAX_ITER>	Max iterations of each LSQR solve (default: 100)
--pad-mm <MM>	Zero-padding in mm per side.
--precision <PRECISION>	Arithmetic of the two LSQR solves: double, or single as the original implementation (default: double)

qsmxt invert star

STAR-QSM (Streaking Artifact Reduction): sparse inversion from a k-space initial guess

qsmxt invert star [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--pad <X> <Y> <Z>	Zero-padding in voxels per side, X Y Z (default: 8 8 20)
--te <TE>	Echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: sum)

qsmxt invert lsqr

Minimally regularised LSQR

qsmxt invert lsqr [OPTIONS] --mask <MASK> --output <OUTPUT> --magnitude <MAGNITUDE>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--magnitude <MAGNITUDE>	Magnitude NIfTI, LSQR’s row weight as the HEIDI package’s (its scale does not change the solution): one 3D volume, a 4D one, or one per echo; several echoes are RSS-combined. Required.
--b0 <B0>	B0 field strength in Tesla — sets the default residual weighting. Default: 3.
--lsqr-residual-weighting <LSQR_RESIDUAL_WEIGHTING>	LSQR residual-field weight w — the only thing regularising the solve. Omit to use qsm-core’s field-strength-scaled default; 0 drops the residual term and streaks badly.
--no-lsqr-global-offset	Do not fit a global field offset alongside χ.
--lsqr-tol <LSQR_TOL>	LSQR convergence tolerance.
--lsqr-max-iter <LSQR_MAX_ITER>	LSQR max iterations.

qsmxt invert heidi

LSQR+HEIDI (Homogeneity Enabled Incremental Dipole Inversion), or HEIDI on another inversion’s map (--seed)

qsmxt invert heidi [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI, the row weight of the LSQR seed as the HEIDI package’s (not needed with –seed): one 3D volume, a 4D one, or one per echo; several echoes are RSS-combined.
--seed <SEED>	A χ map (ppm) another inversion made of this field, to fill the cone of instead of running LSQR (--lsqr-* and --magnitude then do nothing). HEIDI suits a seed with little or no smoothing (TKD and TSVD with a threshold no larger than the cone threshold, small-λ Tikhonov, NDI, RTS); a method that smooths its result also changes the parts of the data HEIDI keeps as they are, and HEIDI cannot undo that.
--lsqr-residual-weighting <LSQR_RESIDUAL_WEIGHTING>	LSQR residual-field weight w — the only thing regularising the solve. Omit to use qsm-core’s field-strength-scaled default; 0 drops the residual term and streaks badly.
--no-lsqr-global-offset	Do not fit a global field offset alongside χ.
--lsqr-tol <LSQR_TOL>	LSQR convergence tolerance.
--lsqr-max-iter <LSQR_MAX_ITER>	LSQR max iterations.
--heidi-cone-threshold <HEIDI_CONE_THRESHOLD>	|D(k)| above which a coefficient is well-conditioned and kept from the seed map.
--heidi-gradient-threshold <HEIDI_GRADIENT_THRESHOLD>	Field-gradient threshold (ppm/mm); below it a direction reads as homogeneous.
--no-heidi-laplacian-correction	Do not open the homogeneity masks where the field’s Laplacian is large.
--heidi-laplacian-threshold <HEIDI_LAPLACIAN_THRESHOLD>	Field-Laplacian threshold (ppm per voxel²)
--heidi-gradient-mask-floor <HEIDI_GRADIENT_MASK_FLOOR>	TV weight kept where every direction reads as an edge (0 reproduces upstream’s shipped binarised masks)
--heidi-continuation-steps <HEIDI_CONTINUATION_STEPS>	NESTA continuation steps on μ.
--heidi-inner-iterations <HEIDI_INNER_ITERATIONS>	Accelerated-gradient iterations per continuation step.
--heidi-mu-min <HEIDI_MU_MIN>	Final (smallest) smoothing parameter μ.
--heidi-tol <HEIDI_TOL>	Relative objective change that ends an inner loop early (the default runs every iteration)
--no-heidi-denoise	Skip the anisotropic-diffusion smoothing of the field before its gradients are taken.
--heidi-denoise-iterations <HEIDI_DENOISE_ITERATIONS>	Anisotropic-diffusion iterations.
--heidi-denoise-time-step <HEIDI_DENOISE_TIME_STEP>	Anisotropic-diffusion time step.
--heidi-denoise-conductance <HEIDI_DENOISE_CONDUCTANCE>	Anisotropic-diffusion conductance.
--heidi-denoise-reference-itk	ITK 3.20’s float arithmetic for the denoising, as the authors’ code runs it (for comparison with it; the default computes its conductance scale in double precision)
--no-heidi-denoise-reference-itk	The denoising in double precision, after a pipeline that turns ITK’s float arithmetic on (the HEIDI pipeline)
--heidi-te <TE>	HEIDI: the echo time the field stands for the phase of, in the conversion to HEIDI’s units: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: last)

qsmxt invert ndi

Nonlinear Dipole Inversion

qsmxt invert ndi [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... --magnitude <MAGNITUDE>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--tau <TAU>	Gradient-descent step size (tau)
--alpha <ALPHA>	L2 regularization weight (alpha)
--max-iter <MAX_ITER>	Max iterations (the maximum, with –auto-stop)
--auto-stop <AUTO_STOP>	Stop automatically once noise starts to build up along the dipole cone (FANSI ndi_auto)
--te <TE>	Echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)
--weight-norm <NORM>	Magnitude-weight normaliser: global-max (the magnitude’s maximum, as FANSI; default) or mask-p999 (its 99.9th percentile inside the mask; more consistent in vivo)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI, the data weight (required): one 3D volume, a 4D one, or one per echo; several echoes are RSS-combined. Required.

