XReGViT: Explainable GAN-ViT Model for Retinal Lesion Detection and Diabetic Retinopathy Severity Classification

Diabetic retinopathy (DR) is one of the causes of preventable blindness and early diagnosis by retinal fundus analysis is essential to the effective treatment of diabetic retinopathy. Conventional convolutional neural network (CNN) methods are usually subject to class imbalance, poor lesion readability, and poor generalization. This paper introduces XReGViT, a framework of Explainable GAN-ViT that integrates generative augmentation, transformer-based classification, and explainable lesion localization as an automated retinal analysis framework. The model uses a Generative Adversarial Network (GAN) to generate highfidelity fundus images to reduce the problem of data imbalances and increase the diversity of lesions. As the main classifier, a Vision Transformer (ViT) is used that uses self-attention to learn global contextual features across five levels of DR, noDR, Mild, Moderate, Severe, and Proliferative DR on the APTOS 2019 dataset. Grad-CAM and attention-rollout visualization explainability is added, which allows localization of lesions microaneurysms, hemorrhages, and exudates of lesions at the clinical level. Through experimental analysis, it is shown that XReGViT attains a 97.18% accuracy, 96.96% precision, 97.18% recall, and $F 1$-score of $97.06 \%$, which is better than the benchmark models.