Enhancing Explainability in Diabetic Retinopathy Detection using Lesion-Based Image Analysis

Diabetic Retinopathy (DR) remains one of the leading causes of vision damage worldwide, emphasizing the urgent need for reliable and interpretable diagnostic tools. While deep learning models have demonstrated remarkable performance in DR detection, their “black-box” nature limits clinical adoption. This study explores an explainable AI (XAI) framework that integrates lesion-level analysis with retinal fundus images to improve both accuracy and interpretability. Lesions such as microaneurysms, hemorrhages, and exudates are detected and highlighted as clinically relevant biomarkers, which are then mapped to disease severity grading. The model not only classifies DR stages but also generates visual explanations that correspond to ophthalmologists’ diagnostic reasoning. Preliminary results suggest that lesion-based explainability enhances clinician trust, facilitates validation of automated outputs, and supports better decision-making in screening programs. This approach underscores the potential of combining AI precision with medical interpretability, paving the way for practical integration of DR screening tools in real-world healthcare settings.
In the primary stage of experiments, the images were categorized into multiple classes representing the severity of diabetic retinopathy (e.g., No DR, Mild, Moderate, Severe, and Proliferative DR). The MobileNet model achieved over 82% classification accuracy, indicating its ability to distinguish between different stages of the disease with reasonable reliability. This performance demonstrates the feasibility of using lesion-focused features for automated DR grading. At the same time , the lesion-based explainability maps generated during these trials showed alignment with clinically relevant structures such as microaneurysms, hemorrhages, and exudates, supporting both the accuracy and interpretability of the predictions.