An explainable ensemble for diabetic retinopathy grading with a novel confidence quality factor and configurable heatmaps

Abstract

While current artificial intelligence (AI) tools aid in detecting diabetic retinopathy (DR), they face significant challenges that limit their clinical utility. Most are restricted to binary (referable vs. non-referable) screening and operate as “black boxes,” lacking the detailed, transparent explanations required for diagnostic confidence. This study addresses these gaps by introducing a novel, explainable ensemble-based approach for detailed DR grading. Our system utilizes a parallel ensemble of two efficient deep learning networks, EfficientNetV2 and ConvNeXt, to perform a full five-class international clinical diabetic retinopathy (ICDR) classification. The proposed model achieves state-of-the-art performance, with 96.7% accuracy and an Area Under the Curve (AUC) over 96% for all classes on a public dataset. More importantly, it provides a comprehensive diagnostic report designed to enhance clinical trust and utility. This report features multiple, configurable superimposed heatmaps, two probability-ordered diagnostic suggestions, and a novel quality factor that estimates the confidence of the prediction. By offering richer, more transparent, and interactive explanations, our system moves beyond simple screening to function as a valuable diagnostic assistance tool for ophthalmologists and other healthcare professionals.