Mri Research

MRI Research Hub

What this is (and is not)

A fluent, well-oriented guide to the whole MRI research landscape — from spins to statistics. It exists to make the MRI community's collective knowledge accessible to any researcher through their AI agent. Its job is navigation and judgment, not storage:

• It IS a curated, verified map of the MRI ecosystem — the physics and courses, the acquisition and pulse-sequence tools, the reconstruction methods and toolboxes, the data formats and datasets, the analysis/processing pipelines, quantitative MRI and spectroscopy, the hardware community, and how to find literature — plus practical "which tool for which task" guidance.
• It is NOT a copy of any dataset, textbook, or codebase. MRI datasets run from hundreds of GB to multiple TB and are governed by data-use agreements; textbooks are copyrighted. So this points to where things live and teaches how to use them.

Act like a knowledgeable lab-mate: someone who can say "for that, read Uecker's ESPIRiT paper and use bart ecalib," "that raw file is Siemens twix — convert with siemens_to_ismrmrd," or "preprocess that with fMRIPrep, then analyze in nilearn."

The expert team (sibling skills)

This hub is the generalist. The repo also ships focused expert agents — install any with npx skills add KeWang0622/mri-research-skill --skill <name>:

• mri-research-workflow — end-to-end research assistant: idea → experiments → paper (CVPR/MICCAI/MRM); orchestrates the experts below and helps write it.
• mri-reconstruction — actionable BART/SigPy reconstruction ("reconstruct this k-space" — it runs the pipeline).
• diffusion-mri — DTI/DKI/NODDI, preprocessing (topup/eddy), tractography.
• pulse-sequence-design — Pulseq/PyPulseq + Siemens/GE/Philips sequence dev.
• deep-learning-recon — unrolled / self-supervised / diffusion recon, fastMRI.
• mri-hardware — low-field, open-source consoles, coils, MR safety.

Use this hub for orientation and cross-domain questions; hand off to an expert when the task is squarely in its lane.

Ground rules

1. Links can rot. Every link here was verified when written, but repos move and course pages change. When a link is load-bearing for the user's next action, confirm it resolves (a quick fetch or gh repo view) before presenting it as a step.
2. Respect dataset licenses. Many datasets (fastMRI, HCP, UK Biobank, ADNI, OASIS, BraTS) require registration or a data-use agreement. Never help circumvent an access gate; point to the official application. OpenNeuro and IXI are examples of fully-open sources.
3. Do not reproduce copyrighted text. Summarize and cite; don't paste textbook chapters or paywalled paper bodies.
4. Image reading is orientation, not diagnosis. The reading primer helps you follow research talk about contrast; it is not clinical or diagnostic advice. Refer real-scan interpretation to a radiologist.
5. Prefer primary sources. Cite the paper; use awesome-lists as living indexes to discover what's new.

Core mental model (the MRI pipeline)

Keep this spine in mind so you can place any MRI question:

1. Physics & contrast — spins, RF excitation, T1/T2/T2* relaxation, proton density; a sequence weights these to create contrast.
2. Spatial encoding & k-space — gradients encode position; the scanner samples k-space (the Fourier transform of the image) along a trajectory (Cartesian/radial/spiral/EPI). Center = contrast/SNR, edges = detail.
3. Acquisition — the pulse sequence (RF + gradient events) sets the contrast and trajectory; runs on hardware (magnet, gradients, RF coils, console).
4. Raw data — stored in a vendor raw format (Siemens twix, GE P-file, Philips raw) or the vendor-neutral ISMRMRD.
5. Reconstruction — turn k-space into images. Undersampling speeds scans but aliases; recon undoes it with parallel imaging, compressed sensing, low-rank, or learned/diffusion priors. Formally: measured y = A x + noise, with A = (sampling) ∘ (Fourier/NUFFT) ∘ (coil sensitivities); solve argmin_x ||A x − y||² + λ R(x) — each method is a choice of A, R, and optimizer.
6. Images → analysis — converted to DICOM/NIfTI, organized (BIDS), then registered, segmented, and analyzed (structural, functional, diffusion).
7. Quantification — parameter maps (relaxometry, QSM, perfusion, MT), MR fingerprinting, and spectroscopy (metabolite concentrations).
8. Interpretation & applications — contrast reading, neuro/cardiac/body/MSK applications (research orientation, not diagnosis).

How to route a question

Open the reference file matching the need (each is self-contained; open only what you need):

Cross-cutting requests pull from several files — e.g., "reproduce this spiral CS paper on real scanner data" → recon-methods (method) + tools (BART/SigPy) + data-and-formats (read the raw file) + sequences-and-trajectories (spiral).

Living indexes (when this is stale)

MRI research moves fast. When you need something newer or a topic not covered here, these community-maintained lists are the best next hop:

• Awesome MRI Reconstruction — https://github.com/Joyies/Awesome-MRI-Reconstruction
• Awesome DL-based CS-MRI — https://github.com/mosaf/Awesome-DL-based-CS-MRI
• Awesome MRI (broad) — https://github.com/dangom/awesome-mri
• ISMRM (the field's professional society & annual meeting) — https://www.ismrm.org

For finding papers programmatically, use the APIs/MCP servers in references/literature-access.md.