DD-CSENet: Dual-Domain Cross-Iteration Squeeze-Excitaton Network for Sparse-View Reconstruction of MRI

Xiongchao Chen, Yale University

This repository contains the PyTorch implementation of Dual-domain Iterative residual sqeze-excitaton network for sparse-view reconstruction of MRI.

Citation

If you use this code for your research or project, please cite:

  @article{chen2022dual,
    title={Dual-Domain Cross-Iteration Squeeze-Excitation Network for Sparse Reconstruction of Brain MRI},
    author={Chen, Xiongchao and Shinagawa, Yoshihisa and Peng, Zhigang and Valadez, Gerardo Hermosillo},
    journal={arXiv preprint arXiv:2210.02523},
    year={2022}
  }

Environment and Dependencies

Requirements:

• Python 3.8.13
• PyTorch 1.11.0
• numpy 1.22.3
• tqdm 4.64.0
• opencv 4.4.0
• scikit-image 0.17.2
• hypy 3.1.0
• pyyaml 6.0

Our code has been tested with Python 3.8.13, PyTorch 1.11.0, CUDA: 10.0.130 on Ubuntu 18.04.6

Dataset Setup

.
Data
├── train_480                # contain training files
|   ├── recon_001.h5
|       ├── kspace.mat  
|       ├── kspace_sv.mat
|       ├── recon_origin.mat
|       ├── recon_origin_sv.mat
|       ├── mask.mat
|       ├── recon_rss.mat
|       ├── recon_rss_sv.mat
|   └── ...  
|
├── valid                # contain validation files
|   ├── recon_001.h5
|       ├── kspace.mat  
|       ├── kspace_sv.mat
|       ├── recon_origin.mat
|       ├── recon_origin_sv.mat
|       ├── mask.mat
|       ├── recon_rss.mat
|       ├── recon_rss_sv.mat
|   └── ... 
|
└── test                 # contain testing files
|   ├── recon_001.h5
|       ├── kspace.mat  
|       ├── kspace_sv.mat
|       ├── recon_origin.mat
|       ├── recon_origin_sv.mat
|       ├── mask.mat
|       ├── recon_rss.mat
|       ├── recon_rss_sv.mat
|   └── ... 
└── ...  

where
kspace: full-view kspace data with a size of Num x W x H.
kspace_sv: downsampled sparse-view kspace data with a size of Num x W x H.
recon_origin: iverse fourier transfrom of 'kspace' with a size of Num x W x H.
recon_origin_sv: iverse fourier transfrom of 'kspace_sv' with a size of Num x W x H.
mask: downsampling mask of the sparse-view kspace data with a size of Num x W x H.
recon_rss: root sum-of-square (RSS) reconstruction of 'recon_origin' with a size of W x H.
recon_rss_sv: root sum-of-square (RSS) reconstruction of 'recon_origin_sv' with a size of W x H.

To Run the Code

Sample training/testing scripts are provided at the root folder as train.sh and train.sh.

• Train the model

python train.py --experiment_name 'train_IRSENet_1' --data_root '../../Data/Data_ArrangeRecon/' --model_type 'model_cnn' --net_G 'DuRDN' --norm 'BN' --n_filters 16 --growth_rate 16 --n_denselayer 3 --n_channels 32 --lr_I1 5e-4 --lr_K1 5e-4 --lr_I2 5e-4 --lr_K2 5e-4 --n_epochs 600 --batch_size 2 --eval_epochs 5 --snapshot_epochs 5 --num_workers 0 --gpu_ids 0

where
--experiment_name experiment name for the code, and save all the training results in this under this "experiment_name" folder.
--data_root: path of the dataset.
--model_type: model type used (default convolutioanl neural networks -- "model_cnn").
--net_G: neural network model used (default: 'DuRegister_DuSE'). \ --norm: normalization (default: 'BN' for Batch Normalization).
--n_filters: number of filters in the dense connected layers (default: 16).
--growth_rate: number of expanded channels in the dense connected layers (default: 16).
--n_denselayer: number of dense layers.
--n_channels: number of input channels.
--lr_I1: learning rate of the Image-Net in the 1^st iteration.
--lr_K1: learning rate of the Kspace-Net in the 1^st iteration.
--lr_I2: learning rate of the Image-Net in the 2^nd iteration.
--lr_K2: learning rate of the Kspace-Net in the 2^nd iteration.
--n_epochs: num of epoches of training.
--batch_size: training batch size.
--gpu_ids: GPU configuration.

• Test the model

python test.py --resume './outputs/train_IRSENet_1/checkpoints/model_399.pt' --experiment_name 'test_IRSENet_1_399' --data_root '../../Data/Data_ArrangeRecon/' --model_type 'model_cnn' --net_G 'DuRDN' --norm 'BN' --n_filters 16 --growth_rate 16 --n_denselayer 3 --n_channels 32 --eval_epochs 5 --snapshot_epochs 5 --num_workers 0 --gpu_ids 0

where
--resume: path of the model to be tested.
--resume_epoch: training epoch of the model to be tested.
--experiment_name: experiment name for the code, and save all the testing results in this under this "experiment_name" folder.

Data Avaliability

The acquisition of the deidentified fastMRI Dataset was conducted retrospectively using human subject data, made available in open access by Center for Advanced Imaging Innovation and Research (CAI²R), in the Department of Radiology at NYU School of Medicine and NYU Langone Health. Curation of these data are part of an IRB approved study.

Contact

If you have any question, please file an issue or directly contact the authors:

Xiongchao Chen: [email protected]