A Hybrid Deep Learning Approach With Explainable AI for Diabetic Retinopathy Classification

Diabetic Retinopathy is one of the major cause of vision impairment around the world, yet its impact can be significantly reduced through early detection and timely intervention. Previous screening techniques were dependent on manual detection, which consumed a lot of time and often provided inconsistent results. However, the successful developments in the area of Deep Learning have facilitated the automated and large-scale analysis of the retinal image with better performance. Irrespective of such developments, the use of explainable AI in many studies has received limited attention. This work, provides an extensive analysis of several deep learning models like ResNet18, EfficientNetB0, DenseNet121 and MobileNetV2 for multiclass diabetic retinopathy classification, using a unified training framework with SMOTE-based class balancing to address severe data imbalance. In addition to individual models, both hard and soft voting ensemble strategies, as well as a feature-level hybrid architecture combining EfficientNetB0 and DenseNet121, were systematically evaluated. The proposed hybrid model achieved an accuracy of 84.02%, while ensemble learning attained the highest overall accuracy of 85.29%, demonstrating improved robustness and balanced performance across severity levels. Model performance was comprehensively evaluated using accuracy, precision, recall, F1-score, and class-wise ROC-AUC metrics, with per-class AUC values exceeding 0.90 for multiple severity stages. Particularly, model interpretability is introduced with the help of Gradient-weighted Class Activation Mapping(Grad-CAM), where visualization in the form of heatmaps of clinically relevant retinal regions is offered. To move beyond purely qualitative analysis, a quantitative Grad-CAM concentration score was employed, demonstrating more spatially localized and diagnostically focused attention in individual and hybrid model representations. Hence, the results indicate that incorporating complementary convolutional neural network architectures, together with ensemble learning and quantitative explainability, yields a more robust, interpretable, and clinically relevant diabetic retinopathy screening framework.