Ocular Microbiology and Immunology Group
Back to OMIG Main Page

< Previous | 2025 Agenda and Abstracts | Next >

2025 OMIG Abstract

External Validation of a Multi-Regionally Trained Deep Learning Model for Trachomatous Inflammation – Follicular Detection

Hady Yazbeck¹, Jad F. Assaf¹, Jeremy Keenan², John Jackson¹, Phit Upaphong¹, Xubo Song¹, Travis Redd<sup>1

1</sup>Oregon Health and Science University, Portland, Oregon; ²University of California, San Francisco, San Francisco, California

Purpose: Using external test datasets to validate a deep learning model for automated detection of trachomatous inflammation – follicular (TF), which was trained on the largest and most geographically diverse dataset of conjunctival images yet reported in the literature.

Methods: A total of 71,206 everted eyelid smartphone images from 15,605 subjects in Ethiopia (Wag Hemra), Niger, and Peru were graded by experts and used to internally train, validate and test a multi-regional deep learning model for TF detection. The model’s generalizability was then assessed on external datasets from Tanzania, Ethiopia (Goncha Siso Enese woreda), The Gambia, and the Solomon Islands. Performance was measured using F1-score, Area Under the Receiver Operating Characteristic Curve (AUROC) and TF predicted prevalence. Class activation heatmaps were generated to visualize areas contributing most to the model’s predictions.

Results: Results are reported on the Tanzania, Ethiopia (Goncha Siso Enese woreda), Solomon Islands and The Gambia test sets in that order. The F1-scores were 0.84; 0.79; 0.75 and 0.97. AUROCs were 0.98; 0.96; 0.99 and 1.00. Prevalence estimates were 0.16; 0.28; 0.05 and 0.32, compared to expert-derived estimates of 0.15; 0.22; 0.04 and 0.34, respectively. Heatmaps seemed to agree with diagnostically relevant TF features.

Conclusions: We successfully develop a deep learning model with good cross-regional generalization for TF detection. Such a model may be used to guide mass distribution of macrolide antibiotics and improve the scalability and cost effectiveness of trachoma eradication campaigns.