Application of conditional lightweight GAN for retinal fundus image synthesis based on diabetic retinopathy severity levels on the IDRiD dataset

Abstract

Diabetic Retinopathy (DR) is a leading cause of preventable blindness, yet the development of automated diagnostic models using Deep Learning is often hindered by the availability of imbalanced medical datasets. This study aims to address this issue by implementing a Conditional Lightweight Generative Adversarial Network (c-LGAN) architecture to synthesize realistic fundus retinal images corresponding to five DR severity levels from the IDRiD dataset. The c-LGAN model was trained on a balanced dataset, and its performance was quantitatively evaluated using Frechet Inception Distance (FID) and Inception Score (IS) metrics. The results demonstrate that the proposed model is capable of generating high-quality images, evidenced by achieving a best FID score of 121.24 at epoch 100. However, further observation identified significant stability challenges in long-term training, marked by a performance collapse after the model reached its optimal point. This phenomenon was attributed to an overpowering discriminator. This study concludes that c-LGAN is a promising approach for data augmentation but emphasizes the critical importance of periodic metric monitoring and model checkpointing strategies to capture peak performance and overcome training stability issues.