Explainable deep learning self-supervised vision transformer for fundus-based myopia classification

To develop and validate a self-supervised vision transformer (ViT) for automated three-class classification of Normal, High Myopia (HM), and Pathologic Myopia (PM) from colour fundus photographs, with anatomically faithful interpretability.

A DINOv2 self-supervised ViT was fine-tuned using a two-stage transfer-learning protocol with class-balanced sampling. Model robustness was assessed using 10-fold stratified cross-validation, and an ensemble of the best fold-specific checkpoints was evaluated once on a strictly held-out independent test set. Comparative performance was assessed against ResNet-50, VGG-16, and EfficientNet-B3 using paired, two-sided Student's t-tests on fold-level macro-F1 scores. Model explainability was examined using a Dual-Rollout Attention approach that fuses early-layer spatial detail with late-layer semantic information.

On the independent test set, the ensemble achieved an overall accuracy of 97.03% (95% CI 96.75-97.32%) and a macro F1-score of 97.08% (95% CI 96.80-97.36%). The DINOv2-based framework outperformed ResNet-50, VGG-16, and EfficientNet-B3 architectures (all p < 0.001). Dual-Rollout Attention highlighted clinically relevant regions including optic disc morphology, peripapillary atrophy, and areas consistent with chorioretinal thinning.

A self-supervised ViT combined with transformer-specific interpretability enables accurate, robust, and transparent automated classification of high and pathologic myopia from fundus photographs. This approach may support standardized assessment and scalable screening, pending external and prospective validation.