v0.8.0 — Evidence-backed MRI research and paper writing

MRI research skills, from experiment evidence to a manuscript

• Practical MICCAI, MRM, JMRI and IEEE TMI writing guides with official author-policy links.
• Editable paper-plan and reviewer-response templates; an illustrative diffusion study links claims, experiments and figures without inventing results.
• Project-local .mri-research/ memory for preferences, run records and evidence-linked lessons.
• Expanded DWI/DTI, EPI and DENSE guidance; task-specific evaluation and original-source attribution.
• Annotated paper/textbook reading lists for all seven skills and a central reference index.
• Agent-owned tool setup using established implementations; cross-skill links that work in standalone installations.
• Refreshed README with the MRI mascot and hosted tutorial links. Large presentation HTML stays outside this repository.

These are research guidance and templates, not a validation claim for every workflow. Follow current venue instructions and verify scientific outputs for the intended task.

Validation: repository structure/frontmatter, existing project-memory unit tests, skill validation and standalone Markdown reference checks.

v0.7.0 — verified review round

Install: npx skills add KeWang0622/mri-research-skill -g

A second adversarial multi-agent review round, this time with every consequential
claim checked against a primary source (Crossref, PyPI, the GitHub/Codeberg APIs)
rather than taken from a reviewer's word. The headline finding: BART moved to
Codeberg and the GitHub repository is archived, so the repo's central execution
dependency was pointing at a frozen tree.

Fixed — correctness (found by verification, not opinion)

• BART now points at its active home, https://codeberg.org/mrirecon/bart
  (docs https://mrirecon.codeberg.page/), everywhere it appears; the archived
  github.com/mrirecon/bart mirror is labelled as such.
• ecalib two-map bug. bart ecalib returns two ESPIRiT map sets by
  default, so pics emitted a soft-SENSE image with a size-2 MAPS dimension — a
  silent wrong answer. bart_recon.sh, the hub, and mri-reconstruction now all
  pass -m1 and explain why (and note that -r 24 is already BART's default).
• Non-Cartesian calibration used the wrong NUFFT. Gridding for ESPIRiT now
  uses the inverse NUFFT (nufft -i); the adjoint (-a) leaves sampling
  density in the data and biases the coil maps.
• EPI is Cartesian. The docs no longer imply EPI needs a trajectory/NUFFT; it
  needs ramp-sampling regridding plus Nyquist-ghost/phase correction upstream.
• bart_recon.sh hygiene: refuses to overwrite an existing output, writes
  intermediates to mktemp scratch and cleans them up on exit, and prints the
  output dimensions so a stray MAPS dimension is visible.
• Raw-data ingest was wrong in two places. bart twixread needs -A (or
  explicit -x/-y/-z/-c/-s/-n); cfl.py ships in BART's python/ directory and
  is not on PyPI (the bartpy package there is an unrelated decision-tree
  library); and bart ismrmrd only exists when BART is built with ISMRMRD=1
  (the Makefile defaults to ISMRMRD=0).
• MaRCoS was miscredited. The system paper is Negnevitsky V, et al.,
  J Magn Reson 2023;350:107424 (doi:10.1016/j.jmr.2023.107424); Guallart-Naval
  et al. is the multi-site benchmarking paper (NMR Biomed 2023;36(1):e4825).
• SPIRiT was described as the SENSE/GRAPPA bridge — that is ESPIRiT. SPIRiT is
  GRAPPA generalized to enforce calibration consistency everywhere, with no
  explicit sensitivity maps.
• mridc is archived (read-only since Apr 2024) and its author redirects to
  ATOMMIC (https://github.com/wdika/atommic); all three mentions updated.
• Papers with Code has been retired — paperswithcode.com now redirects to
  Hugging Face Papers; the workflow agent says so instead of linking a dead index.
• SPM is no longer "SPM12" — development moved to github.com/spm/spm with
  calendar versioning (25.01.02 stable, 26.01 in RC), which is not path- or
  script-compatible with SPM12.
• Removed a garbled GE parenthetical (the GEHealthcare org has no public repos)
  and a private note that had leaked into foundations.md.
• Dropped a stray trailing whitespace and pinned down Tarquin's status (source at
  martin3141/tarquin, no commits since 2021 — dormant, baseline only).