qsmxt invert fansi

FANSI Nonlinear Total Variation

qsmxt invert fansi [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... --magnitude <MAGNITUDE>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI, the data weight (required): one 3D volume, a 4D one, or one per echo; several echoes are RSS-combined. Required.
--alpha1 <ALPHA1>	First-order (TV/TGV) L1 penalty weight (alpha1)
--mu1 <MU1>	Gradient-consistency ADMM weight (mu1), or auto for 100 × alpha1 (default: auto)
--mu2 <MU2>	Fidelity-consistency ADMM weight (mu2)
--alpha0 <ALPHA0>	Second-order L1 penalty weight (alpha0, nlTGV only), or auto for 2 × alpha1 (default: auto)
--mu0 <MU0>	Second-order consistency ADMM weight (mu0, nlTGV only), or auto for 2 × mu1 (default: auto)
--max-iter <MAX_ITER>	Max iterations.
--tol-update <TOL_UPDATE>	Percent-update convergence tolerance.
--tol-delta <TOL_DELTA>	Inner Newton convergence tolerance.
--te <TE>	Echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)

qsmxt invert fansi-tgv

FANSI Nonlinear Total Generalized Variation

qsmxt invert fansi-tgv [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... --magnitude <MAGNITUDE>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI, the data weight (required): one 3D volume, a 4D one, or one per echo; several echoes are RSS-combined. Required.
--alpha1 <ALPHA1>	First-order (TGV gradient) L1 penalty weight (alpha1)
--mu1 <MU1>	Gradient-consistency ADMM weight (mu1), or auto for 100 × alpha1 (default: auto)
--mu2 <MU2>	Fidelity-consistency ADMM weight (mu2)
--alpha0 <ALPHA0>	Second-order L1 penalty weight (alpha0), or auto for 2 × alpha1 (default: auto)
--mu0 <MU0>	Second-order consistency ADMM weight (mu0), or auto for 2 × mu1 (default: auto)
--max-iter <MAX_ITER>	Max iterations.
--tol-update <TOL_UPDATE>	Percent-update convergence tolerance.
--tol-delta <TOL_DELTA>	Inner Newton convergence tolerance.
--te <TE>	Echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: spacing)

qsmxt invert l1qsm

L1 Data-Fidelity QSM

qsmxt invert l1qsm [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... --magnitude <MAGNITUDE>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI, the data weight (required): one 3D volume, a 4D one, or one per echo; several echoes are RSS-combined. Required.
--alpha1 <ALPHA1>	Gradient (TV) L1 penalty weight (alpha1)
--mu1 <MU1>	Gradient-consistency ADMM weight (mu1), or auto for 100 × alpha1 (default: auto)
--mu2 <MU2>	Fidelity-consistency ADMM weight (mu2)
--mu3 <MU3>	L1 proximal ADMM weight (mu3)
--lambda <LAMBDA>	L1 fidelity strength (lambda)
--max-iter <MAX_ITER>	Max iterations.
--tol-update <TOL_UPDATE>	Percent-update convergence tolerance.
--tol-delta <TOL_DELTA>	Inner Newton convergence tolerance.
--te <TE>	Echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)

qsmxt invert whqsm

Weak-Harmonic QSM

qsmxt invert whqsm [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--alpha1 <ALPHA1>	TV regularization weight (alpha1)
--mu1 <MU1>	ADMM penalty for TV splitting (mu1)
--mu2 <MU2>	ADMM penalty for data-fidelity splitting (mu2)
--beta <BETA>	Weak-harmonic ROI penalty (beta)
--muh <MUH>	ADMM penalty for harmonic-field splitting (muh)
--max-iter <MAX_ITER>	Max iterations.
--tol-update <TOL_UPDATE>	Percent-update convergence tolerance.
--tol-delta <TOL_DELTA>	Inner Newton convergence tolerance.
--weight <WEIGHT>	Data weight: flat (1 over the whole volume, WH_nlTV.m’s default), flat-in-mask (1 inside the mask only), max-normalised (the magnitude over its maximum) or in-vivo (min(2 × magnitude / maximum, 1)); the last two need –magnitude (default: flat)
--magnitude <MAGNITUDE>	Magnitude NIfTI, for a magnitude data weight (–weight): one 3D volume, a 4D one, or one per echo; several echoes are RSS-combined.
--te <TE>	Echo time the field is converted to phase at: first, last, spacing (TE2 - TE1), sum (of the echo times), or a time in ms (default: first)

qsmxt invert hdqsm

Hybrid Data-Fidelity QSM

qsmxt invert hdqsm [OPTIONS] --mask <MASK> --output <OUTPUT> --echo-times <ECHO_TIMES>... --magnitude <MAGNITUDE>... <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--magnitude <MAGNITUDE>	Magnitude of every echo (the data weight is their echo-time-weighted combination): one 4D NIfTI, or one 3D NIfTI per echo in echo order. Required.
--alpha-l2 <ALPHA_L2>	L2-stage TV weight (alpha_l2)
--mu1-l2 <MU1_L2>	L2-stage gradient-consistency ADMM weight (mu1_l2), or auto for 100 × alpha_l2 (default: auto)
--mu2 <MU2>	Fidelity consistency weight (mu2)
--max-iter-l1 <MAX_ITER_L1>	Stage-1 (L1) iterations (max_iter_l1)
--max-iter-l2 <MAX_ITER_L2>	Stage-2 (L2) iterations (max_iter_l2)
--tol-update <TOL_UPDATE>	Stage-2 percent-update convergence tolerance.

qsmxt invert amp-pe

AMP-PE (Approximate Message Passing with Parameter Estimation)

qsmxt invert amp-pe [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--magnitude <MAGNITUDE>	Magnitude NIfTI (data weight and morphology image): one 3D volume, a 4D one, or one per echo; several echoes are RSS-combined.
--b0 <B0>	B0 field strength in Tesla (scales the simulated phase) Default: 3.
--wave-order <WAVE_ORDER>	Daubechies wavelet order (1=db1, 2=db2)
--nlevel <NLEVEL>	Wavelet decomposition levels.
--wave-pec <WAVE_PEC>	Morphology-mask energy retention fraction (0.0-1.0)
--simulated-te <SIMULATED_TE>	Simulated echo time (s) used to turn the field into phase.
--max-linearization-ite <MAX_LINEARIZATION_ITE>	Linearization iterations per stage.
--damp-rate-sig <DAMP_RATE_SIG>	GAMP signal-update damping rate.
--damp-rate-par <DAMP_RATE_PAR>	Parameter-estimation learning rate (kappa)
--max-pe-spar-ite <MAX_PE_SPAR_ITE>	Inner sparse-reconstruction iterations.
--max-pe-est-ite <MAX_PE_EST_ITE>	Inner parameter-estimation iterations.
--cvg-thd <CVG_THD>	GAMP inner convergence threshold.
--tikhonov-beta <TIKHONOV_BETA>	L2-seed Tikhonov weight.

qsmxt invert xqsm

xQSM deep-learning dipole inversion (weights downloaded on first use)

qsmxt invert xqsm [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla (scan metadata) Default: 3.

qsmxt invert qsmnet

QSMnet deep-learning dipole inversion (weights downloaded on first use)

qsmxt invert qsmnet [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla (scan metadata) Default: 3.

qsmxt invert qsmnet-plus

QSMnet+ deep-learning dipole inversion (weights downloaded on first use)

qsmxt invert qsmnet-plus [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla (scan metadata) Default: 3.

