Early detection and intervention are necessary for diabetic retinopathy, which is a complication for people suffering from diabetes leading to blindness. This project presents a transparent Explainable (XAI) framework for automatically classifying diabetic retinopathy using deep learning. We employed the VGG16, InceptionV3, and Xception models on retinal image datasets with varying data splits (60–40, 70–30, 80–20) to determine the most effective training configuration. The Xception model, which underwent training with L2 regularization on an 80–20 split, attained an accuracy of 86.22%, outperforming models trained with early stopping and other splits that only reached a maximum accuracy of 82.02%. Integrating Grad-CAM enhances transparency and interpretability in the model’s decision-making process. This technique allowed for the visualization of important areas in retinal images that influenced the model’s predictions. The results highlight the possibility of combining advanced deep learning models with techniques for explainability, creating a reliable tool for clinicians to detect and treat diabetic retinopathy early.