One Block Is Enough: LoopMR for Accelerated MRI Reconstruction
Weights of LoopMR. LoopMR reuses a single Restormer block at every iteration of an unrolled reconstruction. Besides the current image estimate, the block receives data-consistency inputs and latent images. It uses 12 iterations, 4 latents and 5 adjacent slices, with sensitivity maps from PromptMR+'s sensitivity map estimator (SME), and has 10.27M parameters.
Training and evaluation code: the loopmri repository, which downloads these files to its weights/ folder on first use.
| file | dataset |
|---|---|
cc359_rs5_fd_146k.pt |
Calgary-Campinas (CC359) brain, ×5 / ×10 |
knee5_rs5_fd_100k.pt |
fastMRI knee, ×4 / ×8 |
cmr23_rs5_fd_466k.pt |
CMRxRecon2023 cine and mapping, ×4 / ×8 / ×10 |
knee5_sens_prefit/ |
the SME pre-fit the knee training starts from |
promptmr_plus/knee5_promptmr_plus_100k.pt |
PromptMR+ baseline, fastMRI knee |
promptmr_plus/cmr23_promptmr_plus_466k.pt |
PromptMR+ baseline, CMRxRecon2023 |
The two PromptMR+ models are the baselines of the paper's tables, trained by us with the authors' settings on the same data and schedule as LoopMR, because the authors' released knee and CMRxRecon2023 weights were trained on other data (fastMRI's full knee training set, and CMRxRecon2023 P001-P120, which includes the test patients). On CC359 the paper uses the authors' released weights (hellopipu/PromptMR), which are not copied here: the loopmri code downloads and converts them.
Results
PSNR (dB) / SSIM on the test sets, LoopMR and PromptMR+ both evaluated by the loopmri code, each dataset with its official protocol (CMRxRecon2023: all 76 test files at every R, as in the challenge). CC359 and knee were computed on an H100 and CMRxRecon2023 on an A100. The block runs in bf16, so on another GPU type PSNR can differ by about 0.01 dB in these means.
| dataset | R | LoopMR | PromptMR+ |
|---|---|---|---|
| CC359 brain | ×5 | 37.45 / 0.9528 | 37.28 / 0.9513 |
| ×10 | 35.12 / 0.9377 | 34.83 / 0.9346 | |
| fastMRI knee | ×4 | 40.44 / 0.9424 | 40.34 / 0.9410 |
| ×8 | 37.54 / 0.9161 | 37.39 / 0.9138 | |
| CMRxRecon2023 heart | ×4 | 41.93 / 0.9800 | 41.66 / 0.9788 |
| ×8 | 37.24 / 0.9559 | 36.82 / 0.9524 | |
| ×10 | 36.00 / 0.9485 | 35.55 / 0.9440 |
Licence
CC BY-NC 4.0. The models were trained on CC359, fastMRI and CMRxRecon2023 data, whose terms of use apply, and use PromptMR+'s code (Rutgers non-commercial research licence): its SME in LoopMR, the whole network in the PromptMR+ baselines. Research use only, not for clinical use.