Retinopathic conditions like glaucoma, diabetic retinopathy, and age-related macular degeneration are some of the leading causes of visual impairment and blindness. Automated and timely methods of detection can enhance the patient’s quality of care tremendously. This work evaluates the effectiveness of four modern CNN architectures, ResNet101, DenseNet121, DenseNet169, and Xception, in binary classification of fundus images using the ODIR-5K dataset. All models were subjected to transfer learning and were assessed using well-defined benchmarks. Out of all models tested, the Xception model performed the binary classification task the best with an accuracy, precision, recall, F1, and Jaccard score of 96%, and ROC-AUC of 97%, which suggests strong diagnostic capability. DenseNet121 also performed well and maintained an ROC-AUC score of 98%, highlighting its effective feature learning. In order to tackle the ‘black-box’ problem associated with deep learning models, the current research implements Grad-CAM, Grad-CAM plus plus, Score-CAM, Faster Score-CAM, and Layer-CAM, to provide visual evidence interpreting the decisions made by the models. The generated heatmaps consistently highlighted clinically relevant retinal regions that contributed to model predictions, thereby enhancing interpretability and supporting clinical validation. The use of Explainable AI methods offered clear visual justifications for the models’ outputs, fostering greater trust in the decision-making process. Integrating high-performing CNNs with XAI techniques not only facilitates accurate automated diagnosis but also improves model transparency and practical applicability in ophthalmic settings. The results underscore the potential deep learning interpretable frameworks have for increasing accuracy in diagnosis.