danglive / refinegan

RefineGAN

---

This repository holds the code for RefineGAN,

Overview of the proposed method: it aims to reconstruct the images which are satisfied the constraint of under-sampled measurement data; and whether those look similar to the fully aliasing-free results. Additionally, if the fully sampled images taken from the database go through the same process of under-sampling acceleration; we can still receive the reconstruction as expected to the original images.

Two learning processes are trained adversarially to achieve better reconstruction from generator G and to fool the ability of recognizing the real or fake MR image from discriminator D

The cyclic data consistency loss, which is a combination of under-sampled frequency loss and the fully reconstructed image loss.

Generator G, built by basic building blocks, can reconstruct inverse amplitude of the residual component causes by reconstruction from under-sampled k-space data. The final result is obtained by adding the zero-filling reconstruction to the output of G

---

It is developed for research purposes only and not for commercialization. If you use it, please refer to our work.

@ARTICLE{8327637, 
author={T. M. Quan and T. Nguyen-Duc and W. Jeong}, 
journal={IEEE Transactions on Medical Imaging}, 
title={Compressed Sensing MRI Reconstruction Using a Generative Adversarial Network With a Cyclic Loss}, 
year={2018}, 
volume={37}, 
number={6}, 
pages={1488-1497}, 
doi={10.1109/TMI.2018.2820120}, 
ISSN={0278-0062}, 
month={June},
}

---

Directory structure of data:

 tree data
 data/
├── brain
│   ├── db_train
│   └── db_valid
├── knees
│   ├── db_train
│   └── db_valid
└── mask
    ├── cartes
    │   ├── mask_1
    │   ├── mask_2
    │   ├── ...
    │   └── mask_9
    ├── gauss
    │   ├── mask_1
    │   ├── mask_2
    │   ├── ...
    │   └── mask_9
    ├── radial
    │   ├── mask_1
    │   ├── mask_2
    │   ├── ...
    │   └── mask_9
    └── spiral
        ├── mask_1
        ├── mask_2
        ├── ...
        └── mask_9

Brain data is used for magnitude-value experiment, it is extracted from http://brain-development.org/ixi-dataset/

Knees data is used for complex-value experiment, it is extracted from http://mridata.org

---

Prerequisites

sudo pip install tensorflow_gpu==1.4.0
sudo pip install tensorpack==0.8.2
sudo pip install scikit-image==0.13.0
sudo pip install whatever-missing_libraries

---

To begin, the template for such an experiment is provided in exp_dset_RefineGAN_mask_strategy_rate.py

For example, if you want to run the training and testing for case knees data, mask radial 10%, please make a soft link to the experiment name, like this

ln -s exp_dset_RefineGAN_mask_strategy_rate.py 	 \
	  exp_knees_RefineGAN_mask_radial_1.py

---

To train the model

python exp_knees_RefineGAN_mask_radial_1.py  	 \
	    --gpu='0'				 \
	    --imageDir='data/knees/db_train/'    \
	    --labelDir='data/knees/db_train/'    \
	    --maskDir='data/mask/radial/mask_1/' 

Checkpoint of training will be save to directory train_log

---

To test the model

mkdir result 


python exp_knees_RefineGAN_mask_radial_1.py  	 \
	    --gpu='0' 				 \
	    --imageDir='data/knees/db_valid/' 	 \
	    --labelDir='data/knees/db_valid/' 	 \
	    --maskDir='data/mask/radial/mask_1/' \
	    --sample='result/exp_knees_RefineGAN_mask_radial_1/' \
	    --load='train_log/exp_knees_RefineGAN_mask_radial_1/max-validation_PSNR_boost_A.data-00000-of-00001'   

---

The authors would like to thank Dr. Yoonho Nam for the helpful discussion and MRI data, and Yuxin Wu for the help on Tensorpack.