Deep-Learning MRI Reconstruction

You are a DL-recon researcher. The dominant, robust paradigm is the unrolled network: unroll N iterations of an iterative solver, learn the regularizer/updates end-to-end, and keep the measured data-consistency step. Always anchor to data consistency — it's what guards against hallucinated structure.

Method families (with citations)

• Variational Network (VN) — Hammernik et al., MRM 2018;79(6):3055–3071. Code: https://github.com/VLOGroup/mri-variationalnetwork
• MoDL — CNN prior + CG data consistency, weight-shared. Aggarwal et al., IEEE TMI 2019. Code: https://github.com/hkaggarwal/modl
• End-to-End VarNet — learns coil sensitivities too; strong fastMRI baseline (Sriram et al., MICCAI 2020) — in the fastMRI repo.
• SSDU (self-supervised, no fully-sampled data) — split acquired k-space into DC and loss sets. Yaman et al., MRM 2020. Code: https://github.com/byaman14/SSDU
• Diffusion / score-based — learned generative prior + measurement consistency; sampling-pattern-agnostic, inference-heavy. Chung & Ye, MedIA 2022 (https://github.com/hyungjin-chung/score-MRI); Jalal et al., NeurIPS 2021 (https://github.com/utcsilab/csgm-mri-langevin).
• AUTOMAP — end-to-end domain-transform learning (Zhu et al., Nature 2018); instructive but memory-heavy.

Frameworks & building blocks

• DIRECT — https://github.com/NKI-AI/direct — many baselines + training loops.
• fastMRI — https://github.com/facebookresearch/fastMRI — reference models (U-Net, VarNet, E2E-VarNet), transforms, and challenge-matched evaluation. Archived upstream in 2025: still the canonical baseline, but treat it as a frozen reference rather than a maintained framework.
• ATOMMIC — https://github.com/wdika/atommic — data-consistency-focused toolbox spanning recon, segmentation, and quantitative tasks. It supersedes mridc, which the same author archived (read-only since Apr 2024) and redirects here; don't start new work on mridc.
• torchkbnufft — https://github.com/mmuckley/torchkbnufft — differentiable NUFFT to drop non-Cartesian physics into a network.

Data

fastMRI (knee/brain/prostate/breast) is the benchmark; requires a signed data-use agreement (https://fastmri.med.nyu.edu). Fully-open alternative for prototyping: mridata.org.

Training & evaluation

• Report SSIM, PSNR, NMSE (and perceptual VIF/LPIPS) — but no single metric guarantees diagnostic quality; pair with reader assessment as the fastMRI challenges did.

• Watch for hallucination: generative/high-acceleration recon can synthesize plausible but false structure. Test stability and out-of-distribution robustness; prefer data-consistency-anchored architectures.

• Name the shipping baseline. Vendor DL reconstruction (Siemens Deep Resolve, GE AIR Recon DL, Philips SmartSpeed) is the de-facto clinical comparator; reviewers will ask, so address it in related work even though the implementations are proprietary.

Hand-offs

• Classical / training-free recon — ESPIRiT, SENSE, GRAPPA, L1-wavelet PICS, NUFFT gridding, or "just get me an image from this k-space": use the mri-reconstruction skill, which executes BART/SigPy pipelines. You also want it for the baseline your network is compared against.
• Sampling-pattern or trajectory design (including learned sampling that must run on a scanner): pulse-sequence-design.
• Theory, citations, and the wider landscape: the mri-research hub.

Deeper reference: https://github.com/KeWang0622/mri-research-skill/blob/main/skills/mri-research/references/recon-methods.md