qsmxt invert qsmnetme

QSMnetME deep-learning dipole inversion, the χ-separation toolbox’s QSM network (weights downloaded on first use)

qsmxt invert qsmnetme [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla (scan metadata) Default: 3.

qsmxt invert autoqsm

AutoQSM single-step reconstruction — input is the TOTAL field (weights downloaded on first use)

qsmxt invert autoqsm [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla (scan metadata) Default: 3.

qsmxt invert qsmgan

QSMGAN deep-learning dipole inversion (weights downloaded on first use)

qsmxt invert qsmgan [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla (scan metadata) Default: 3.

qsmxt invert ir2qsm

IR2QSM deep-learning dipole inversion (weights downloaded on first use)

qsmxt invert ir2qsm [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla (scan metadata) Default: 3.

qsmxt invert lpcnn

LPCNN deep-learning dipole inversion (weights downloaded on first use)

qsmxt invert lpcnn [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla (scan metadata) Default: 3.

qsmxt invert modl-qsm

MoDL-QSM deep-learning dipole inversion → χ33/STI component (weights downloaded on first use)

qsmxt invert modl-qsm [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla (scan metadata) Default: 3.

qsmxt invert nextqsm

NeXtQSM single-step reconstruction — input is the TOTAL field (weights downloaded on first use)

qsmxt invert nextqsm [OPTIONS] --mask <MASK> --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input local field NIfTI file. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla (scan metadata) Default: 3.

qsmxt invert iqsm

iQSM single-step reconstruction — input is every echo’s WRAPPED phase (weights downloaded on first use)

qsmxt invert iqsm [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... <INPUT>

Argument	Description
<INPUT>	Wrapped phase NIfTI file in radians: 4D (x,y,z,echo), or 3D for one echo. Required.
-m, --mask <MASK>	Binary brain mask NIfTI file. Required.
-o, --output <OUTPUT>	Output NIfTI file in ppm (susceptibility for iqsm/iqsm-plus, the local field for iqfm) Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI file with the same echoes as the phase (4D), to weight the echo combination by magnitude; without it the echoes are weighted by TE² alone.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)
--echo-weighting <ECHO_WEIGHTING>	Echo weighting of the multi-echo combination: mag2-te2 weights each echo by (magnitude × TE)², the inverse of its noise variance, as in the authors’ Python code; mag-te2 by magnitude × TE², as in the authors’ MATLAB code. Values: mag-te2, mag2-te2. Default: mag2-te2.

qsmxt invert iqsm-plus

iQSM+ orientation-adaptive single-step reconstruction — input is every echo’s WRAPPED phase; echoes are combined with (magnitude × TE)² weights (weights downloaded on first use)

qsmxt invert iqsm-plus [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-times <ECHO_TIMES>... <INPUT>

Argument	Description
<INPUT>	Wrapped phase NIfTI file in radians: 4D (x,y,z,echo), or 3D for one echo. Required.
-m, --mask <MASK>	Binary brain mask NIfTI file. Required.
-o, --output <OUTPUT>	Output NIfTI file in ppm (susceptibility for iqsm/iqsm-plus, the local field for iqfm) Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI file with the same echoes as the phase (4D), to weight the echo combination by magnitude; without it the echoes are weighted by TE² alone.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds, every echo’s, in order. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--tile-size <TILE_SIZE>	DL tiling: output core size per patch, in voxels (enables tiling)
--tile-halo <TILE_HALO>	DL tiling: context margin per side, in voxels (default 8)

qsmxt invert cosmos

COSMOS — multi-orientation least squares over 2+ co-registered orientations

qsmxt invert cosmos [OPTIONS] --input <INPUTS> --mask <MASK> --output <OUTPUT>

Argument	Description
-i, --input <INPUTS>	Local field NIfTI in ppm — repeat once per orientation, all on one common grid. Required.
--b0-direction <B0_DIRECTION>	B0 direction for one orientation in the common grid’s voxel frame (3 values) — repeat once per –input, in the same order. Omitted, each volume’s own affine is used instead, which is only valid when the orientations were separately prescribed.
--b0-dirs <B0_DIRS>	Text file of B0 directions, one “x y z” per line, in –input order.
-m, --mask <MASK>	Binary mask NIfTI file, on the common grid. Required.
--force	Reconstruct even when the B0 directions look degenerate. A degenerate set yields a plausible-looking map that is not the method you asked for; prefer fixing the metadata.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--lambda <LAMBDA>	L2 regularization on the k-space denominator. 0 (default) is plain least squares and is right for 3+ covering orientations; 1e-3..1e-2 helps with 2, or single-axis rotations.
--magnitude <MAGNITUDES>	Magnitude NIfTI for one orientation — repeat once per –input to use the magnitude-weighted formulation (conjugate gradient) instead of the closed form.
--tol <TOL>	Conjugate-gradient tolerance (weighted formulation only)
--max-iter <MAX_ITER>	Maximum conjugate-gradient iterations (weighted formulation only)

qsmxt invert sti

STI (susceptibility tensor imaging) — 6+ co-registered orientations to a rank-2 tensor

qsmxt invert sti [OPTIONS] --input <INPUTS> --mask <MASK> --output <OUTPUT>

Argument	Description
-i, --input <INPUTS>	Local field NIfTI in ppm — repeat once per orientation, all on one common grid. Required.
--b0-direction <B0_DIRECTION>	B0 direction for one orientation in the common grid’s voxel frame (3 values) — repeat once per –input, in the same order. Omitted, each volume’s own affine is used instead, which is only valid when the orientations were separately prescribed.
--b0-dirs <B0_DIRS>	Text file of B0 directions, one “x y z” per line, in –input order.
-m, --mask <MASK>	Binary mask NIfTI file, on the common grid. Required.
--force	Reconstruct even when the B0 directions look degenerate. A degenerate set yields a plausible-looking map that is not the method you asked for; prefer fixing the metadata.
-o, --output <OUTPUT>	Output prefix; writes {prefix}_desc-sti_Chimap.nii (the 4D tensor, components 11 12 13 22 23 33), {prefix}_desc-mms_Chimap.nii, {prefix}_desc-msa_Chimap.nii and {prefix}_desc-sti_evec.nii (the 4D principal eigenvector) — the pipeline’s names. Required.
--lambda <LAMBDA>	Ridge term on the 6x6 normal equations at each k-point. Raise when orientations are few, clustered or near-coplanar.

