GAN-BASED RETINAL IMAGES SYNTHESIS AND EYE DISEASE DIAGNOSIS

Abstract

This paper presents a retinal fundus image analysis framework that combines vessel segmentation, GANbased minority-class augmentation, disease classification, and explainable inference for automated ocular screening. The study focuses on four classes: normal, diabetic retinopathy, glaucoma, and age-related macular degeneration. ODIR-5K was used for classification, while a retinal vessel segmentation dataset was used to train PUNet-33 and a lightweight CycleGAN-style generator. The generator learned a vesselmask to fundus-image mapping and was used to synthesize additional glaucoma and AMD images, increasing those classes from 162 and 164 images to 300 each in the training set. The strongest practical classifier in the project was a weighted EfficientNet-B3 ensemble, while the GAN augmentation experiment retrained an Inception-v3 classifier for comparison. Fixed-split testing reached 73.65% accuracy for the weighted ensemble, whereas patient-aware grouped crossvalidation provided a more reliable pooled accuracy of 67.84%. The GAN-augmented retraining run reached 70.61% test accuracy, showing that synthetic balancing improved class availability but did not yet outperform the original baseline. The project also integrates GradCAM, SHAP-style attribution, and a layman-readable explanation to support interpretable deployment.