This work presents an end–to–end diabetic retinopathy classification framework built on the APTOS 2019 dataset, which contains more than thirteen thousand retinal fundus images. The system combines an EfficientNetB3 backbone with a carefully designed training strategy and multiple lay ers of interpretability. Along with conventional Grad–CAM based visual explanations, the study includes adversarial sensitivity analysis, counterfactual perturbations, and clini cally–motivated robustness tests. These evaluations highlight how the model reacts to changes in pathological regions and whether its predictions remain grounded in medically relevant cues. Experimental results show that the proposed approach achieves strong predictive performance while providing stable and clinically meaningful explanations. The findings indi cate that combining large–scale data, modern convolutional architecture, and structured interpretability testing improves reliability and transparency in automated DR grading. These insights support the use of explainable deep learning tools in real screening pipelines, where trustworthy behaviour is as important as accuracy.

Diabetic Retinopathy,Explainable AI,Adversarial Robustness,EfficientNetB3,Grad-CAM,LIME,FGSM,Counterfactual Explanations