qsmxt separate

Susceptibility source separation

Subcommand	Description
r2star-qsm	R2*-QSM closed-form (Dimov 2022) — QSM + R2* (GRE-only)
decompose	DECOMPOSE-QSM signal-domain fit (Chen 2021) — a QSM per echo + multi-echo magnitude (GRE-only)
chi-sep-ilsqr	χ-separation iLSQR (Shin 2021) — local field + R2’ + QSM
chi-sep-medi	χ-separation MEDI — local field + R2’ + multi-echo magnitude + QSM
wavesep	WaveSep wavelet-L1 (Fang 2023) — QSM + R2’
hc-chisep	Hollow-cylinder χ-separation (Stewart 2026) — QSM + R2’ + multi-echo magnitude
susep-net	SUSEP-Net deep-learning separation — QSM + R2’ + local field (weights downloaded on first use)
chi-sepnet	χ-sepnet deep-learning separation — local field + R2’ or R2* (weights downloaded on first use)

qsmxt separate r2star-qsm

R2*-QSM closed-form (Dimov 2022) — QSM + R2* (GRE-only)

qsmxt separate r2star-qsm [OPTIONS] --qsm <QSM> --mask <MASK> --output <OUTPUT> <--r2star <R2STAR>|--magnitude <MAGNITUDE>...>

Argument	Description
--qsm <QSM>	Conventional QSM (χ_total) NIfTI in ppm. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output prefix; writes {prefix}_desc-paramagnetic_Chimap.nii, {prefix}_desc-diamagnetic_Chimap.nii and {prefix}_desc-total_Chimap.nii — the pipeline’s names. Required.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Default: 3.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--r2star <R2STAR>	R2* map NIfTI in Hz (else fit from –magnitude + –echo-times)
--magnitude <MAGNITUDE>	Multi-echo magnitude: one 4D file or several 3D files (space-separated). Fits R2* if –r2star absent.
--echo-times <ECHO_TIMES>	GRE echo times in seconds (space- or comma-separated), required with –magnitude.
--r-const-3t <R_CONST_3T>	Relaxometric constant at 3T (Hz/ppm)

qsmxt separate decompose

DECOMPOSE-QSM signal-domain fit (Chen 2021) — a QSM per echo + multi-echo magnitude (GRE-only)

qsmxt separate decompose [OPTIONS] --qsm <QSM> --mask <MASK> --output <OUTPUT> --magnitude <MAGNITUDE>... --qsm-per-echo <QSM_PER_ECHO>... --echo-times <ECHO_TIMES>...

Argument	Description
--qsm <QSM>	Conventional QSM (χ_total) NIfTI in ppm. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output prefix; writes {prefix}_desc-paramagnetic_Chimap.nii, {prefix}_desc-diamagnetic_Chimap.nii and {prefix}_desc-total_Chimap.nii — the pipeline’s names. Required.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Default: 3.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--magnitude <MAGNITUDE>	Multi-echo magnitude: one 4D file or several 3D files (space-separated) Required.
--qsm-per-echo <QSM_PER_ECHO>	QSM of each echo (ppm), which sets each echo’s phase: one 4D file or one 3D file per echo (space-separated; the same file may be given for every echo) Required.
--echo-times <ECHO_TIMES>	GRE echo times in seconds (space- or comma-separated) Required.
--n-inner <N_INNER>	Inner alternating passes per voxel.
--chi-bound <BOUND>	Bound on |χ| in the fit: a number in ppm (default: 0.5), or echo-spacing to compute it from the field strength and echo spacing.
--unwrap-phase <BOOL>	Unwrap the phase along the echoes (default: false, as in the authors’ code)
--fit <FIT>	The fit: reference-code (Tim Ho’s script; default) or paper (Chen et al. 2021: amplitudes summing to one, per-voxel S(0) = 1 scale) Values: reference-code, paper.
--max-iterations <MAX_ITERATIONS>	Maximum solver iterations of each fit (lsqcurvefit MaxIterations; default: 400)
--max-function-evaluations <MAX_FUNCTION_EVALUATIONS>	Maximum function evaluations of each fit (lsqcurvefit MaxFunctionEvaluations; default: 100 × the fit’s parameters)
--function-tolerance <FUNCTION_TOLERANCE>	Relative cost change each fit stops at (lsqcurvefit FunctionTolerance; default: 1e-6)
--step-tolerance <STEP_TOLERANCE>	Relative step each fit stops at (lsqcurvefit StepTolerance; default: 1e-6)
--optimality-tolerance <OPTIMALITY_TOLERANCE>	First-order optimality each fit stops at (lsqcurvefit OptimalityTolerance; default: 1e-6)
--finite-difference-step <FINITE_DIFFERENCE_STEP>	Relative finite-difference step of the Jacobian (lsqcurvefit FiniteDifferenceStepSize; default: √eps)
--typical-x <TYPICAL_X>	Typical magnitude of each fitted value (lsqcurvefit TypicalX; default: 1)

qsmxt separate chi-sep-ilsqr

χ-separation iLSQR (Shin 2021) — local field + R2’ + QSM

qsmxt separate chi-sep-ilsqr [OPTIONS] --qsm <QSM> --mask <MASK> --output <OUTPUT> --local-field <LOCAL_FIELD> --r2prime <R2PRIME>

Argument	Description
--qsm <QSM>	Conventional QSM (χ_total) NIfTI in ppm. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output prefix; writes {prefix}_desc-paramagnetic_Chimap.nii, {prefix}_desc-diamagnetic_Chimap.nii and {prefix}_desc-total_Chimap.nii — the pipeline’s names. Required.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Default: 3.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--local-field <LOCAL_FIELD>	Local (tissue) field NIfTI in ppm. Required.
--r2prime <R2PRIME>	R2’ map NIfTI in Hz. Required.
--dr-pos <DR_POS>	Paramagnetic relaxometric constant (Hz/ppm) (default: 137 × B0 / 2.9 T)
--dr-neg <DR_NEG>	Diamagnetic relaxometric constant (Hz/ppm) (default: 137 × B0 / 2.9 T)
--pad <X> <Y> <Z>	Zero-padding at both ends of each axis in voxels, X Y Z (default: 16 16 16)
--cg-max-iter-lsqr <CG_MAX_ITER_LSQR>	Maximum CG iterations of the LSQR step (default: 50)
--cg-tol-lsqr <CG_TOL_LSQR>	CG relative tolerance of the LSQR step (default: 0.02)
--cg-max-iter-sa <CG_MAX_ITER_SA>	Maximum CG iterations of the streaking-removal step (default: 50)
--cg-tol-sa <CG_TOL_SA>	CG relative tolerance of the streaking-removal step (default: 0.02)
--d2-thr <D2_THR>	The streaking cone is where |D| is below this (default: 0.11)
--prc-w-i-min <PRC_W_I_MIN>	Percentile of the field’s Laplacian at which the LSQR weight is 1 (default: 60)
--prc-w-i-max <PRC_W_I_MAX>	Percentile of the field’s Laplacian at which the LSQR weight is 0 (default: 99.8)
--prc-w-g-min <PRC_W_G_MIN>	Percentile of each gradient’s magnitude at which the streaking-removal weight is 1 (default: 50)
--prc-w-g-max <PRC_W_G_MAX>	Percentile of each gradient’s magnitude at which the streaking-removal weight is 0 (default: 70)

