An Explainable Deep Learning Approach for Diabetic Retinopathy Classification and Precise Lesion Segmentation

Abstract

Diabetic Retinopathy (DR) is a leading cause of vision impairment and blindness among diabetic patients worldwide, necessitating early and accurate diagnosis for effective treatment. This study proposes an explainable deep learning approach for diabetic retinopathy classification and precise lesion segmentation using a hybrid architecture that integrates Convolutional Neural Networks (CNNs) and Transformer-based attention mechanisms. The proposed framework is designed to automatically identify the severity level of DR from retinal fundus images and accurately localize pathological regions such as microaneurysms, hemorrhages, and exudates through advanced segmentation techniques.The CNN component extracts hierarchical spatial features from input retinal images, while the Transformer module captures long-range contextual dependencies to enhance global feature representation. To improve clinical trust and interpretability, Explainable Artificial Intelligence (XAI) techniques such as Grad-CAM and attention map visualization are incorporated, enabling clinicians to understand model decision-making and validate highlighted lesion regions.The segmentation module employs a hybrid encoder–decoder architecture with multi-scale feature fusion to precisely delineate infected areas, improving diagnostic transparency and reliability. Experimental evaluation on benchmark retinal image datasets demonstrates superior performance in terms of classification accuracy, sensitivity, specificity, Dice coefficient, and Intersection-over-Union (IoU) compared to conventional deep learning models.The proposed explainable framework not only enhances diagnostic accuracy but also strengthens clinical interpretability, making it a reliable decision-support tool for automated diabetic retinopathy screening and early intervention in real-world healthcare settings