Cycle-consistent generative adversarial network (Cycle-GAN) is an unsupervised approach for image synthesis. In medical imaging, it promises to provide a tool for intricate data augmentation. Although some clinical studies report that GANs may generate unrealistic features in some cases, they are still promising tools for image generation with high visual fidelity. This repository provides the code for 3D patch-based synthesis of medical images using Cycle-GAN.

Create a virtual environment to install the prerequisites. If there are issues with tensorflow-gpu, cuda and cuDNN version mismatch, use Anaconda or conda-forge to install the requirements. The working versions of the Nvidia cuda driver, tensorflow-gpu, cudatoolkit and cudnn can be found here.

conda create -n tfgpu
conda activate tfgpu
conda install tensorflow-gpu -c conda-forge

The code benefits from Monai which is a Torch-based medical imaging library for custom preprocessing and caching of medical images. The preprocessing operations are performed in the dataloader.py file with specifications defned in the config.py file. The training specifications can be defined via the arguments defiend in the main.py file.

The expected directories are as follows:

|--directory
       |----train
              |----A
                   |----xxx.nii.gz, xxx.nii.gz, ...
              |----B
                   |----xxx.nii.gz, xxx.nii.gz, ...
       |----test
              |----A
                   |----xxx.nii.gz, xxx.nii.gz, ...
              |----B
                   |----xxx.nii.gz, xxx.nii.gz, ...

For using other image formats, make the desired changes in the dataloader.py file.

Train:

python main.py path/to/data/directory path/to/save/results

The inference code will be released soon.