UnrollingNets-dMRI

( If you find this code is helpful to you, could you please kindly give me a star )

This repo provides the TensorFlow 2.x implementation for several SOTA unrolling networks in accelerated MRI reconstruction.

It contains JotlasNet (MRI 2025), T2LR_Net (CIBM 2024), L+S_Net (MedIA 2021), SLR_Net (TMI 2021), ISTA_Net (CVPR 2018), DCCNN (TMI 2017).

Method	Paper	GitHub
JotlasNet (MRI 2025)	Link	Link
T2LR_Net (CIBM 2024)	Link	Link
L+S_Net (MedIA 2021)	Link	Link
SLR_Net (TMI 2021)	Link	Link
ISTA_Net (CVPR 2018)	Link	Link
DCCNN (TMI 2017)	Link	Link

1. Getting Started

Environment Configuration

• we recommend to use docker

  # pull the docker images
  docker pull yhaoz/tf:2.9.0-bart
  # then you can create a container to run the code, see docker documents for more details
  

• if you don't have docker, you can still configure it via installing the requirements by yourself

  pip install -r requirements.txt # tensorflow is gpu version
  

Note that, we only run the code in NVIDIA GPU. In our implementation, the code can run normally in both Linux & Windows system.

Dataset preparation

Actually, two dataset involves in this code, i.e., OCMR and CMRxRecon. Here we just provides our pre-processing pipeline for OCMR dataset. Since CMRxRecon is a MICCAI challenge and has its own leader board, we do not provide our own code.

You could find the the single-coil and multi-coil dataset pre-processing and creating code in yhao-z/ocmr-preproc-tf. You may need to put the pre-processed data into a file folder that contains four subfolders.

# the data needs to be arranged into four sub file folders, and you may set the datadir in the code.
- train
- val
- test
- masks

2. Run the code

Test only

We provide the training weights on OCMR dataset for many sampling cases and both single-coil and multi-coil scenarios as in weights-ocmr. Note that the provided weights are only applicable in our data pre-processing implementation. If you are using other different configuration, retraining from scratch is needed.

# Please refer to main.py for more configurations.
python main.py --mode 'test'

Training

# Please refer to main.py for more configurations.
python main.py --mode 'train'

3. To Do

• Add VISTA mask generation code.
• OCMR has release more data in last three years, but this code still uses the initial released limited data. Add more training data or make a new dataset in the future.

4. Note

• If this code have any error, feel free to raise an issue.
• If you have built your own model with this code, welcome to pull request and merge your model into this ModelZoo, of course, if you want.