unofficial pytorch implementation of RefineGAN (https://arxiv.org/abs/1709.00753) for CSMRI reconstruction, the official code using tensorpack can be found at https://github.com/tmquan/RefineGAN

• run the original tensorpack code (sorry, can't run tensorpack on my GPU)
• pytorch implementation and experiments on brain images with radial mask
• bug fixed. the mean psnr of zero-filled image is not exactly the same as the value in original paper, although the model improvement is similar
• experiments on different masks

python>=3.7.11 is required with all requirements.txt installed including pytorch>=1.10.0

git clone https://github.com/hellopipu/RefineGAN.git
cd RefineGAN
pip install -r requirements.txt

cd run_sh
sh train.sh

the model will be saved in folder weight, tensorboard information will be saved in folder log. You can change the arguments in script such as --mask_type and --sampling_rate for different experiment settings.

check the training curves while training

tensorboard --logdir log

the training info of my experiments is already in log folder

test after training, or you can download my trained model weights from google drive.

cd run_sh
sh test.sh

cd run_sh
sh visualize.sh

sampling rates : 10%(light orange), 20%(dark blue), 30%(dark orange), 40%(light blue). You can check more loss curves of my experiments using tensorboard.

loss_G_loss_total	loss_recon_img_Aa

PSNR on training set over 500 epochs, compared with results shown in original paper.

my_train_psnr	paper_train_psnr

mean PSNR on validation dataset with radial mask of different sampling rates, batch_size is set as 4;

rate	from left to right: mask, zero-filled, prediction and ground truth	error (zero-filled) and error (prediction)
10%
20%
30%
40%

• data processing before training : complex value represents in 2-channel , each channel rescale to [-1,1]; accordingly the last layer of generator is tanh()
• Generator uses residual learning for reconstruction task
• Generator is a cascade of two U-net, the U-net doesn't do concatenation but addition when combining the enc and dec features.
• each U-net is followed by a Data-consistency (DC) module, although the paper doesn't mention it.
• the last layer of generator is tanh layer on two-channel output, so when we revert output to original pixel scale and calculate abs, the pixel value may exceed 255; we need to do clipping while calculating psnr
• while training, we get two random image samples A, B for each iteration, RefineGAN calculates a large amount of losses (it may be redundant) including reconstruction loss on different phases of generator output in both image domain and frequency domain, total variantion loss and WGAN loss
• one special loss is D_loss_AB, D is trained to only distinguish from real samples and fake samples, so D should not only work for (real A, fake A) or (real B, fake B), but also work for (real A, fake B) input
• WGAN-gp may be used to improve the performance
• small batch size MAY BE better. In my experiment, batch_size=4 is better than batch_size=16

I will appreciate if you can find any implementation mistakes in codes.