MRI reconstruction stops learning from clinical raw data: SIM trains on physics simulation and transfers zero-shot across 14 datasets

Accelerated MRI reconstruction changes its training source: instead of clinical raw k-space gathered per institution, the model learns from samples generated by the physics of the imaging process and reconstructs directly on unseen scanners, populations, anatomies and protocols. Supervised reconstruction models degrade sharply outside their training distribution, while untrained methods rely on severely undersampled single-case measurements and are capped by them; existing simulation-based methods mostly only matched models trained on moderate volumes of real data. Of the 14 datasets, only fastMRI Brain serves as the in-distribution test bed and the other 13 are out-of-distribution zero-shot tests; SIM outperforms every competing baseline on every out-of-distribution benchmark, exceeding the second-best method by 1.06 to 3.95 dB in PSNR and 0.008 to 0.045 in SSIM.

The simulation engine treats arbitrary natural images as imaging objects and synthesizes undersampled measurements with fully sampled references through the multi-coil MRI forward encoding, so the model learns the inverse mapping rather than anatomy- or contrast-specific correlations. The simulation-to-real gap is narrowed in two ways: noise levels and the undersampling operator are matched to evaluation, and a fresh coil configuration is randomly generated for each training sample instead of reproducing fixed hardware. Training data come from SA-1B natural images with randomly assigned real and imaginary channels plus phase and magnitude alignment; evaluation spans M4Raw at 0.3 T through AHEAD at 7 T and PCrCEST at 5 T, and covers brain, knee, spine, heart and liver.

Model capacity and data volume scale together: on simulated data, test loss falls smoothly as dataset and model size grow in tandem, following an approximate power-law relationship. Scaling model size on the largest public fastMRI Brain real dataset yielded no sustained improvement and pronounced overfitting; real data carry substantial redundancy in anatomy, contrast and acquisition characteristics. Models up to 567 million parameters trained on 1.1 million simulated samples kept gaining on real out-of-distribution test sets including CC-359, AHEAD Echo1 and Stanford 2D without saturation, whereas models trained solely on real data degraded beyond modest size on out-of-distribution evaluations.

Validation centers on retrospective undersampling of 2D multi-slice, magnitude-only images: fastMRI Brain is the only in-distribution training and test source among the 14 datasets, and the other 13 are used for zero-shot transfer; clinical usability comes from a blinded reader study of 150 scans from 50 subjects in an Alzheimer's disease cohort at Peking University Shenzhen Hospital. The authors call for prospective, multicenter studies to evaluate integration with clinical scanners and establish clinically appropriate acceleration factors.

All acceleration evaluations use retrospective undersampling and have not been validated prospectively in a live scanning workflow; the study is limited to 2D multi-slice, magnitude-only reconstruction, leaving 3D isotropic and phase-sensitive imaging uncovered; training data contain no medical anatomy, and the authors note that the precise extent to which this improves fairness warrants systematic evaluation.