[Note that results with PyTorch implementation may vary slightly from the paper.]

How to use the PyTorch code?

• Download the iSEG-2017 data and place it in data folder. (Visit this link to download the data. You need to register for the challenge.)
• Do the preprocessing as mentioned in the tensorflow readme. (optional)
• To run the code for your own dataset, change utils/preprocess.py according to the dataset. Originally it is adapted for iSEG-2017 dataset.

How to run 3D U-Net?

• Configure the unet.json file inside config directory according to your experiment.
• To run training

$ python main.py configs/unet.json

• This will train your model and save the best checkpoint according to your validation performance inside checkpoint_dir mentioned in the json file.
• You can also resume training from saved checkpoint by setting the load_chkpt as True and running the same command as in Step 2.
• To run testing, configure the phase as "testing" in json file and run the command mentioned in Step 2.
• This version of code only compute dice coefficient to evaluate the testing performance.
• Note that the U-Net used here is modified according to the U-Net used in proposed model.(To stabilise the GAN training)

How to run GAN based 3D U-Net?

• Configure the fmgan.json or badgan.json file inside config directory according to your experiment.
• To run training (example: Feature Matching GAN)

$ python main.py configs/fmgan.json

• This will train your model and save the best checkpoint according to your validation performance inside checkpoint_dir mentioned in the json file.
• You can also resume training from saved checkpoint by setting the load_chkpt as True and running the same command as in Step 2.
• To run testing, configure the phase as "testing" in json file and run the command mentioned in Step 2.
• This version of code only compute dice coefficient to evaluate the testing performance.

This PyTorch code is built over the template from this repo, for more details about the structure, kindly refer to it.