qsmxt separate chi-sep-medi

χ-separation MEDI — local field + R2’ + multi-echo magnitude + QSM

qsmxt separate chi-sep-medi [OPTIONS] --qsm <QSM> --mask <MASK> --output <OUTPUT> --local-field <LOCAL_FIELD> --r2prime <R2PRIME> --magnitude <MAGNITUDE>... --echo-times <ECHO_TIMES>...

Argument	Description
--qsm <QSM>	Conventional QSM (χ_total) NIfTI in ppm. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output prefix; writes {prefix}_desc-paramagnetic_Chimap.nii, {prefix}_desc-diamagnetic_Chimap.nii and {prefix}_desc-total_Chimap.nii — the pipeline’s names. Required.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Default: 3.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--local-field <LOCAL_FIELD>	Local (tissue) field NIfTI in ppm. Required.
--r2prime <R2PRIME>	R2’ map NIfTI in Hz. Required.
--magnitude <MAGNITUDE>	Multi-echo magnitude: one 4D file or one 3D file per echo (space-separated) Required.
--echo-times <ECHO_TIMES>	GRE echo times in seconds (space- or comma-separated) Required.
--csf-mask <FILE|none>	CSF mask NIfTI for the CSF-uniformity term, or none for no CSF term (the toolbox’s GUI without a CSF mask) (default: the CSF found from R2* of the echoes, as the toolbox’s script does; see –csf-r2star)
--csf-r2star <METHOD>	The R2* the CSF mask is found from when no –csf-mask is given: arlo (MEDI’s ARLO; the default) or nnls (the χ-separation toolbox’s r2star_nnls, which its script runs)
--lambda <LAMBDA>	Weight of the data terms against the L1 terms (default: 1)
--lambda-csf <LAMBDA_CSF>	Weight of the CSF-uniformity term (default: 1)
--dr-pos <DR_POS>	Paramagnetic relaxometric constant (Hz/ppm) (default: 137 × B0 / 2.9 T)
--dr-neg <DR_NEG>	Diamagnetic relaxometric constant (Hz/ppm) (default: 137 × B0 / 2.9 T)
--pad <PAD>	Zero-padding at both ends of the third axis, in voxels (default: 5)
--tol <TOL>	Stop when the relative change falls below this (default: 0.03)
--max-iter <MAX_ITER>	Maximum Gauss-Newton iterations (default: 30)
--cg-tol <CG_TOL>	Inner CG tolerance (default: 0.05)
--cg-max-iter <CG_MAX_ITER>	Inner CG max iterations (default: 100)

qsmxt separate wavesep

WaveSep wavelet-L1 (Fang 2023) — QSM + R2’

qsmxt separate wavesep [OPTIONS] --qsm <QSM> --mask <MASK> --output <OUTPUT> --r2prime <R2PRIME>

Argument	Description
--qsm <QSM>	Conventional QSM (χ_total) NIfTI in ppm. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output prefix; writes {prefix}_desc-paramagnetic_Chimap.nii, {prefix}_desc-diamagnetic_Chimap.nii and {prefix}_desc-total_Chimap.nii — the pipeline’s names. Required.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Default: 3.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--r2prime <R2PRIME>	R2’ map NIfTI in Hz. Required.
--dr-pos <DR_POS>	Paramagnetic relaxivity (Hz/ppm)
--dr-neg <DR_NEG>	Diamagnetic relaxivity (Hz/ppm)
--alpha <ALPHA>	Proximal-gradient step size.
--lambda <LAMBDA>	Wavelet L1 soft-threshold weight.
--wavelet-order <WAVELET_ORDER>	Daubechies wavelet order.
--max-iter <MAX_ITER>	ISTA max iterations.
--tol <TOL>	Relative-change stop tolerance.

qsmxt separate hc-chisep

Hollow-cylinder χ-separation (Stewart 2026) — QSM + R2’ + multi-echo magnitude

qsmxt separate hc-chisep [OPTIONS] --qsm <QSM> --mask <MASK> --output <OUTPUT> --r2prime <R2PRIME> --magnitude <MAGNITUDE>... --echo-times <ECHO_TIMES>...

Argument	Description
--qsm <QSM>	Conventional QSM (χ_total) NIfTI in ppm. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output prefix; writes {prefix}_desc-paramagnetic_Chimap.nii, {prefix}_desc-diamagnetic_Chimap.nii and {prefix}_desc-total_Chimap.nii — the pipeline’s names. Required.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Default: 3.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--r2prime <R2PRIME>	R2’ map NIfTI in Hz. Required.
--magnitude <MAGNITUDE>	Multi-echo magnitude: one 4D file or several 3D files (space-separated) Required.
--echo-times <ECHO_TIMES>	GRE echo times in seconds (space- or comma-separated) Required.
--se-magnitude <SE_MAGNITUDE>	Multi-echo spin-echo magnitude: one 4D file or several 3D files (optional)
--se-echo-times <SE_ECHO_TIMES>	Spin-echo times in seconds (comma-separated) for –se-magnitude.
--dr-pos-3t <DR_POS_3T>	Paramagnetic relaxivity at 3T (Hz/ppm)
--bin-hz <BIN_HZ>	R2’ bin width for the anchored grid search (Hz)

qsmxt separate susep-net

SUSEP-Net deep-learning separation — QSM + R2’ + local field (weights downloaded on first use)

qsmxt separate susep-net [OPTIONS] --qsm <QSM> --mask <MASK> --output <OUTPUT> --local-field <LOCAL_FIELD> --r2prime <R2PRIME>

