Detection of Structural Glaucoma Progression with Deep Learning on Serial Optic Disc Photographs

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Early detection and monitoring of glaucoma progression are crucial for timely intervention.
• Serial optic disc photographs are key for assessing structural changes.

Purpose of the Study:

• To design a supervised deep learning (DL) model for detecting glaucoma progression.
• To utilize serial optic disc photographs (DPs) as input for the DL model.

Main Methods:

• A retrospective longitudinal cohort study included 1,510 eyes from 916 patients with at least 2 years of follow-up.
• A twin convolutional neural network (CNN) was developed to analyze pairs of DPs.
• The dataset was split into training, validation, and testing sets (80/10/10 ratio).

Main Results:

• The DL model achieved an Area Under the Curve (AUC) of 0.821 for detecting glaucoma progression.
• The model demonstrated an overall accuracy of 72%, with 87% sensitivity and 68% specificity.
• 22% of eyes showed deterioration based on clinical review.

Conclusions:

• The developed twin CNN model can detect glaucoma progression with clinically relevant accuracy.
• Deep learning shows potential as an adjunctive tool for clinical decision-making in identifying structural glaucoma progression.