MDPET: A Unified Motion Correction and Denoising Adversarial Network for Low-dose Gated PET

Bo Zhou, Yu-Jung Tsai, Xiongchao Chen, James S. Duncan and Chi Liu

IEEE Transactions on Medical Imaging (TMI), 2021

[Paper]

This repository contains the PyTorch implementation of MDPET.

Citation

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

@article{zhou2021mdpet,
  title={MDPET: A Unified Motion Correction and Denoising Adversarial Network for Low-dose Gated PET},
  author={Zhou, Bo and Tsai, Yu-Jung and Chen, Xiongchao and Duncan, James S and Liu, Chi},
  journal={IEEE Transactions on Medical Imaging},
  year={2021},
  publisher={IEEE}
}

Environment and Dependencies

Requirements:

• Python 3.7
• Pytorch 1.4.0
• scipy
• scikit-image
• itertools
• tqdm
• csv
• json

Our code has been tested with Python 3.7, Pytorch 1.4.0, CUDA 10.0 on Ubuntu 18.04.

Dataset Setup

./preprocess/              # data setup 
│
├── train                  # training data -- each .h5 contains 6 x high-dose gated data and 6 x low-dose gated data
│   ├── case_1.h5     
│   ├── case_2.h5       
│   ├── case_3.h5 
│   ├── ...    
│   └── case_N.h5
│
├── test                   # test data -- each .h5 contains 6 x high-dose gated data and 6 x low-dose gated data
│   ├── case_N+1.h5     
│   ├── case_N+2.h5       
│   ├── case_N+3.h5 
│   ├── ...    
│   └── case_M.h5         
└── 

Each .h5 file should contain 6 gated image at high-dose level and 6 gated image at low-dose level.
The variable name in the .h5 should be 'G1LD' / 'G2LD' / 'G3LD' / 'G4LD' / 'G5LD' / 'G6LD' for low-dose gated images, and 'G1HD' / 'G2HD' / 'G3HD' / 'G4HD' / 'G5HD' / 'G6HD' for high-dose gated images.

To Run Our Code

• Train the model

python train.py \
--experiment_name 'experiment_train_svrhddn_dp_gan_mse' \
--model_type 'model_svrhd_dp_gan' \
--data_root './preprocess/' \
--net_G1 'svr_dp' \
--net_G2 'unet' \
--net_D 'patchGAN' \
--weight_dn_recon 100 \
--weight_reg 1000 \
--image_loss 'mse' \
--batch_size 1 \
--zoomsize_train 106 106 106 \
--cropsize_train 96 96 96 \
--rotate_train 30 \
--AUG \
--eval_epochs 20 \
--save_epochs 20 \
--snapshot_epochs 20 \
--lr 1e-4

where
--experiment_name provides the experiment name for the current run, and save all the corresponding results under the experiment_name's folder.
--data_root provides the data folder directory (with structure and files illustrated above).
--net_G1 specifies the motion estimation network type.
--net_G2 specifies the denoising network type.
--net_D specifies the discriminator network type.
--AUG enables the data augumentation during the training.
Other hyperparameters can be adjusted in the code as well.

• Test the model

python test.py \
--resume './outputs/experiment_train_svrhddn_dp_gan_mse/checkpoints/model_xxx.pt' \
--experiment_name 'experiment_test_svrhddn_dp_gan_mse' \
--model_type 'model_svrhd_dp_gan' \
--data_root './preprocess/' \
--net_G1 'svr_dp' \
--net_G2 'unet' \
--net_D 'patchGAN'

Sample training/test scripts are provided under './scripts/' and can be directly executed.

Test Data

We provide one test data which includes the input to MDPET and output of the MDPET.
Please see the following link to request the test data.
[Drive Link]

Contact

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

Bo Zhou: [email protected]