Diabetic Retinopathy (DR) is a significant threat to eyesight and blindness globally, particularly for those with a more extended diabetes history. Deep learning has achieved high accuracy for DR detection using fundus images; however, the "black-box" nature hinders its application in clinical practice, where interpretability is important. In this work, we propose a solution to the model transparency problem by introducing an XAI-enhanced diagnostic framework utilizing CNNs. We present an explainable deep learning framework based on a convolutional neural network (CNN), specifically ResNet-50, which has been fine-tuned on the APTOS 2019 Blindness Detection dataset. To narrow the interpretability gap, we utilize the Grad-CAM and SHAP visualization methods, which generate class-discriminative heatmaps and feature-attribution plots, respectively. The multi-class diabetes retinopathy (DR) classification result yielded an overall accuracy of 83% for the model. Importantly, the explanation agreement score with ophthalmologists is over 78%, indicating a high correlation between the AI-based saliency maps and expert-annotated lesion regions. Our findings show that XAI can not only maintain diagnostic accuracy but also enhance model interpretability, rendering AI-based DR screening systems more acceptable and usable in clinical practice. This study reinforces the importance of explainability as an integral part of implementing medical AI.
Turki Alghamdi (Mon,) studied this question.