Dual-Domain Coarse-to-Fine Progressive Network for Simultaneous Denoising, Limited-View Reconstruction, and Attenuation Correction of Cardiac SPECT (IEEE Transcations on Medical Imaging 2024)

Xiongchao Chen, Bo Zhou, Xueqi Guo, Huidong Xie, Qiong Liu, James S. Duncan, Albert J. Sinusas, Chi Liu

Overview

Projection-domain framework

Image-domain framework

This repository contains the PyTorch implementation of the Dual-Domain Coarse-to-Fine Progressive Network (DuDoCFNet).

Citation

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

xxx (Under review ...)

Environment and Dependencies

Requirements:

• Python 3.6.10
• Pytorch 1.2.0
• numpy 1.19.2
• scipy
• scikit-image
• h5py
• tqdm

Our code has been tested with Python 3.6.10, Pytorch 1.2.0, CUDA: 10.0.130 on Ubuntu 18.04.6.

Dataset Setup

Data
├── train                # contains training files
|   ├── data1.h5
|       ├── Amap.mat  
|       ├── Recon_LD_LA_EM.mat
|       ├── Proj_FD_FA_EM.mat  
|       ├── Proj_LD_FA_EM.mat
|       ├── Proj_LD_LA_EM.mat
|   └── ...  
| 
├── valid               # contains validation files
|   ├── data1.h5
|       ├── Amap.mat  
|       ├── Recon_LD_LA_EM.mat
|       ├── Proj_FD_FA_EM.mat  
|       ├── Proj_LD_FA_EM.mat
|       ├── Proj_LD_LA_EM.mat
|   └── ...  
|
├── test                # contains testing files
|   ├── data1.h5
|       ├── Amap.mat  
|       ├── Recon_LD_LA_EM.mat
|       ├── Proj_FD_FA_EM.mat  
|       ├── Proj_LD_FA_EM.mat
|       ├── Proj_LD_LA_EM.mat
|   └── ...  
└── ...  

where
Amap: CT-derived attenuation maps with a size of 72 x 72 x 40.
Recon_LD_LA_EM: reconstructed lose-dose and limited-angle images with a size of 72 x 72 x 40.
Proj_FD_FA_EM: full-dose and full-angle projections with a size of 32 x 32 x 20.
Proj_LD_FA_EM: low-dose and full-angle projections with a size of 32 x 32 x 20.
Proj_LD_LA_EM: lose-dose and limited-angle projections with a size of 32 x 32 x 20.
Mask_Proj: binary projection mask with a size of 32 x 32 x 20. 1 refers to the central limited-angle regions. \

To Run the Code

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

• Train the model

python train.py --experiment_name 'train_1' --model_type 'model_cnn' --data_root './xxx' --norm 'BN' --net_filter 32 --n_denselayer 4 --growth_rate 16 --lr_G1 1e-3 --lr_G2 1e-4 --step_size 1 --gamma 0.99 --n_epochs 150 --batch_size 2 --eval_epochs 5 --snapshot_epochs 5 --gpu_ids 0

where
--experiment_name experiment name for the code, and save all the training results in this under this "experiment_name" folder.
--model_type: model type used (default convolutional neural networks).
--data_root: the path of the dataset.
--norm: batch normalization in the CNN modules (default: 'BN').
--net_filter: num of filters in the densely connected layers (default: 32).
--n_denselayer: num of the densely connected layers (default: 4).
--growth_rate: growth rate of the densely connected layers (default: 16).
--lr_G1: learning rate of projection-domain modules (default: 1e-3).
--lr_G2: learning rate of image-domain modules (default: 1e-3). --step_size: num of epoch for learning rate decay .
--gamma: learning decay rate.
--n_epochs: num of epochs of training.
--batch_size: training batch size.
--test_epochs: number of epochs for periodic validation.
--save_epochs: number of epochs for saving trained model.
--gpu_ids: GPU configuration.

• Test the model

python test.py --resume './outputs/train_1/checkpoints/model_xx.pt' --experiment_name 'test_1_xx' --model_type 'model_cnn' --data_root '../xxx' --norm 'BN' --net_filter 32 --n_denselayer 4 --growth_rate 16 --batch_size 2 --gpu_ids 0

where
--resume: the path of the model to be tested.
--resume_epoch: training epoch of the model to be tested.

Data Availability

The original dataset in this study is available from the corresponding author (chi.liu@yale.edu) upon reasonable request and approval of Yale University.

Contact

If you have any questions, please file an issue or directly contact the author via xiongchao.chen@yale.edu or cxiongchao9587@gmail.com.