Added

• The SNR cost of acceleration, which the repo previously never quoted: a
  g-factor section (SNR_accel = SNR_full/(g·√R)), noise prewhitening / optimal
  coil combination (Roemer 1990), and pseudo-replica SNR for GRAPPA/ESPIRiT/
  nonlinear recon where g has no analytic form (Robson 2008). The recon agent is
  now required to report it.
• Structured low-rank split into its three distinct ideas — SAKE
  (block-Hankel, inter-coil dependency), LORAKS (limited support + phase, plus
  P-LORAKS), ALOHA (annihilating filters) — each with its own citation.
• SMASH (the k-space ancestor GRAPPA generalizes) and a fuller
  partial-Fourier entry with the real-image/Hermitian caveat (Noll 1991).
• Controlled aliasing, properly separated: CAIPIRINHA (multi-slice and 2D),
  blipped-CAIPI for SMS-EPI, and Wave-CAIPI for 3D.
• Dixon/IDEAL untangled: IDEAL is the species decomposition, graph cuts solve
  the field map and its water/fat swap ambiguity, and quantitative PDFF needs a
  multi-peak fat model with simultaneous R2*.
• ASL needs a kinetic model: the subtraction image is not CBF (Buxton 1998),
  plus PCASL/PASL/VSASL, the post-labeling delay, the consensus implementation
  (Alsop 2015), and ASLPrep.
• QSM dipole inversion is the ill-posed step (zero cone in k-space) — review
  and MEDI citations, with a note on why two pipelines disagree.
• MRSHub (https://mrshub.org) as the MRS community index, and a full citation
  for HFS-SDE diffusion recon.
• Explicit, bidirectional hand-offs in every expert skill, so trained recon,
  classical recon, acquisition design, hardware, and image analysis each have one
  unambiguous owner. The hub now states that it owns image-level analysis
  (fMRI/GLM, BIDS, DICOM/NIfTI, FreeSurfer, segmentation, registration) — no
  sibling skill covered it, and the hub's description didn't trigger on it.

Changed

• The hub description now fits the 1024-character limit skill validators
  enforce (it was over) and includes the analysis triggers it was missing.
• All seven skills are on version 0.7.0 (previously five were still on 0.1.0).

CI

• validate_repo.py now enforces what this round had to catch by hand:
  description length ≤ 1024, one source of truth for the version across all seven
  skills + CITATION.cff + README, frontmatter name matching its directory, and
  routing-table rows that resolve to real files.
• New freshness job: flags linked upstreams that have been archived, renamed,
  or moved (this is what would have caught the BART migration), and checks that
  every PyPI package and DOI the repo names actually exists.
• Pinned the third-party ShellCheck action to a commit SHA.

v0.6.0 — More tools, red-team pass, branch protection

More tools, a multi-agent red-team pass, and stronger governance

More tools

• Sequence design: gradient/trajectory optimization (GrOpt, Lustig
  minTimeGradient, safe_pns_prediction), GIRF (MRI-gradient/GIRF, GIRFReco.jl),
  Pulseq ecosystem (TOPPE for GE, pulseq-CEST, MR-Physics-with-Pulseq),
  simulators (JEMRIS, MRiLab, sycomore, EPG-X, MRzero-Core), and RF design
  (pulpy, Spectral-Spatial-RF-Pulse-Design, Multiband-RF, kpTx).
• Hardware / RF: coil design (CoilGen, pyCoilGen), EM/SAR (MARIE, mariepy,
  CoSimPy, scikit-rf), B0 shimming (Shimming Toolbox), the MaRCoS ecosystem
  (marcos_client/server, marga, ocra-pulseq), GPA-FHDO, MRI4ALL, and OSI² ONE.
• Reconstruction / quantitative: CG-SENSE, partial-Fourier; fat–water
  (Dixon / IDEAL / PDFF) and CEST.

Multi-agent red-team pass — three agents (two sourcing/verifying tools, one
adversarial reviewer) drove concrete fixes: the bundled bart_recon.sh is now
honest about its assumptions (Cartesian, coils on dim 3, ACS required) and
handles non-Cartesian data via an explicit trajectory path — no more "just works"
overclaim; the generalist hub was rescoped so it no longer over-triggers against
the experts; and several gaps and citations were tightened.

