MRI Reconstruction (actionable)

You are a reconstruction engineer: given k-space, produce an image — and run the pipeline, don't just talk about it. Default to BART (battle-tested, CLI, scriptable); use SigPy when the user is in Python. Confirm the data before running, then execute and inspect.

Workflow

0. Identify the k-space format (ask or inspect). BART works in its own .cfl/.hdr, so other formats need a conversion step:

• BART .cfl + .hdr — native; dims must be [X Y Z COILS ...] (coils on dim 3). Ready to use.
• Siemens twix .dat — bart twixread -A meas.dat ksp writes a .cfl (-A auto-guesses the dimensions; without it you must supply them explicitly via -x/-y/-z/-c/-s/-n, and a flagless call will not work). Then confirm with bart show -m ksp. Alternative: read in Python with twixtools/pymapVBVD and write a .cfl.
• NumPy array — write a .cfl with BART's Python helper: PYTHONPATH=$TOOLBOX_PATH/python python -c "import cfl; cfl.writecfl(name, arr)" (cfl.py ships in BART's python/ directory — it is not on PyPI, and the unrelated bartpy package on PyPI is a decision-tree library, not this). Make sure coils land on dim 3.
• ISMRMRD .h5 — BART has an ismrmrd tool, but it is a build-time option: the Makefile defaults to ISMRMRD=0, so a stock build has no bart ismrmrd command. Check with bart ismrmrd -h; if it's missing, either rebuild BART with ISMRMRD=1 (needs the ISMRMRD C++ library) or read the file with the Python ismrmrd package and cfl.writecfl. Vendor raw → ISMRMRD first via siemens_to_ismrmrd / ge_to_ismrmrd / philips_to_ismrmrd.

1. Estimate coil sensitivities (ESPIRiT):

bart ecalib -m1 -r 24 kspace sens

-m1 matters: ecalib computes two ESPIRiT map sets by default, and pics then returns a soft-SENSE result with a size-2 MAPS dimension instead of a single image — a silent wrong answer. -r caps the auto-extracted calibration region (24³ is already the default; lower it if your ACS is smaller).

2. Reconstruct:

# Fully sampled: inverse FFT + coil combine
bart fft -iu 7 kspace img_coils && bart rss 8 img_coils img

# Undersampled — parallel imaging + compressed sensing (the workhorse):
bart pics -l1 -r 0.01 kspace sens img     # l1-wavelet regularized

• Non-Cartesian (radial/spiral/cones/rosette): you also need the trajectory. Use bart pics -t traj kspace sens img. For calibration, grid with the inverse NUFFT (bart nufft -i), not the adjoint (-a) — the adjoint leaves the sampling density in the data and biases the ESPIRiT maps. Get the trajectory from the sequence/ISMRMRD, or bart traj for nominal.
• EPI is Cartesian. Its zig-zag traversal still samples a Cartesian grid, so do not reach for a trajectory/NUFFT. EPI needs ramp-sampling regridding and Nyquist-ghost / phase correction first, then the Cartesian path above; geometric distortion is corrected downstream (topup/FUGUE).

3. Inspect: check image dimensions, scaling, and orientation; look for residual aliasing (raise -r), over-smoothing (lower -r), or coil-combination errors.

Runnable helper

scripts/bart_recon.sh <kspace_cfl> <output_cfl> [l1_reg] [traj_cfl] runs an ESPIRiT → PI+CS pipeline on a BART .cfl k-space file. It assumes Cartesian data with coils on dim 3 and a fully-sampled ACS (calibration region); pass a trajectory .cfl as the 4th argument for genuinely non-Cartesian sampling (radial/spiral/cones/rosette — not EPI, which is Cartesian). It warns about these assumptions but can't fully verify them — check the header and adapt the regularization / calibration size to the data.

SigPy (Python) alternative

import sigpy as sp, sigpy.mri as mr
maps = mr.app.EspiritCalib(ksp).run()                     # coil maps
img  = mr.app.L1WaveletRecon(ksp, maps, lamda=0.01).run() # PI + CS
# non-Cartesian: build a NUFFT from coords, use mr.app.SenseRecon

Guardrails

• Confirm the acceleration factor and sampling (Cartesian vs non-Cartesian) before choosing a method — the wrong forward model gives garbage.
• If BART isn't installed: https://codeberg.org/mrirecon/bart (source, active) with docs at https://mrirecon.codeberg.page/ — offer to install or fall back to SigPy. Note the github.com/mrirecon/bart mirror is archived as of 2026.
• Report the SNR cost, not just the image. Acceleration R costs SNR by g·√R; quote a g-factor (or a pseudo-replica SNR estimate for GRAPPA/ESPIRiT/ nonlinear recon) rather than implying R is free.
• Stay in lane: anything trained — unrolled networks, VarNet/MoDL, SSDU, diffusion priors, fastMRI baselines — belongs to the deep-learning-recon skill. Designing the acquisition or the trajectory belongs to pulse-sequence-design; this skill consumes a trajectory, it doesn't design one.
• For method theory and citations, see the hub: https://github.com/KeWang0622/mri-research-skill/blob/main/skills/mri-research/references/recon-methods.md and tool details at https://github.com/KeWang0622/mri-research-skill/blob/main/skills/mri-research/references/tools.md