Generative Image Priors for MRI Reconstruction Trained from Magnitude-Only Images

This folder includes the scripts that implement the workflow proposed in our paper (https://arxiv.org/abs/2308.02340). This project has a certain capacity to handle a large dataset (~100k images) on a Linux-based platform. It provides functionalities for preprocessing the data and training generative models using the dataset, and it was tested on a local GPUs workstation and HPC cluster. With this project, users can efficiently extract prior information from large datasets for MRI reconstruction.

Workflow

1. Preprocess dataset

2. Phase augmentation

3. Train generative priors

4. Image reconstruction with priors

Highlights

1. Distributed training
2. Interruptible training
3. Efficient dataloader for medical images
4. Customizable models with a configuration file
5. Parallelized processing

Related repositories

1. bart
2. spreco
3. bart tutorials

Citation

1. Luo, G, Blumenthal, M, Heide, M, Uecker, M. Bayesian MRI reconstruction with joint uncertainty estimation using diffusion models. Magn Reson Med. 2023; 1-17
2. Blumenthal, M, Luo, G, Schilling, M, Holme, HCM, Uecker, M. Deep, deep learning with BART. Magn Reson Med. 2023; 89: 678- 693.
3. Luo, G, Zhao, N, Jiang, W, Hui, ES, Cao, P. MRI reconstruction using deep Bayesian estimation. Magn Reson Med. 2020; 84: 2246-2261.