PromptMR+

[Paper] [Supplementary] [Slides] [Video]

This repository contains the official implementation of the paper Rethinking Deep Unrolled Models for Accelerated MRI Reconstruction. In this work, we propose effective gradient-based learning and memory-efficient sensitivity map estimation to enhance deep unrolled models for multicoil MRI reconstruction. PromptMR+ is an example of applying these simple yet efficient techniques to the deep unrolled model PromptMR.

News

• Jan 23, 2025: Fully Reproducible code, model weights and reconstruction results were released.
• Oct 10, 2024: PromptMR+ secured 1st place in both tasks of the MICCAI CMRxRecon2024 challenge.
• Aug 12, 2024: Paper accepted as an Oral presentation at ECCV 2024.

Training GPU Memory on cc-brain dataset.

Test GPU Memory on CMRxRecon2023 dataset.

Model Weights

We provide the model weights and reconstruction results for PromptMR/PromptMR+ trained on the FastMRI-knee, FastMRI-brain, Calgary-Campinas-brain, CMRxRecon2023-cardiac and CMRxRecon2024-cardiac datasets, which can be downloaded from the HuggingFace.

CC-brain

Model	Cas.	Trained on	Training Acc	PSNR/SSIM 5x	PSNR/SSIM 10x
PromptMR	12	train	5x and 10x	36.98/0.9496	34.32/0.9302
PromptMR+	12	train	5x and 10x	37.32/0.9516	34.87/0.9350

Note: test result is on a subset of official val set (10/20), results are better than those reported in the paper.

CMRxRecon2023-cardiac

Model	Cas.	Trained on	Training Acc	Cine LAX PSNR/SSIM 10x	Cine SAX PSNR/SSIM 10x	Mapping T1w PSNR/SSIM 10x	Mapping T2w PSNR/SSIM 10x
PromptMR	12	train	4x, 8x and 10x	38.28/0.9560	39.18/0.9615	38.99/0.9661	37.21/0.9622
PromptMR+	12	train	4x, 8x and 10x	39.13/0.9605	39.99/0.9658	40.37/0.9719	38.22/0.9670

Note: test result is on the official validation set.

FastMRI-knee

Model	Cas.	Trained on	Training Acc	NMSE/PSNR/SSIM 4x	NMSE/PSNR/SSIM 8x
PromptMR	12	train	4x and 8x	0.0051/39.71/0.9264	0.0080/37.78/0.8984
PromptMR+	12	train	4x and 8x	0.0050/39.92/0.9276	0.0078/38.09/0.9012

Note: test result is on a subset of official val set (100/199).

CMRxRecon2024-cardiac

Model	Cas.	Trained on	Training Acc	Task1 Avg PSNR/SSIM	Task2 Avg PSNR/SSIM
PromptMR	12	train	4x~24x	38.28/0.9560	39.18/0.9615
PromptMR+	12	train	4x~24x	39.13/0.9605	39.99/0.9658

Note: test result is on the split subset from the official training set (17%). (reported in the STACOM24 paper)

FastMRI-Brain

Model	Cas.	Trained on	Training Acc	NMSE/PSNR/SSIM 4x	NMSE/PSNR/SSIM 8x
PromptMR	12	train+val	4x and 8x	0.0033/41.59/0.9609	0.0063/38.82/0.9465
PromptMR+	12	train	4x and 8x	0.0031/41.84/0.9615	0.0055/39.46/0.9494

Note: test result is on the official test set. (Not reported in the paper)

Install

python>=3.8.16, torch>=2.3, lightning>=2.2.4, wandb>=0.17.0

Data Preparation

check configs/data_split

Train

For example, train PromptMR+ on fastmri-knee dataset

python main.py fit --config configs/base.yaml --config configs/model/pmr-plus.yaml --config configs/train/pmr-plus/fm-knee.yaml 

Training on fastmri-knee requires at least 17G GPU memory if set use_checkpoint and compute_sens_per_coil in config file.

Test

For exmaple, run inference of PromptMR+ on fastmri-knee dataset

python main.py predict --config configs/inference/pmr-plus/fm-knee.yaml

Citation

@inproceedings{xin2025rethinking,
  title={Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction},
  author={Xin, Bingyu and Ye, Meng and Axel, Leon and Metaxas, Dimitris N},
  booktitle={European Conference on Computer Vision},
  pages={164--181},
  year={2025},
  organization={Springer}
}