Abstract
Purpose :
Despite intensive efforts, trachoma remains the most important infectious cause of vision loss and one of the overall leading causes of global blindness. Current methods for detection may be inadequate for elimination programs, so we sought to develop an artificial intelligence classifier (AI) for follicular trachoma (TF) using conjunctival images to allow for rapid, low-cost remote detection.
Methods :
A smartphone camera collected 2614 upper eyelid images during a 2019 district survey in Chamwino, Tanzania. The subset of 2285 gradable images with concordant field and photo grades were re-sized to 224 x 224 pixels and randomly divided into equal training and test sets. To enrich the training set, oversampling of TF cases with data augmentation methods (horizontal flipping, rotation, perspective shift, and color jitter) were applied. Additionally, color space exploration and follicle enhancement methods were employed. Two commonly used pre-trained convolutional neural networks (VGG16 and Resnet101) were further trained on “training set.” Class weighting and batch accumulation were used to improve model performance.
Results :
With a goal of maximizing recall, a ResNet101 model generated the best results, with a recall of 89%. This translated to a 75% reduction in skilled grader burden as only positive images require confirmation. Additionally, the misclassification of images in this work is typically due to poor image quality (e.g., failure to capture entire tarsal plate, image focus, etc.).
Conclusions :
AI methods show promise in identifying TF, though the training data was limited by low case numbers of TF and non-standardized images. Future studies will retrain the model with larger datasets with more TF and will develop standardized photographic techniques (improved image quality, centration and lighting). While trachoma efforts are decentralized and organized at the country level, we believe photographic standardization using smartphone camera technology is reasonable given the widespread utilization of a single mHealth app (Tropical Data) to support TF prevalence surveys. To allow for practical value in areas of low network connectivity, edge computing strategies such as in-handset classification may be needed.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.