Ocular Microbiology and Immunology Group
Back to OMIG Main Page

< Previous | 2026 Agenda and Abstracts | Next >

2026 OMIG Abstract

POSTER PRESENTATION


Class-Conditional Correlation Alignment for Generalizable Automated Trachoma Detection

Ifrah Khurram¹, Sourav Kumar¹, Hady Yazbeck², and Travis Redd¹

¹University of Colorado School of Medicine, Aurora, Colorado; ²Oregon Health & Science University, Homestead, Oregon

Purpose: Automated trachoma grading from conjunctival imaging could improve the consistency and scalability of trachoma screening, but poor generalizability remains a barrier due to regional differences in imaging conditions and disease prevalence, particularly in low prevalence regions with limited positive examples. We evaluated whether a statistical image alignment technique we call Class-Conditional CORelation ALigment (CC-CORAL) could improve generalizability across diverse regions, including a near-zero-prevalence population, without requiring data to be shared across sites.

Methods: We analyzed 11,358 conjunctival images from three sites with varying trachoma prevalence: Ethiopia (31%), Peru (26%), and Niger (0.64%). Deep learning models were trained on data from each site individually, and evaluated on data from other sites. We then applied CC-CORAL and evaluated the impact on model performance across each dataset.

Results: CC-CORAL substantially improved cross-site performance, increasing the average F1 for models evaluated on external sites from 0.61 to 0.79. The largest gains were seen when transferring to Ethiopia, where F1 improved from 0.40 to 0.87 for the Niger-trained model and from 0.40 to 0.78 for the Peru-trained model. Ethiopia in-domain performance decreased from 0.87 to 0.82. Performance on Niger, a near-zero-prevalence site with only 53 positive-class images, also improved: F1 increased from 0.62 to 0.65 for the Ethiopia-trained model and from 0.56 to 0.66 for the Peru-trained model. Niger in-domain performance increased from 0.53 to 0.61. Performance on Peru improved from 0.88 to 0.91 for the Ethiopia-trained model, 0.77 to 0.89 for the Niger-trained model, and stayed the same at 0.91 for Peru in-domain performance.

Conclusions: Statistical feature alignment can substantially improve the cross-site generalizability of automated trachoma grading, including in low-prevalence field settings like Niger.