A Clinically-Aligned Multi-Family Explainable AI Framework for Diabetic Retinopathy Detection on Fundus Images

Automated diabetic retinopathy (DR) screening has achieved expert-level accuracy, yet clinical adoption remains limited by the opacity of deep neural networks. We address this gap with a DenseNet121-based binary classifier trained on 3,662 retinal fundus images from APTOS 2019, optimised through a two-phase transfer-learning and fine-tuning strategy with focal loss and class-balanced sampling. The model achieves 95.45% test accuracy, an AUC-ROC of 0.9881, sensitivity of 93.91%, and specificity of 95.94%. To make these predictions interpretable, we integrate and systematically benchmark six complementary explainable AI (XAI) techniques drawn from three theoretical families: perturbation (Occlusion Sensitivity, LIME, RISE), gradient (Integrated Gradients), and activation-based (Grad-CAM++, Score-CAM). Each method is evaluated on processing time, memory footprint, and agreement with expert-annotated anatomical structures. The six methods converge on clinically meaningful regions, including the optic disc (85% average agreement), major vessels (78%), and macula (66%), indicating that the network's decisions are grounded in established DR pathology rather than spurious correlations. Statistical tests (DeLong, McNemar, bootstrap with 10,000 resamples) confirm significant gains over standard CNN baselines. The framework demonstrates that strong screening performance and clinical interpretability can be jointly achieved, providing a deployment-ready template for DR decision support.