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

AI / informaticsActa ophthalmologica· Jul 2026

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

Eirini Maliagkani, Christos Tsoutsas, Nikolaos Papageorgiou … Elpiniki Papageorgiou

Read paper

Cataract & RefractiveAI / informaticsQ1IF 3.0

AI vision transformer classified normal, high, and pathologic myopia from fundus photos with high accuracy.

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

Eirini Maliagkani … Elpiniki Papageorgiou

Acta ophthalmologica · Jul 2026 · Q1 in Ophthalmology

Purpose

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

Methods

A DINOv2 self-supervised vision transformer was fine-tuned using a two-stage transfer-learning protocol with class-balanced sampling.

97.03%

Results

97.03%

correct fundus-photo myopia classification in about 97 of every 100 cases

More results

The DINOv2-based framework outperformed ResNet-50, VGG-16, and EfficientNet-B3 architectures (all p < 0.001).

“

Conclusion · 1 of 2

A self-supervised vision transformer combined with transformer-specific interpretability enables accurate, robust, and transparent automated classification of high and pathologic myopia from fundus photographs.

Conclusion · 2 of 2

This approach may support standardized assessment and scalable screening, pending external and prospective validation.

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

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

Related papers