chunmeifeng / DONet

【TNNLS 2021】DONet: Dual-Octave Network for Fast MR Image Reconstruction

DONet

DONet: Dual-Octave Network for Fast MR Image Reconstruction (IEEE Transactions on Neural Networks and Learning Systems)

Dependencies

• Python 3.7
• Tensorflow 1.14
• numpy
• h5py
• skimage
• matplotlib
• tqdm

Install dependencies as follows:

wget https://repo.anaconda.com/archive/Anaconda3-2020.07-Linux-x86_64.sh
bash Anaconda3-2020.07-Linux-x86_64.sh.sh
source ~/.bashrc
conda install python=3.7
conda install tensorflow-gpu==1.14
conda install numpy
conda install scikit-learn
conda install scikit-image
conda install tqdm
conda install opencv

Dataset and Prepartion

All data that we used for our experiments are released at GLOBUS(https://app.globus.org/file-manager?origin_id=15c7de28-a76b-11e9-821c-02b7a92d8e58&origin_path=%2F). Before training, we recommend you to process data into .tfrecords to accelerate the progress. File ./data_preparation/data2tfrecords.py specifies the route of data processing.

How to train and test on DONet

Unpack the dataset file to the folder you defined. Then, change the data_dst argument in ./option.py to the place where datasets are located.

Enter in the folder /DONet/code

Train

CUDA_VISIBLE_DEVICES=0 python run.py --n_GPU 1 --name Dual-Oct_dense_B10_lrb3_a0.125_cpd320_1Un3X --n_blocks 10 --n_feats 64 --lr 1e-3 --alpha 0.125 --data_dst coronal_pd_320 --epoch 50 --mask_name 1Un3_320

Test

CUDA_VISIBLE_DEVICES=0 python tester.py --n_GPU 1 --rsname Dual-Oct_dense_B10_lrb3_a0.125_cpd320_1Un3X --n_blocks 10 --n_feats 64 --alpha 0.125 --data_dst coronal_pd_320 --mask_name 1Un3_320 --test_only --save_gt --save_results

Change other arguments in option.py or in the shell that you can train your own model.

If one GPU will be out of memory, you can change the --n_blocks and --n_feats to compress the model, empirically we set --n_resblocks 10 and --n_feats 64. Moreover, you can set CUDA_VISIBLE_DEVICES=0, 1, --n_GPU 2and try using two or more GPUs for this training.

Citation

If you find DONet useful for your research, please consider citing the following papers:

@inproceedings{feng2021DONet,
  title={DONet: Dual-Octave Network for Fast MR Image Reconstruction},
  author={Feng, Chun-Mei and Yang, Zhanyuan and Fu, Huazhu and Xu, Yong and Yang, Jian and Shao, Ling},
  booktitle={IEEE Transactions on Neural Networks and Learning Systems},
  year={2021}
}