Ocular Microbiology and Immunology Group
Back to OMIG Main Page

< Previous | 2026 Agenda and Abstracts | Next >

2026 OMIG Abstract

Multiregional Deep Learning for Active Trachoma Using Latent Class Analysis-Derived Probabilistic Labels

Jad F. Assaf^1,2, Hady Yazbeck¹, Phit Upaphong¹, John Jackson¹, Xubo Song³, Jeremy Keenan⁴,
and Travis K. Redd⁵

¹Casey Eye Institute, Department of Ophthalmology, OHSU, Portland, Oregon; ²Department of Ophthalmology, Nazareth Hospital, Philadelphia, Pennsylvania; ³Department of Medical Informatics and Clinical Epidemiology and Program in Computer Science and Electrical Engineering, OHSU, Portland, Oregon; ⁴Francis I. Proctor Foundation, UCSF, San Francisco, California; ⁵Department of Ophthalmology, University of Colorado Anschutz Medical Campus, Aurora, Colorado

Purpose: Most automated trachoma models are trained to detect trachomatous inflammation–follicular (TF) alone, although active trachoma may manifest as TF and/or trachomatous inflammation–intense (TI). We used latent class analysis (LCA) to integrate TF and TI grades into a probabilistic active-trachoma label and trained a multiregional deep learning model against this label.

Methods: We retrospectively analyzed 71,206 everted upper-eyelid photographs from 15,605 children aged 0-9 years in Ethiopia, Niger, and Peru from 2014 to 2020. Certified graders assessed TF and TI, and polymerase chain reaction results were available for a subset. LCA integrated these indicators into image-level probabilities of active trachoma. A MobileNetV3 model was trained on 90% of the data using fivefold cross-validation and evaluated on independent 10% hold-out sets.

Results: In Ethiopia, Niger, and Peru, respectively, test-set accuracy was 85%, 96%, and 88%, and the area under the receiver operating characteristic curve was 0.94, 0.89, and 0.91. Model-derived versus evaluator-based prevalence estimates were 28.70% versus 28.61%, 0.86% versus 1.15%, and 33.77% versus 31.17%, respectively.

Conclusions: Training against an LCA-derived probabilistic label incorporating both TF and TI grades enabled the model to estimate active trachoma rather than TF alone. Regional prevalence estimates closely matched evaluator-based estimates in held-out testing. Validation in independent populations is required before use in trachoma surveillance or mass drug administration planning.