Automated ocular disease screening must resolve microscopic lesions while tracking whole-retina structural change, handle severe class imbalance, and generalize across posterior- and anterior-segment conditions. Existing hybrid CNN–transformer approaches train two full backbones end-toend, obscuring each stream’s contribution, and fuse representations through fixed concatenation regardless of input. We propose OculoViT, combining a frozen DINOv2 stream with a trainable Swin Transformer V2 stream via cross-attention and an attentionguided, input-dependent fusion gate, trained under a unified protocol of Symmetric Cross-Entropy, luminance-channel CLAHE, and training-restricted oversampling. Evaluated on 16,242 fundus and anterior-segment images across ten diagnostic categories (34.5:1 class imbalance) against DeiT III-Small and Swin V2-Tiny baselines under identical conditions, OculoViT attained 91.75% accuracy, 93.77% macro F1-score, a Cohen’s κ of 0.9037, and a macro ROC-AUC of 0.997, modestly exceeding both baselines despite overlapping confidence intervals. Notably, a configuration in which one stream received no gradient updates matched fully trained alternatives, with residual errors concentrated in visually similar glaucoma–healthy and glaucoma–myopia pairs. Grad-CAM++ and SHAP attributions localized to anatomically relevant regions, though only qualitatively, without expertannotated validation. Results suggest representational capacity is largely saturated across architectures, leaving decision-boundary refinement as the primary remaining lever for improvement.
 

Vision Transformer,Deep Learning,Explainable AI,Ocular Disease Classification