Enhanced Diabetic Retinopathy Assessment Using Multiple Instance Learning and Explainable Vision Transformers Integrated with Generative Adversarial Networks

Diabetic Retinopathy (DR) is a leading cause of preventable blindness, necessitating early detection for effective intervention. Conventional DR screening is labor-intensive and subject to inter-observer variability, making automated solutions essential. This study proposes an enhanced DR assessment framework integrating Multiple Instance Learning (MIL) and Vision Transformers (ViT) to improve detection accuracy and explainability. The MIL-ViT model leverages self-attention mechanisms to capture long-range dependencies in retinal fundus images, enhancing classification performance. Additionally, YOLOv8-based segmentation identifies microaneurysms, hemorrhages, and exudates, while EigenCAM explainability visualizations provide interpretable heat maps. The system was evaluated using multiple deep learning models, with EfficientNetB0 achieving 93.09% accuracy and an AUC-ROC of 0.96, outperforming traditional CNN architectures. The framework is deployed via a Streamlit-based interactive interface, enabling real-time DR grading and segmentation for clinical and teleophthalmology applications. This research advances automated, interpretable, and scalable DR screening solutions, which are particularly beneficial for resourceconstrained settings.