paper can be found in here
-
base : Denoising without reversal loss. Denoising model can be [dncnn, unet, edsr].
python main.py --way base --model [dncnn, unet, edsr] --source ge --vgg_weight 0.1 --l_weight 1 -
rev : Denoising with reversal loss. Gradient reversal of target domain is included. Denoising model can be [dncnn, unet, edsr]. you can choose domain classifer input(dc_input) and reversal stage(style_stage).
python main.py --way rev --model [dncnn, unet, edsr] --source ge --target mayo --test_every 500 --vgg_weight 0.01 --l_weight 1 --rev_weight 0.1 --dc_mode [mse, bce, wss] --dc_input [img, noise, feature, c_img, c_noise, c_feature] --style_stage [1,2,3,4,5,6] (--content_randomization) -
wgan : Denoising with wasserstein loss.
python main.py --way wgan --source ge --target mayo --vgg_weight 0.1 --l_weight 1 -
wganrev : Denoising with wasserstein loss and reversal loss. Gradient reversal of target domain is included. Denoising model is wganvgg. you can choose domain classifer input(dc_input) and reversal stage(style_stage).
python main.py --way wganrev --source ge --target mayo --test_every 500 --vgg_weight 0.1 --l_weight 1 --rev_weight 0.001 --dc_mode [mse, bce, wss] --dc_input [img, noise, feature, c_img, c_noise, c_feature] --style_stage [1,2,3,4,5,6] (--content_randomization) -
out2src : Denoising with fake_target low dataset. You have to specify the fake_dir_name of source dataset (only base name of dir, not the full path).
python main.py --way wganrev --source ge --target mayo --domain_sync out2src --fake_dir fake_dir -
ref2trg : Denoising with fake_target low & high dataset. You have to specify the fake_dir name (only base name of dir, not the full path).
python main.py --way wganrev --source ge --target mayo --domain_sync ref2trg --fake_dir fake_dir
python main.py --mode test --target mayo --thickness 3
- OS: The package development version is training on Linux and tested on Windows operating systems with Anaconda.
- Python : 3.7.1
- Pytorch : 1.4.0
- Source data : Phantom dataset (vendor : GE)
- Target data : Mayo dataset (vendor : SIEMENS)