Argument	Description
--qsm <QSM>	Conventional QSM (χ_total) NIfTI in ppm. Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output prefix; writes {prefix}_desc-paramagnetic_Chimap.nii, {prefix}_desc-diamagnetic_Chimap.nii and {prefix}_desc-total_Chimap.nii — the pipeline’s names. Required.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Default: 3.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--local-field <LOCAL_FIELD>	Local (tissue) field NIfTI in ppm. Required.
--r2prime <R2PRIME>	R2’ map NIfTI in Hz. Required.
--patch <D> <H> <W>	Sliding-window patch in voxels, D H W, each a multiple of 8 (default: the whole volume at once)

qsmxt separate chi-sepnet

χ-sepnet deep-learning separation — local field + R2’ or R2* (weights downloaded on first use)

qsmxt separate chi-sepnet [OPTIONS] --mask <MASK> --output <OUTPUT> --local-field <LOCAL_FIELD> <--r2prime <R2PRIME>|--r2star <R2STAR>>

Argument	Description
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output prefix; writes {prefix}_desc-paramagnetic_Chimap.nii, {prefix}_desc-diamagnetic_Chimap.nii and {prefix}_desc-total_Chimap.nii — the pipeline’s names. Required.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Default: 3.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--local-field <LOCAL_FIELD>	Local (tissue) field NIfTI in ppm. Required.
--r2prime <R2PRIME>	R2’ map NIfTI in Hz (χ-sepnet-R2’)
--r2star <R2STAR>	R2* map NIfTI in Hz (χ-sepnet-R2*: R2’ predicted by R2PRIMEnet; --route toolbox only)
--qsm <QSM>	Conventional QSM (χ_total) NIfTI in ppm; --route native only (the toolbox route makes its own with QSMnetME)
--route <ROUTE>	Input route: toolbox (default; the authors’: training frame, QSMnetME susceptibility channel) or native (the network alone on the acquired grid, given –qsm) Values: toolbox, native. Default: toolbox.
--patch <D> <H> <W>	--route native: sliding-window patch in voxels, D H W, each a multiple of 16 (default: 192 192 128, the authors’)
--frame-patch <D> <H> <W>	--route toolbox: run the networks as a sliding window of this patch over the 1 mm frame, D H W, each a multiple of 16 (default: the whole frame at once, as the authors do)
--ignore-field-mismatch	With –r2star: run R2PRIMEnet even when the field strength is not 3 T (uncalibrated)

qsmxt qsmart

QSMART vessel-aware reconstruction: total field -> susceptibility

qsmxt qsmart [OPTIONS] --mask <MASK> --output <OUTPUT> --field-strength <FIELD_STRENGTH> --echo-time <ECHO_TIME> <INPUT>

Argument	Description
<INPUT>	Input total field NIfTI file (ppm) Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output susceptibility map NIfTI file. Required.
--b0-direction <B0_DIRECTION>	B0 direction (3 values) Default: 0 0 1.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla. Required.
--echo-time <ECHO_TIME>	Echo time in seconds. Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI (optional; drives vasculature detection). With –qsmart-vessel-magnitude mean-n4 (the default) every echo’s magnitude (4D), averaged and N4 bias-corrected; with combined, a combined magnitude (3D), used as it is.
--reliability <RELIABILITY>	QSMART’s reliability map from its echo fit (qsmxt fieldmap laplacian --b0-estimation qsmart-echofit --reliability-output): both stages are reconstructed inside it.
--intermediates <DIR>	Also write QSMART’s intermediate maps into this directory, as <name>.nii: vessel_mask (1 = vessel), localfield_stage1, chimap_stage1, localfield_stage2 and chimap_stage2 (ppm), and vessel_magnitude when QSMART makes its own (mean-n4)
--ilsqr-precision <PRECISION>	Arithmetic of the inner iLSQR’s two LSQR solves (--qsmart-inversion ilsqr): double, or single as STI Suite’s QSM_iLSQR, which QSMART calls (default: double; qsmxt run --ilsqr-precision)
--qsmart-ilsqr-max-iter <QSMART_ILSQR_MAX_ITER>	QSMART iLSQR max iterations per stage. The iLSQR tolerance is –ilsqr-tol.
--qsmart-ilsqr-pad-mm <MM>	Zero-padding of QSMART’s inner iLSQR solves, in mm per side (default: 6)
--qsmart-vasc-sphere-radius <QSMART_VASC_SPHERE_RADIUS>	QSMART vasculature detection sphere radius, in –qsmart-vessel-units (default 8)
--qsmart-inversion <QSMART_INVERSION>	QSMART inner dipole inversion algorithm (default: ilsqr) Values: rts, tv, tkd, tsvd, tgv, tikhonov, nltv, medi, tfi, ilsqr, star, lsqr, heidi, qsmart, ndi, fansi, fansi-tgv, l1qsm, whqsm, hdqsm, amp-pe, xqsm, qsmnet, qsmnet-plus, qsmnetme, autoqsm, qsmgan, ir2qsm, lpcnn, modl-qsm, nextqsm, iqsm, iqsm-plus.
--qsmart-sdf-sigma1-stage1 <QSMART_SDF_SIGMA1_STAGE1>	QSMART SDF sigma1, stage 1 (voxels)
--qsmart-sdf-sigma2-stage1 <QSMART_SDF_SIGMA2_STAGE1>	QSMART SDF sigma2, stage 1 (voxels)
--qsmart-sdf-sigma1-stage2 <QSMART_SDF_SIGMA1_STAGE2>	QSMART SDF sigma1, stage 2 (voxels)
--qsmart-sdf-sigma2-stage2 <QSMART_SDF_SIGMA2_STAGE2>	QSMART SDF sigma2, stage 2 (voxels)
--qsmart-sdf-lower-lim <QSMART_SDF_LOWER_LIM>	QSMART SDF proximity lower limit.
--qsmart-sdf-curv-constant <QSMART_SDF_CURV_CONSTANT>	QSMART SDF curvature constant.
--qsmart-frangi-scale-min <QSMART_FRANGI_SCALE_MIN>	QSMART Frangi min vessel radius, in –qsmart-vessel-units (default 0.5)
--qsmart-frangi-scale-max <QSMART_FRANGI_SCALE_MAX>	QSMART Frangi max vessel radius, in –qsmart-vessel-units (default 6)
--qsmart-frangi-scale-ratio <QSMART_FRANGI_SCALE_RATIO>	QSMART Frangi scale step, in –qsmart-vessel-units (default 0.5)
--qsmart-frangi-c <QSMART_FRANGI_C>	QSMART Frangi C noise threshold.
--qsmart-vessel-units <QSMART_VESSEL_UNITS>	Units of QSMART’s vessel sizes (–qsmart-vasc-sphere-radius and the Frangi scales): voxels (default, as QSMART gives them) or mm (converted to voxels with the mean voxel size) Values: voxels, mm.
--qsmart-vessel-magnitude <QSMART_VESSEL_MAGNITUDE>	The magnitude QSMART detects vessels on: mean-n4 (the mean over echoes, N4 bias-corrected, as QSMART prepares it; default) or combined (the root sum of squares over echoes, inhomogeneity-corrected unless –no-inhomogeneity-correction) Values: mean-n4, combined.
--qsmart-offset-grid <QSMART_OFFSET_GRID>	The dipole grid of QSMART’s offset adjustment: fft (default) or qsmart-reference (adjust_offset.m’s own, half a step off the FFT grid on odd matrix sizes, as the QSMART pipeline runs it) Values: fft, qsmart-reference.

