RobustDRNet: A clinically-aligned hybrid ensemble model with multi-method explainability for lesion-aware diabetic retinopathy grading

Diabetic retinopathy (DR) screening requires artificial intelligence (AI) models that are not only highly accurate in grading five clinical stages but are also capable of generating quantitatively evaluated lesion-aware explanations to earn the trust of clinicians. We propose RobustDRNet , a hybrid ensemble model that combines local convolutional features from Residual Network-34 (ResNet-34) and ConvNeXt-Tiny with global transformer embeddings from Vision Transformer Base/16 (ViT-B16) via two-stage feature fusion and a disentangled multilayer perceptron (MLP), followed by a logistic regression stacking meta-learner for prediction aggregation. To address severe class imbalance, our training pipeline employs stratified sampling, contrast-limited adaptive histogram equalization (CLAHE) for contrast enhancement, strong data augmentation, and class-weighted focal loss. Evaluated on the Asia Pacific Tele-Ophthalmology Society (APTOS) 2019 dataset, RobustDRNet achieved 88.4% validation accuracy, a 0.967 macro-averaged area under the receiver operating characteristic curve (macro-AUC), and Cohen’s kappa of 0.823, outperforming individual backbones and simple voting ensembles. In addition to classification performance, we integrated six complementary explainable AI (XAI) techniques: Gradient-weighted Class Activation Mapping++ (Grad-CAM++), Integrated Gradients, attention rollout, SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), and Testing with Concept Activation Vectors (TCAV). Each technique was quantitatively benchmarked against expert-annotated lesion maps from the Indian Diabetic Retinopathy Image Dataset (IDRiD). Saliency maps achieved mean Intersection over Union (IoU) scores of 0.06 for Grad-CAM++ and approximately 0.10 for Integrated Gradients; SHapley Additive exPlanations (SHAP) perturbations showed a deletion drop of 0.25 and an insertion gain of 0.22; and TCAV achieved complete classifier-level TCAV alignment (score = 1.0) with clinically coherent, grade-wise importance trajectories. By combining competitive grading performance with multi-perspective and quantitatively evaluated interpretability, RobustDRNet provides a promising DR screening framework whose decisions are supported by lesion-aware explanatory evidence.