RetinalXAI-FundusSys: An XAI-Based Deep Learning System for Timely Diabetic Retinopathy Identification

Diabetic Retinopathy (DR), a serious microvascular consequence of Diabetes Mellitus (DM), remains one of the primary causes of avoidable vision impairment globally. Early and timely DR identification is essential, yet traditional clinical screening relies on manual grading of Retinal Fundus Images (RFIs), making the process subjective, labor-intensive, & prone to inter-observer variability. This paper introduces RetinalXAI-FundusSys, an XAI-based DL technique for timely DR identification. The proposed system leverages an optimized Fundus Image Optimization Workflow (FIOW) with CLAHE, Gaussian filtering, and artifact suppression, followed by robust feature extraction using EfficientNet-B5 CNN. To overcome the "black-box" nature of DL, the framework integrates Explainable AI (XAI) modules—Grad-CAM, Grad-CAM++, and SHAP—for clinically aligned lesion-level interpretability. Quantitative evaluation across multiple benchmark datasets demonstrates superior performance in terms of ACC (Accuracy), PREC (Precision), REC (Recall), F1-Score, and SPEC (Specificity), outperforming several DR models. The proposed system provides a acceptable, transparent, and scalable solution for tele-ophthalmology, automated screening, and large-scale DR surveillance programs.