Abstract

Purpose: To compare the performance of a foundation model-based artificial intelligence (AI) system and glaucoma specialists to triage glaucoma based on severity.
Methods: Retrospective cohort study of 920 participants evaluated by glaucoma specialists at the University of Southern California. Eye-level reference labels for glaucoma were derived from the clinical notes and visual fields. RETFound was fine-tuned on a training dataset (N= 1,397 eyes) to classify color fundus photographs (CFPs) as glaucoma suspect, mild glaucoma (MD≥-6 dB), or moderate-to-severe glaucoma (MD<-6 dB). Four glaucoma specialists graded CFPs reserved in a held-out test dataset (N= 283 eyes) by the same three classes. Decision boundaries were assessed using a one-versus-rest strategy: discriminating any stage glaucoma from glaucoma suspect and moderate-to-severe glaucoma from glaucoma suspect and mild glaucoma …

Date

2026

Authors

Kyle Bolo, Abhijith Shaji, Zhiwei Li, Van D Nguyen, Brian Song, Jiun Do, Jose-Luis Ambite, Carl Kesselman, Benjamin Xu

Journal

Investigative Ophthalmology & Visual Science

Volume

67

Issue

7

Pages

577-577

Publisher

The Association for Research in Vision and Ophthalmology