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

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Primary open angle glaucoma (POAG) diagnosis relies on identifying optic disc phenotypes from optic disc photographs (ODPs).
• Accurate classification of these phenotypes is crucial for effective glaucoma management and treatment planning.
• Existing methods may lack the precision or interpretability needed for comprehensive analysis.

Purpose of the Study:

• To develop a self-supervised learning (SSL) model for classifying optic disc phenotypes in POAG.
• To explore novel phenotypic patterns using optic disc photographs (ODPs).
• To enhance glaucoma diagnosis and patient care through advanced AI analysis.

Main Methods:

• Collected 850 ODPs from POAG patients, augmented to 10,493 images.
• Utilized DINO Vision Transformer for SSL to extract 2048-dimensional features.
• Employed Random Forest and XGBoost for classification, UMAP for visualization, and attention maps for interpretability.

Main Results:

• The DINO-EYE model achieved 91% accuracy in phenotype classification (92.1% after merging similar phenotypes).
• Unsupervised clustering identified distinct groupings, notably for concentric thinning and extensive Peripapillary Atrophy (PPA).
• The model demonstrated superior performance and interpretability compared to the RETFound SSL model.

Conclusions:

• DINO-EYE effectively extracts clinically meaningful features from fundus images for accurate POAG optic disc phenotype classification.
• The model surpasses existing SSL methods in performance and interpretability, offering potential for clinical decision support.
• This AI approach promises to advance individualized glaucoma care planning and patient management.