Abstract
Purpose :
Trachoma, an eye disease caused by Chlamydia trachomatis, is responsible for severe visual impairment/blindness of 1.9 million people worldwide. To treat this disease, WHO recommends community-wide prevalence surveys with trained health workers - a challenge in many resource-limited countries. One possible solution aims to use deep learning (DL), a subsect of artifical intelligence (AI) machine learning. Researchers from the Proctor Foundation at UCSF have developed a DL algorithm to identify trachoma from conjunctival photographs, but the algorithm has only been trained on photographs from Ethiopia - which limits its diagnostic accuracy in other populations. After further training with conjunctival photographs from the Peruvian Amazon, we hypothesize the overall diagnostic accuracy of the algorithm will improve.
Methods :
For the present study, 700 conjunctival photographs were selected from a de-identified photo repository of 0–9-year-old kids from the Peruvian Amazon taken during routine trachoma monitoring. The Peruvian photographs were split into a training set (n=600) and a testing set (n=100), and all 700 photos were manually graded and given follicular annotations based on the WHO simplified grading scale for trachomatous inflammation. To test the sensitivity and specificity, the algorithm provided a grade of trachoma (i.e., presence or absence of TF) for each photograph in the test set, and sensitivity and specificity of the algorithm was calculated twice, both pre- and post-exposure to the Peruvian pictures.
Results :
When the existing algorithm (trained only on Ethiopian photographs) was run on the 100-photograph Peruvian test set before training with Peruvian photographs, the overall accuracy was 0.50, with a sensitivity of 4% and specificity of 96%. After the addition of the Peruvian training set, the algorithm’s accuracy for the same Peruvian test set improved to 0.82, with a sensitivity of 88% and specificity of 76%.
Conclusions :
Our results are consistent with our hypothesis that with increased amounts of diverse population photos, the AI DL algorithm would consolidate and more accurately diagnose follicular trachoma overall. Using a diverse range of populations to improve machine learning for trachoma screening is important to accommodate for the diversity of populations affected by trachoma globally.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.