Controllable retinal image synthesis using conditional StyleGAN and latent space manipulation for improved diagnosis and grading of diabetic retinopathy

This research leverages ConditionalStyleGAN to generate synthetic fundus images, which are then used to enhance the performance of a retinopathy grading classifier across various scenarios. The pretrained ConditionalStyleGAN model is located in the "CStyleGAN" folder.

Our findings indicate that the best results for the grading classifier are achieved when it is initially pretrained with synthetic images and subsequently fine-tuned with real images. Detailed results, including all images used in training, the trained model, and the code to reproduce the results, are available in the "4.3.2. DR Severity Grading" folder.

Another aspect of this research involves using a SeFa-based algorithm to manipulate images by removing lesions. These manipulated images are then utilized to train a diabetic retinopathy detection classifier. The model used by the SeFa algorithm is stored in the "SeFa-based Manipulation" folder. This model is a `.pth` version of the `.pkl` model produced by ConditionalStyleGAN.

The best results for the diabetic retinopathy detection classifier are documented in the "4.3.3. DR Detection" folder. This folder contains the trained model, images, and the code necessary to reproduce the results presented in this research.

All code is implemented using the Fastai library, and the folder organization is structured according to the code requirements.