qsmxt swi

Susceptibility-weighted imaging

qsmxt swi [OPTIONS] --phase <PHASE>... --magnitude <MAGNITUDE>... --output <OUTPUT>

Argument	Description
--phase <PHASE>	Wrapped phase NIfTI files, one per echo, in echo order. Required.
--magnitude <MAGNITUDE>	Magnitude NIfTI files, one per echo, in the same order as –phase. Required.
--tes <TES>	Echo times in seconds, one per echo (read from the phase JSON sidecars when omitted; not needed for a single echo)
-m, --mask <MASK>	Mask the phase is filtered within (default: CLEAR-SWI’s own, from the first echo’s magnitude)
-o, --output <OUTPUT>	Output SWI NIfTI file. Required.
--mip	Also compute minimum intensity projection.
--mip-output <MIP_OUTPUT>	Output path for MIP (default: the –output name with its swi suffix replaced by minIP, or _minIP appended — sub-01_swi.nii → sub-01_minIP.nii, as the pipeline names them)
--filtered-phase	Also write CLEAR-SWI’s high-pass filtered phase, the phase the SWI is weighted by: Hz for several echoes, radians for one, 0 outside the mask (qsmxt run --swi-filtered-phase)
--filtered-phase-output <FILTERED_PHASE_OUTPUT>	Output path for the filtered phase (default: the –output name with its swi suffix replaced by desc-swi_phase, or _desc-swi_phase appended — sub-01_swi.nii → sub-01_desc-swi_phase.nii, as the pipeline names it). Implies –filtered-phase.
--swi-unwrapping <SWI_UNWRAPPING>	SWI phase unwrapping before the echoes are combined [default: laplacian]. Values: laplacian, laplacian-slice, romeo.
--swi-hp-sigma <SWI_HP_SIGMA>	SWI high-pass filter sigma (3 values, in voxels)
--swi-scaling <SWI_SCALING>	SWI phase-mask scaling. Values: tanh, negative-tanh, positive, negative, triangular.
--swi-strength <SWI_STRENGTH>	SWI phase scaling strength.
--swi-mip-window <SWI_MIP_WINDOW>	SWI MIP window size in slices.

qsmxt segment

Whole-brain segmentation of a magnitude image with SynthSeg

qsmxt segment [OPTIONS] --output <OUTPUT> <MAGNITUDE>

Argument	Description
<MAGNITUDE>	Input magnitude NIfTI file. Not a susceptibility map: SynthSeg reads anatomy, and χ has none of the tissue contrast it was trained to recognise. Required.
-o, --output <OUTPUT>	Output label NIfTI file (FreeSurfer label ids) Required.
--lookup <LOOKUP>	Also write the BIDS label lookup table (index/name) to this path.
--synthseg-version <VERSION>	SynthSeg weights generation: v1 (32 labels, ships with SynthSeg) or v2 (33 labels, adds CSF; weights not published yet, so refused unless synthseg-2.0.onnx is in $QSM_MODEL_DIR)
--synthseg-crop <SYNTHSEG_CROP>	Cubic centre-crop in voxels at 1 mm before inference (rounded up to a multiple of 32). Omit to process the whole volume, as the SynthSeg CLI does.
--synthseg-patch <VOXELS>	Run SynthSeg as a sliding window of half-overlapping cubes this many voxels on a side at 1 mm (at least 64, rounded up to a multiple of 32), so its memory follows the patch, not the volume. Omit to run the whole volume at once, as the SynthSeg CLI does.
--no-synthseg-flip-averaging	Skip the left-right flipped second pass. Halves inference time, as SynthSeg’s –fast does.
--no-synthseg-topology-cleanup	Skip resetting each topological class to its largest connected component.
--synthseg-sigma <SYNTHSEG_SIGMA>	Gaussian blur on the posteriors, in voxels (0 disables it)
--use-custom-dseg <TOOL>	Prefer a bring-your-own segmentation from <bids>/derivatives/<TOOL>/ when present. Bare flag = first matching tool alphabetically. Always falls back to running SynthSeg.

qsmxt r2star

R2* mapping from multi-echo magnitude data

qsmxt r2star --mask <MASK> --output <OUTPUT> --echo-times <ECHO_TIMES> <ECHO_TIMES> <ECHO_TIMES>... <INPUTS>...

Argument	Description
<INPUTS>	Multi-echo magnitude: one 4D file or several 3D files (3+ echoes; space-separated) Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output R2* map NIfTI file. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds (must match number of inputs) Required.

qsmxt t2star

T2* mapping from multi-echo magnitude data

qsmxt t2star --mask <MASK> --output <OUTPUT> --echo-times <ECHO_TIMES> <ECHO_TIMES> <ECHO_TIMES>... <INPUTS>...

Argument	Description
<INPUTS>	Multi-echo magnitude: one 4D file or several 3D files (3+ echoes; space-separated) Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output T2* map NIfTI file. Required.
--echo-times <ECHO_TIMES>	Echo times in seconds (must match number of inputs) Required.

qsmxt r2

R2 mapping from multi-echo spin-echo (MESE) magnitude data (EPG)

qsmxt r2 [OPTIONS] --mask <MASK> --output <OUTPUT> --echo-times <ECHO_TIMES>... <INPUTS>...