Governance & CI — main is now branch-protected (PRs must pass the CI
checks before merge), and CI gained committed repo-structure validation and a
"skills CLI detects all skills" job.

Install unchanged:

npx skills add KeWang0622/mri-research-skill --all

v0.5.1 — Industry-standard hardening (community health + CI)

Industry-standard hardening

Community health + CI + a clean audit — no changes to skill content.

• Community health: Code of Conduct (Contributor Covenant 2.1), Security
  policy, issue templates, and a PR template — GitHub "community standards" green.
• CI (.github/workflows/validate.yml): validates skills.sh.json and every
  SKILL.md's frontmatter, runs ShellCheck on the helper script, and
  link-checks every Markdown file (tolerant of publisher bot-blocks) on each
  change and weekly. The full pipeline is green.
• Audit: full link sweep — 172 URLs return 200; the rest are publisher pages
  that bot-block (403/429) but load in a browser. Two third-party links with
  broken TLS certs are excluded from CI (kept in the docs) via a documented
  .lycheeignore. Fixed the README stars-badge link.
• Added .editorconfig.

Install is unchanged:

npx skills add KeWang0622/mri-research-skill --all

v0.5.0 — End-to-end research assistant (idea → paper)

End-to-end research assistant

v0.5.0 adds mri-research-workflow — an agent that takes an MRI (or MRI-ML)
project from idea to a published paper, and helps you write it.

It shepherds the whole flow — literature → hypothesis → experiment design →
running experiments → analysis → figures/tables → manuscript → submission &
rebuttal — and is venue-aware: CVPR / NeurIPS, MICCAI, and Magnetic
Resonance in Medicine (MRM), each with its own template, review style, and
emphasis. It orchestrates the expert agents (e.g., mri-reconstruction to run
BART/SigPy, deep-learning-recon to train) and points to the right LaTeX
template (cvpr-org/author-kit, Springer
LNCS, the MRM class) and reproducibility standards (Papers with Code, CLAIM,
COBIDAS).

Install it, or the whole team:

npx skills add KeWang0622/mri-research-skill --skill mri-research-workflow
npx skills add KeWang0622/mri-research-skill --all

All links verified. The repo now ships seven skills (generalist hub +
research-workflow orchestrator + five domain experts).

v0.4.0 — Multi-agent MRI expert team

A multi-agent MRI team

v0.4.0 turns the repo into a team of expert agents you install with one
command. Beyond the generalist mri-research hub, five focused experts:

• mri-reconstruction — actionable: runs ESPIRiT + PI/CS with BART/SigPy on
  your k-space (bundled scripts/bart_recon.sh). Say "reconstruct this k-space
  with BART" and it runs the pipeline.
• diffusion-mri — DTI/DKI/NODDI, preprocessing (topup/eddy), CSD,
  tractography (MRtrix3, DIPY, FSL, AMICO, TractSeg, QSIPrep).
• pulse-sequence-design — Pulseq/PyPulseq + KomaMRI, RF/SMS, and
  Siemens (IDEA/ICE) / GE (EPIC/Orchestra) / Philips (Paradise) sequence dev.
• deep-learning-recon — unrolled / self-supervised / diffusion recon;
  DIRECT, fastMRI, mridc.
• mri-hardware — low-field, open-source consoles, coils, MR safety.

Install the whole team, or just one expert:

npx skills add KeWang0622/mri-research-skill --all
npx skills add KeWang0622/mri-research-skill --skill mri-reconstruction

All links verified; every expert cross-links back to the hub's reference files.

v0.3.0 — Deeper coverage + publishing guidance

Deeper coverage + publishing guidance

v0.3.0 acts on a full review pass (P0 + P1) and adds a new publishing reference.

New — publishing

• MR journals + author guidelines (MRM, JMRI, NMR in Biomedicine, MAGMA, IEEE
  TMI, Medical Image Analysis, NeuroImage, Imaging Neuroscience, Radiology),
  LaTeX templates (official MRM class, IEEEtran, elsarticle, Springer, Overleaf,
  arXiv), reporting/reproducibility standards (COBIDAS, CLAIM, TRIPOD+AI, ISMRM
  RRSG), and abstract submission.

