model card

README.md

@@ -0,0 +1,40 @@
+---
+license: cc-by-nc-4.0
+library_name: pytorch
+tags:
+- mri-reconstruction
+- medical-imaging
+- accelerated-mri
+- looped-transformer
+---
+
+# LoopMR weights
+
+Trained models of LoopMR, a looped (weight-tied) SDUM-Restormer for multicoil MRI reconstruction: one denoiser block applied 12 times
+with E2E-VarNet's k-space state and soft data consistency, PromptMR+'s cascade inputs and sensitivity estimator, 4 carried latent
+images and 5 adjacent slices (CMRxRecon: 5 adjacent time frames). 10.27M parameters. Trained with a fixed depth of 12 loops, gradient
+through all 12, 1 - SSIM on the final output.
+
+The files are read by the loopmri code (`loopmri.evaluate.load_model`), which downloads them to its `weights/` folder on first use.
+
+| file | data | test results (SSIM / PSNR dB) |
+|---|---|---|
+| `cc359_rs5_fd_146k.pt` | CC359 12-channel brain, 2-D Poisson-disc R 5 / 10 | PromptMR+'s 10 test volumes, challenge metric: R 5 0.9528 / 37.45, R 10 0.9377 / 35.12 |
+| `knee5_rs5_fd_100k.pt` | fastMRI knee (159 train volumes of `multicoil_val`), random lines R 4 / 8 | 40 held-out volumes: 0.9326 / 39.35 (R 4 0.9424 / 40.44, R 8 0.9161 / 37.54) |
+| `cmr23_rs5_fd_466k.pt` | CMRxRecon2023 cine / mapping (P001-P100), equispaced R 4 / 8 / 10 | 76 held-out files (P101-P120), challenge ranking protocol: 0.9629 / 38.62 |
+| `knee5_sens_prefit/` | sensitivity estimator pre-fitted to plain calibration maps | the starting point of the knee training run |
+
+Each `.pt` holds `model` (state dict), `cfg` (the training config) and `step`.
+
+## Training
+
+AdamW lr 3e-4 (estimator 1e-4, CMRxRecon 2e-4), weight decay 0.01, 500 warm-up steps, gradient-norm clip 0.01, batch 1, bf16 inside the
+block, single seed. The knee model was trained exactly as the code's `configs/knee5_rs5_fd.yaml`. The CC359 and CMRxRecon models used
+the same model, objective, optimiser and schedule, except that their first 90,000 (CC359) / 350,000 (CMRxRecon) steps used a
+sampled-depth objective (4-32 loops per step, gradient through the last 8, every iterate scored) before fixed-depth training took over
+to the same final step.
+
+## Licence and data
+
+CC BY-NC 4.0 (non-commercial). The models were trained on Calgary-Campinas (CC359), fastMRI and CMRxRecon2023 data, whose terms of use
+apply, and use PromptMR+'s sensitivity estimator (Rutgers non-commercial research licence). Research use only, not for clinical use.