Highly Accurate and Precise Automated Cup-to-Disc Ratio Quantification for Glaucoma Screening

Area of Science:

• Ophthalmology and Artificial Intelligence
• Medical Imaging Analysis
• Glaucoma Diagnostics

Background:

• Enlarged cup-to-disc ratio (CDR) is a key indicator of glaucomatous optic neuropathy.
• Manual CDR assessment can be subjective, time-consuming, and less accurate than automated methods.

Purpose of the Study:

• To develop and validate a deep learning algorithm for automated CDR determination from fundus images.
• To assess the algorithm's utility in glaucoma screening.

Main Methods:

• A convolutional neural network (CNN) model was trained using FastAI and PyTorch on fundus images from the UK Biobank (UKBB).
• The model was developed for both image gradability classification and CDR estimation (regression analysis).
• Validation was performed using multiethnic datasets from EyePACS and Drishti_GS.

Main Results:

• The gradability model achieved 97.13% accuracy, 99.26% precision, and an AUC of 96.56%.
• The best-performing regression model for CDR estimation achieved a coefficient of determination of 0.8514.
• External validation on the EyePACS dataset for glaucoma classification yielded 82.49% accuracy, 72.02% sensitivity, and 82.83% specificity.

Conclusions:

• The developed AI models demonstrate high precision in image gradability and CDR estimation.
• The AI-derived CDR estimates are accurate, but the threshold for glaucoma screening requires consideration of other clinical factors.
• The algorithm shows potential for improving the efficiency and accuracy of glaucoma screening.