Quantifying Confidence in Diabetic Retinopathy Diagnosis: A Comparative XAI Study of Deep Learning and Bayesian Neural Networks

Abstract

Diabetic Retinopathy remains the primary microvascular complication of diabetes and a leading cause of irreversible blindness globally. While deep learning models offer high diagnostic accuracy, their widespread clinical integration is profoundly limited by two fundamental, unresolved deficiencies in previous literature: the absence of comprehensive, fair comparative analysis across diverse architectures and the pervasive lack of transparent, quantifiable prediction confidence necessary for clinical acceptance. This study directly addresses these challenges by presenting a highly optimized and rigorous comparative evaluation of three powerful models: the high-capacity EfficientNetB0, the computationally efficient MobileNetV3Small, and a novel Custom Bayesian Neural Network (BNN) framework. Through robust methodology, all models achieved exceptional generalization, stabilizing with impressive final F1-Score > 0.91. The Custom BNN demonstrated clear superiority as the most reliable diagnostic tool, securing the highest Accuracy 0.9294 and F1-score 0.9289 on the objective test set. Most significantly, this work delivers a breakthrough in safety assurance by integrating sophisticated Explainable AI (XAI) and probabilistic modeling: Grad-CAM and Local Interpretable Model-agnostic Explanations (LIME) confirmed anatomically grounded decision-making, while the BNN uniquely provides quantifiable uncertainty metrics, offering a crucial 95% confidence interval (CI) for every diagnosis. These results validate a new generation of high-performance models, led by a transparent BNN architecture, that are ready for implementation to deliver reliable, trusted, and efficient Diabetic Retinopathy screening solutions worldwide.