# DINO-EYE: self-supervised learning for identification of different optic disc phenotypes in primary open angle glaucoma

## TL;DR

A self-supervised learning model called DINO-EYE accurately classifies optic disc phenotypes in glaucoma patients and outperforms existing models in performance and interpretability.

## Contribution

The novel DINO-EYE model uses self-supervised learning to classify optic disc phenotypes in primary open angle glaucoma with high accuracy and interpretability.

## Key findings

- DINO-EYE achieved 91% accuracy in classifying optic disc phenotypes using Random Forest and 92.1% after merging similar phenotypes.
- Unsupervised clustering revealed coherent groupings for concentric thinning and extensive Peripapillary Atrophy.
- DINO-EYE outperformed the RETFound SSL model in phenotype classification and provided interpretable attention regions.

## Abstract

To develop a self-supervised learning (SSL) model that classifies optic disc phenotypes in primary open angle glaucoma (POAG) and explores novel phenotypic patterns with optic disc photographs (ODPs). We collected 850 ODPs from patients with POAG and applied data augmentation to address class imbalances, yielding 10,493 images. Using the DINO Vision Transformer as the backbone model, we trained an SSL model to extract 2048-dimensional latent features. These features were used for both supervised classification of six known phenotypes and unsupervised clustering. Classification performance was evaluated with Random Forest and XGBoost models. UMAP (Uniform Manifold Approximation and Projection) was used for dimensionality reduction and feature visualization, and attention maps were generated for model interpretability. The DINO-EYE model features enabled phenotype classification with 91% accuracy with Random Forest and 92.1% after merging clinically similar phenotypes. Unsupervised clustering revealed coherent groupings, particularly for concentric thinning and extensive Peripapillary Atrophy (PPA), though no new phenotypes were unanimously confirmed by clinicians. The proposed model outperformed the RETFound SSL model in phenotype classification and demonstrated interpretable attention regions consistent with expert criteria. Our DINO-EYE effectively extracts clinically meaningful features from fundus images and enables accurate classification of optic disc phenotypes in POAG. It surpasses existing SSL models in performance and interpretability, offering promise for real-world glaucoma decision support and individualized care planning.

The online version contains supplementary material available at 10.1038/s41598-025-33140-1.