wenqihuang / LS-Net-Dynamic-MRI

This the code repository for paper "Deep Low-rank plus Sparse Network for Dynamic MR Imaging".

Deep Low-rank plus Sparse Network (L+S-Net) for Dynamic MR Imaging

This repository provides a tensorflow implementation used in our publication

Huang, Wenqi, et al., Deep low-rank plus sparse network for dynamic MR imaging., Medical Image Analysis 73 (2021): 102190.

If you use this code and provided data, please refer to:

@article{huang2021deep,
  title={Deep low-Rank plus sparse network for dynamic MR imaging},
  author={Huang, Wenqi and Ke, Ziwen and Cui, Zhuo-Xu and Cheng, Jing and Qiu, Zhilang and Jia, Sen and Ying, Leslie and Zhu, Yanjie and Liang, Dong},
  journal={Medical Image Analysis},
  volume={73},
  pages={102190},
  year={2021},
  publisher={Elsevier}
}

Requirements

The framework was build on Tensorflow 2.4 and roughly tested on Tensorflow 2.7.

We provide an requirements file requirements.txt. You can create a new conda environment or virtual environment and execute install the requirements using the following command.

pip install -r requirements.txt

If you want to visualize the dynamic images in gif in tensorboard, please install ffmpeg in advance. On Ubuntu, it can be simply installed by

sudo apt install ffmpeg

Training the L+S-Net usually needs large GPU memory. If you come across the OOM problem, please try to reduce the niter and n_f of the network, or reduce the number of coils by coil compression.

Data

We provided a jupyter notebook for creating a dummy dataset. The corresponding dataloader is also provided in dataset.py. New dataset can be easily made by modifying the demo code.

Training

After creating the dummy dataset, the network can be easily traind by

python main.py --gpu=0 --data='DUMMY' --niter=10 --num_epoch=50

in which --gpu specifies the GPU for training, --data specifies the dataset used for training, --niter is for number of iteration blocks and --num_epoch specifies the training epoches.

Please refer to main.py for more configurations.

You can oberserve the progress of the training in Tensorboard using the specified log directory. The command is as follow

tensorboard --logdir=./logs

Model files will be stored in ./models

Testing

To test a trained model with the name <MODEL_ID>

python test.py [SAME_PARAMS_AS_TRAINIGN] --weight=models/<MODEL_ID>/epoch-50/ckpt

Please note that do not omit the ckpt at the end of the weight path.