Abstract
Purpose :
Glaucoma screening using artificial intelligence (AI) has emerged as a transformative method to increase access to care. There is a critical need to increase the diversity and representativeness of the datasets used to train and validate AI models. We aimed to develop a deep learning model that can accurately diagnose glaucoma from fundus photography in a large African ancestry population.
Methods :
Our dataset included 64,129 fundus photographs from the Primary Open-Angle African American Glaucoma Genetics (POAAGG) study (n=42,914 images from 1,471 cases, n=21,215 images from 673 controls). We used a vision transformer (ViT), with a 70/15/15 data split for training, validation, and testing. To select the most informative images from the series of photographs taken from each subject at a single time point, we designed and tested several heuristics. These heuristics included segmentation-based classification (segments cup and disc; selects 6 images with the largest cup-disc ratios) and a binary classifier (determines whether an image is informative; selects top 6 images). To make final glaucoma predictions, we averaged the predicted probability of glaucoma across selected images and calculated the AUROC. The robustness of our model was tested in the REFUGE dataset of Chinese ancestry individuals (n=400 images in training, validation, and testing datasets, with 40 case images and 360 control images per dataset). Conformal prediction was used to obtain rigorous uncertainty quantification on the end-to-end pipeline.
Results :
The trained model had a strong AUROC (Fig. 1), with the highest AUROC (0.932) in the testing set achieved when using a segmentation-based method (Table 1). Training in the POAAGG dataset and testing in the REFUGE dataset resulted in a high AUROC (0.888); however, the opposite approach was not true (0.683). We constructed prediction sets for each image that are guaranteed to contain the true label with high probability.
Conclusions :
We developed a deep learning model that accurately predicts glaucoma in an African ancestry population. Our results highlight the importance of inclusive datasets to ensure that AI advancements benefit a global and diverse population. This model has future applications in public settings such as primary care offices, as well as low-resource settings.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.