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

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Myopia, a leading cause of vision impairment, requires accurate classification for timely intervention.
• Distinguishing between high myopia (HM) and pathologic myopia (PM) is crucial for patient management.
• Automated analysis of colour fundus photographs offers potential for efficient myopia assessment.

Purpose of the Study:

• To develop and validate a self-supervised vision transformer (ViT) for automated classification of Normal, High Myopia (HM), and Pathologic Myopia (PM).
• To ensure anatomically faithful interpretability of the classification model.
• To assess the model's robustness and compare its performance against existing architectures.

Main Methods:

• Fine-tuning a DINOv2 self-supervised ViT using a two-stage transfer-learning protocol with class-balanced sampling.
• Assessing model robustness via 10-fold stratified cross-validation and evaluating an ensemble on an independent test set.
• Examining model explainability using Dual-Rollout Attention to fuse early and late-layer information.

Main Results:

• The ensemble achieved 97.03% accuracy and 97.08% macro F1-score on the independent test set.
• The DINOv2-based framework significantly outperformed ResNet-50, VGG-16, and EfficientNet-B3 (p < 0.001).
• Dual-Rollout Attention identified clinically relevant features like optic disc morphology and peripapillary atrophy.

Conclusions:

• A self-supervised ViT with transformer-specific interpretability enables accurate, robust, and transparent automated myopia classification.
• This approach has the potential to support standardized assessment and scalable screening for myopia.
• Further external and prospective validation is recommended before widespread clinical adoption.