Hybrid eventIn-person & virtual
March 15–17, 2027Singapore · Village Hotel Changi

Talk Details · IOC 2027

High-sensitivity artificial intelligence triage for referable glaucoma using color fundus photographs

Oral · In-Person · Day 1

Abstract

About this talk

Purpose: To evaluate a deep learning model for high-sensitivity triage of referable glaucoma from color fundus photographs and assess its potential role in a clinical risk-assessment pathway.

Methods: In this retrospective diagnostic accuracy study, we used 17,403 photographs from a subset of the public JustRAIGS training dataset. Images were partitioned into training (10,433), validation (3,493) and held-out test (3,477) sets, keeping linked fellow eyes in the same partition. We fine-tuned an EfficientNet-B0 model with weighted binary cross-entropy to address class imbalance. A decision threshold targeting at least 95% sensitivity was selected on validation data and fixed before test evaluation. The model used images alone; clinical risk factors were not inputs.

Results: The test set contained 132 referable and 3,345 non-referable images. The area under the receiver operating characteristic curve was 0.937, and the area under the precision-recall curve was 0.597. Sensitivity was 94.7% (125/132; 95% CI 90.3–98.4%) and specificity was 70.3% (2,352/3,345; 95% CI 68.7–71.9%). Negative and positive predictive values were 99.7% and 11.2%, respectively, at a referable prevalence of 3.8%. Overall, 2,359 images (67.8%) were classified as non-referable, including seven false negatives; 1,118 (32.2%) were flagged for referral, including 993 false positives.

Conclusions: Image-based triage detected most referable glaucoma images while classifying approximately two-thirds of the test set as non-referable. These findings support further evaluation within a clinical risk-assessment pathway incorporating older age, measured intraocular pressure and first-degree family history of primary open-angle glaucoma. These factors would guide screening and follow-up alongside the image-based score. The combined pathway was not tested in this study. Prospective external validation is required to establish clinical utility, referral burden and safeguards against missed disease.

Speaker

About the speaker

Amna Ikram

Amna Ikram

Shaikh Zayed Postgraduate Medical Institute Lahore, Pakistan

View full speaker profile →

Join the IOC 2027 delegation

Singapore · March 15–17, 2027· Hybrid · in-person & virtual

Register as Delegate →

Attend without presenting.

Chat with us