CGAN

Patch-based CGAN on OCT images

The CGAN model was based on Kugelman et al. paper "Data augmentation for patch-based oct chorio-retinal segmentation using generative adversarial networks". Our code was based on their implementation that you can find on their github repository.

Dependencies

• Python
• tensorflow
• matplotlib

The file tensorflow-env.yml is a working environment for training the CGAN.

Structure

• cgan_NxN_patchbased.py trains a conditional CGAN on NxN patches.
• evalfid_NxN_patchbased.py evaluate trained generators of a GAN, using the Frechet Inception Distance.
• cgan_NxN_patchbased_genfid.py constructs synthetic patches and save the images as required.
• patches.ipynb is the notebook generating original patches of a given shape.
• script_evalfid.sh and script.sh are shells to launch the jobs on the clusters.
• fid.py, Minibatch.py and training_patchbased.py are helpers for the CGAN training and the synthetic patches generation. They were given by Kugelman's et al. implementations.

Example: instructions for 70x70 patches

Follow the following steps to train and generate synthetic patches:

1. Train a conditional GAN on 70x70 patches using cgan_70x70_patchbased.py. Load data using the load_data function.
2. Evaluate trained generators of a GAN, using the Frechet Inception Distance (FID), using evalfid_70x70_patchbased.py and by specifying the folder containing the generators as load_path
3. Construct synthetic patches using cgan_70x70_patchbased_genfid.py. Save the generated images as required.