Development of deep learning model to screen for primary open-angle glaucoma in African ancestry individuals

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Primary open-angle glaucoma (POAG) is a leading cause of blindness, particularly affecting individuals of African ancestry.
• Current artificial intelligence (AI) models for POAG screening lack representation from understudied populations.
• There is a critical need for equitable AI solutions in glaucoma detection.

Purpose of the Study:

• To develop and validate a deep learning model for POAG screening using fundus photography.
• To address the under-representation of African ancestry individuals in AI datasets for glaucoma.
• To create a robust AI tool applicable in diverse clinical and low-resource settings.

Main Methods:

• A deep learning model was trained on 64,129 fundus images from the Primary Open-Angle African American Glaucoma Genetics (POAAGG) cohort.
• The pipeline involved image selection using a binary classifier and POAG probability prediction via a Vision Transformer.
• Final predictions were derived by averaging probabilities from selected images.

Main Results:

• The model achieved an Area Under the Curve (AUC) of 0.925 on the POAAGG dataset.
• Validation on the REFUGE-1 dataset (Chinese ancestry) yielded an AUC of 0.920.
• The AI model demonstrated high accuracy in screening for POAG.

Conclusions:

• The developed AI model shows significant potential for effective POAG screening.
• The model's performance across different ancestries highlights its generalizability.
• This AI tool can be deployed in various settings, including primary care and resource-limited areas, to enhance glaucoma detection.