EyeNet: An Explainable AI-Based Retinal Disease Classification by Using Hybrid Deep Learning Approach

Timely and accurate diagnosis of eye diseases is crucial for preventing irreversible vision impairment, particularly in resource-constrained healthcare environments. With recent advancements in deep learning, automated retinal image analysis has demonstrated strong potential for supporting largescale screening. In this study, two hybrid deep learning models Xception+Vision Transformer (ViT) and VGG19+ViT are proposed and evaluated for four-class retinal disease classification, including cataract, diabetic retinopathy, glaucoma, and normal cases. A publicly available dataset of 4,217 retinal fundus images was used, and transfer learning was applied to fine-tune each hybrid model. Experimental results show that the Xception + ViT model achieved the strongest performance, with an overall accuracy of 96%, precision of 0.99, recall of 0.98, and an $F 1$-score of 0.98, including near-perfect detection of diabetic retinopathy (recall $=1.00$). The VGG19 + ViT model reached 90% accuracy but exhibited lower precision when distinguishing glaucoma from normal cases. An ablation study confirmed the complementary roles of the CNN backbone for local feature extraction and the transformer encoder for global contextual reasoning. To enhance clinical interpretability, Grad-CAM explainability was employed, demonstrating that the hybrid models consistently focus on diagnostically relevant retinal regions such as the optic disc, macula, and vascular structures. Overall, the findings indicate that CNN-ViT hybrid architectures provide a robust and interpretable framework for multi-class retinal disease classification, outperforming conventional CNN baselines and offering strong potential for real-world ophthalmic decision-support systems.