Deepened

• Acquisition: SMS/multiband (Moeller 2010, Setsompop 2012, CMRR), B0 field
  mapping & distortion correction (FSL topup/FUGUE), B1 mapping, off-resonance
  correction, prospective motion.
• Reconstruction: GRASP/XD-GRASP (motion-resolved), subspace/low-rank qMRI,
  denoising (NORDIC, MP-PCA, Patch2Self), and an evaluation & image-quality-
  metrics section (with a DL-hallucination caveat).
• Analysis: quality control & motion (MRIQC, MCFLIRT); cardiac/body/MSK +
  radiomics (pyradiomics; ACDC and M&Ms benchmarks).
• Data: ACDC and M&Ms cardiac datasets.
• Hardware: a dedicated, non-clinical MRI safety section.
• README: a Mermaid pipeline diagram.

All new links verified. Install: npx skills add KeWang0622/mri-research-skill.

v0.2.0 — General MRI + npx skills

Now a general MRI research hub

v0.2.0 broadens the project from reconstruction-focused to the whole MRI
pipeline — physics, acquisition, reconstruction, analysis, quantitative MRI,
spectroscopy, and hardware — and switches distribution to the open skills CLI.

Install

npx skills add KeWang0622/mri-research-skill

Works across Claude Code, Codex, Cursor, OpenCode, and many more agents.

What's new

• New references: image analysis & processing (structural / functional /
  diffusion MRI, segmentation, registration — FreeSurfer, FSL, SPM, AFNI, ANTs,
  fMRIPrep, MRtrix3, DIPY, nnU-Net, MONAI) and quantitative MRI &
  spectroscopy (relaxometry, QSM, perfusion/ASL, MT; LCModel, Osprey, FSL-MRS).
• Expanded data & formats: DICOM/NIfTI/BIDS plus image-level datasets
  (HCP, OpenNeuro, UK Biobank, ADNI, OASIS, IXI, BraTS) with access-term notes.
• Expanded foundations: canonical handbooks (McRobbie, Westbrook, Bernstein
  Handbook of MRI Pulse Sequences, Haacke).
• Restructured to the skills layout with skills.sh.json; MRI-first README
  with working badges; every link verified.

See CHANGELOG.md for the full list.

v0.1.0 — Initial release

First public release of the mri-research Claude skill — a curated, verified
reference hub that turns Claude into a fluent MRI reconstruction research
assistant.

Highlights

• Navigator SKILL.md — a core MR mental model, ground rules (link rot,
  dataset licenses, non-diagnostic scope, no copyrighted text), and a
  question → reference routing table.
• Eight reference files:
  • foundations — MR physics / k-space and where to learn (Berkeley EE225E,
    Stanford EE369B/C, Hornak, MRIquestions, ISMRM).
  • recon-methods — landmark-paper reading list: parallel imaging (SENSE,
    GRAPPA, ESPIRiT, SPIRiT, NLINV), compressed sensing, low-rank & dynamic
    (L+S, k-t), deep learning (VarNet, MoDL, E2E-VarNet, SSDU, AUTOMAP),
    diffusion / score-based recon, and MR fingerprinting — with a "which method
    for which task" guide.
  • tools — decision guide + usage for BART, SigPy, MIRT.jl, MRIReco.jl,
    torchkbnufft, gpuNUFFT, mri-nufft, pygrappa, DIRECT, fastMRI, DeepInPy,
    TensorFlow MRI, Gadgetron.
  • data-and-formats — ISMRMRD & vendor raw (Siemens twix, GE P-file, Philips
    raw) + open datasets (fastMRI, mridata.org, OCMR, SKM-TEA, Calgary-Campinas,
    CMRxRecon, M4Raw).
  • sequences-and-trajectories — Pulseq/PyPulseq, vendor SDKs, RF pulse design,
    trajectory design, KomaMRI simulation.
  • hardware — OSI², MaRCoS, OCRA, low-field, coils/gradients, safety.
  • literature-access — APIs, API keys, and paper-search MCP servers.
  • radiology-primer — T1/T2/FLAIR/DWI contrast, non-diagnostic background.

Philosophy

A pointer / reference skill: it links to and credits the community's
primary papers, tools, courses, and datasets rather than bundling datasets or
copyrighted text.

See CHANGELOG.md for the full list.