<![CDATA[Visual impairment affects over 2.2 billion people globally, with 1 billion cases preventable through early detection. This research aims to address early detection challenges by explicitly developing a classification model for eye diseases using fundus images. We employed the Vision Transformer (ViT-B/16) with fine-tuning transfer learning, trained on a balanced dataset enhanced by light data augmentation techniques such as random rotation, horizontal flipping, and brightness/contrast adjustment. Attention maps were integrated to provide visual explanations of the regions influencing model decisions, thereby supporting interpretability in clinical contexts. The fine-tuned ViT-B/16 model achieved 95% accuracy, 94% precision, 95% recall, and 94% F1-score, with AUC values of 1.00 for cataract and diabetic retinopathy, 0.99 for glaucoma, and 0.98 for normal conditions. The attention map analysis consistently highlighted clinically relevant retinal areas aligned with disease characteristics. This research contributes to the field of informatics by delivering an interpretable, high-performance AI tool for early eye disease detection and supporting scalable medical applications.]]>
Copyrights © 2026