Argument	Description
<INPUTS>	Multi-echo spin-echo (MESE) magnitude: one 4D file or several 3D files (space-separated) Required.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
-o, --output <OUTPUT>	Output R2 map NIfTI file. Required.
--echo-times <ECHO_TIMES>	Spin-echo times in seconds (space- or comma-separated) Required.
--t1 <T1>	Assumed T1 in seconds (EPG dictionary)
--b1-map <B1_MAP>	Optional B1 map NIfTI (refocusing efficiency); fitted from the dictionary if absent.

qsmxt r2prime

R2’ mapping: measured as R2* - R2, or estimated from R2* by R2PRIMEnet

qsmxt r2prime [OPTIONS] --r2star <R2STAR> --mask <MASK> --output <OUTPUT>

Argument	Description
--r2star <R2STAR>	R2* map NIfTI in Hz. Required.
--r2 <R2>	R2 map NIfTI in Hz: R2’ is measured, as R2* - R2.
--r2primenet	Estimate R2’ from the R2* map with R2PRIMEnet (deep learning, trained at 3 T; weights downloaded on first use) instead of measuring it from an R2 map. Needs –field-strength.
-m, --mask <MASK>	Binary mask NIfTI file. Required.
--r2-coverage <R2_COVERAGE>	Binary mask of where the R2 acquisition reached, if it does not cover the whole mask (a MESE slab). Without it, coverage is inferred from R2 > 0.
-o, --output <OUTPUT>	Output R2’ map NIfTI file. Required.
--output-coverage <OUTPUT_COVERAGE>	Also write the mask of voxels where R2’ is a measurement.
--field-strength <FIELD_STRENGTH>	B0 field strength in Tesla (–r2primenet: R2PRIMEnet refuses a field strength other than 3 T unless –r2primenet-ignore-field-mismatch is given)
--r2primenet-ignore-field-mismatch	Run R2PRIMEnet even when the acquisition is not at 3 T, the only field strength its weights were trained at. The R2’ is then uncalibrated.
--r2primenet-patch <D> <H> <W>	R2PRIMEnet sliding-window patch in voxels, D H W, each a multiple of 16 (default: 192 192 128)

qsmxt homogeneity

Inhomogeneity correction on magnitude data (makehomogeneous or N4)

qsmxt homogeneity [OPTIONS] --output <OUTPUT> <INPUT>

Argument	Description
<INPUT>	Input magnitude NIfTI file. Required.
-o, --output <OUTPUT>	Output corrected magnitude NIfTI file. Required.
--method <METHOD>	Correction method: makehomogeneous (–sigma, –nbox) or n4 (N4 bias field correction, as ANTs’ N4BiasFieldCorrection; –n4-*) Values: makehomogeneous, n4. Default: makehomogeneous.
--sigma <SIGMA>	makehomogeneous: smoothing sigma in mm. Default: 7.
--nbox <NBOX>	makehomogeneous: boxes per dimension for the tissue segmentation the receive sensitivity is estimated from. Default: 15.
--n4-shrink-factor <N4_SHRINK_FACTOR>	N4: subsample the image by this factor before fitting the bias field (default 4)
--n4-max-iterations <ITERATIONS>	N4: maximum iterations at each fitting level, one number a level (default 50 50 50 50); the B-spline mesh doubles from one level to the next.
--n4-convergence-threshold <N4_CONVERGENCE_THRESHOLD>	N4: stop a level once the change in the bias field (its coefficient of variation) falls to this (default 0: every iteration runs)
--n4-mesh <X> <Y> <Z>	N4: B-spline mesh elements per axis at the first level, X Y Z (default 1 1 1)
--n4-spline-order <N4_SPLINE_ORDER>	N4: order of the B-spline bias field, 1 to 3 (default 3)
--n4-bias-field-fwhm <N4_BIAS_FIELD_FWHM>	N4: FWHM of the Gaussian modelling the bias field’s blurring of the log-intensity histogram (default 0.15)
--n4-wiener-noise <N4_WIENER_NOISE>	N4: noise term of the Wiener deconvolution of the histogram (default 0.01)
--n4-histogram-bins <N4_HISTOGRAM_BINS>	N4: bins of the log-intensity histogram (default 200)
--n4-geometry <GEOMETRY>	N4: the image geometry it is given: scanner (the scan’s header, as ANTs reads the acquired NIfTI; the default) or make-nii (the axis-aligned header MATLAB’s make_nii writes, as QSMART runs N4). N4’s shrunk grid falls where the header puts it.

qsmxt resample

Resample oblique volume to axial orientation

qsmxt resample [OPTIONS] [INPUT]

Argument	Description
<INPUT>	Input NIfTI file. Treated as continuous data (magnitude, an unwrapped field map, chi). For wrapped phase pass –phase instead, with its magnitude, so the interpolation can go through the complex domain and survive the wraps.
-o, --output <OUTPUT>	Output resampled NIfTI file (for the positional input)
--phase <PHASE>	Wrapped phase to resample together with –magnitude (complex-domain interpolation)
--magnitude <MAGNITUDE>	Magnitude paired with –phase.
--phase-out <PHASE_OUT>	Where to write the resampled phase (default: <phase stem>_axial.nii)
--magnitude-out <MAGNITUDE_OUT>	Where to write the resampled magnitude (default: <magnitude stem>_axial.nii)
--noise-fill	Fill the empty corners of the new grid with seeded noise, as QSMxT 8.x did. Only the –phase / –magnitude pair can be noise-filled; by default the corners are zero.

qsmxt quality-map

Compute ROMEO phase quality map

qsmxt quality-map [OPTIONS] --output <OUTPUT> <PHASE>

Argument	Description
<PHASE>	Input phase NIfTI file (first echo) Required.
-o, --output <OUTPUT>	Output quality map NIfTI file. Required.
--magnitude <MAGNITUDE>	Magnitude image (improves quality estimation)
--phase2 <PHASE2>	Second echo phase image (improves quality estimation)
--te1 <TE1>	Echo time of first phase in seconds. Default: 0.02.
--te2 <TE2>	Echo time of second phase in seconds (if –phase2 provided) Default: 0.04.

qsmxt example

Download an example BIDS dataset to try the pipeline on

qsmxt example [OPTIONS] [OUTPUT_DIR]

Argument	Description
<OUTPUT_DIR>	Directory to create or add the BIDS dataset in [default: qsmxt-example].
-n, --name <NAME>	Which acquisition to fetch; repeat to add several to the one dataset [default: prisma-bridge-run1].
--list	List the available acquisitions and exit.
--force	Rewrite acquisitions that are already present in the dataset.

qsmxt tui

Launch interactive TUI for pipeline configuration

qsmxt tui

qsmxt update

Check for updates and optionally install the latest version

qsmxt update [OPTIONS]

Argument	Description
-y, --yes	Update without prompting for confirmation.
--force	Replace a development build (version 0.0.0) with the latest release. Without it, qsmxt update refuses on a dev build, since every release would look newer.
--main	Build the latest commit on the main branch from source and install it in place of this qsmxt. Needs the Rust toolchain (https://rustup.rs). The build is kept in ~/.qsmxt/build, so the next update rebuilds only what changed. qsmxt update --force goes back to the latest release.

qsmxt news

Show the announcements from the QSMxT team, newest first

qsmxt news