Abstract

Purpose: To compare the performance of a fine-tuned foundation model against a supervised learning-based transfer learning model in detecting referable glaucoma from fundus photographs in a screening population.
Methods: Fundus photographs taken on 3 different cameras (Topcon NW400, Topcon NW8, Canon CR-2 AF Digital) were obtained from the Los Angeles County Department of Health Services (LAC DHS) Teleretinal Screening Program and labeled for referable glaucoma (cup-to-disc ratio≥ 0.6). Four deep learning models were trained on full and cropped images from this dataset (Training N= 8,996; Validation N= 3,002) using two different architectures: a large vision transformer with self-supervised pretraining on fundus photographs (RETFound) and a deep convolutional neural network (VGG-19). Model performance was evaluated on a separate internal test set (N= 1,000) using area under the receiver operating characteristic curve (AUC-ROC) metrics while varying the size of the training set and stratifying the test set by race.
Results: The RETFound model trained on full fundus images (N= 8,996) achieved the highest AUC-ROC (0.915 [0.898-0.930]), followed by the cropped image RETFound model (0.911 [0.898-0.930]), the cropped image VGG-19 model (0.893 [0.875-0.920]), and the full image VGG-19 model (0.881 [0.861-0.899]). All pairwise comparisons against the full image RETFound model reached significance in two-sided t-tests of the bootstrapped test results (p< 0.001). The performance advantage was more pronounced when training on datasets of N≤ 500: AUC-ROCs for the best RETFound and VGG-19 models …

Date

June 30, 2025

Authors

Kyle Bolo, Sreenidhi Munimadugu, Zhiwei Li, Tran Huy Nguyen, Van Nguyen, Jiun Do, Brandon Wong, Lauren Daskivich, Michael Pazzani, Jose-Luis Ambite, Carl Kesselman, Benjamin Xu

Journal

Investigative Ophthalmology & Visual Science

Volume

66

Issue

8

Pages

422-422

Publisher

The Association for Research in Vision and Ophthalmology