A Hybrid Deep Learning Model Fusing DenseNet-169 and Vision Transformer for Diabetic Retinopathy Classification

Diabetic Retinopathy (DR) is a major eye disease in diabetic patients that can cause blindness if not treated in time. It affects almost one-third of people with diabetes and is one of the top preventable causes of vision loss worldwide. DR progresses step by step - from Mild, Moderate, and Severe Non-Proliferative Diabetic Retinopathy (NPDR) to Proliferative Diabetic Retinopathy (PDR). If not identified early, the damage can be irreversible. Timely intervention requires early detection and proper grading. Convolutional Neural Networks (CNNs) work well on fundus images because they can capture local lesion-level features. However, CNNs mainly focus on nearby spatial patterns and often miss long-range connections. Vision Transformers (VITs), on the other hand, use Self-Attention (SA) to capture global relationships across the whole image.In this paper, we propose a Hybrid Model that combines CNNs and VITs for DR classification. CNNs are used to extract local lesion details, while VITs learn long-range dependencies for better severity prediction. Our approach brings out the strengths of both models and achieves high accuracy of 95.26% on the Messidor-2 dataset.These results prove that a CNN-VIT Hybrid system is reliable, scalable, and clinically helpful for early DR screening and diagnosis