Active trachoma: enhancing image classification using pretrained SOTA models and explainable AI

Abstract

BackgroundTrachoma, an infectious disease that leads to blindness, continues to pose a significant public health challenge in over 40 countries as of 2023. The initial phase of this disease, “active trachoma” is characterized by inflammation and can be effectively treated with non-surgical interventions. However, if left untreated, it progresses to the “scarring” phase, often requiring surgical intervention. Earlier detection of “active trachoma” is critical to prevent unnecessary surgery and also to reduce the transmission of the infection. Developing accessible tools for a region with limited resources is necessary. Deep neural networks have proven their effectiveness in numerous image and vision-related tasks, yet research on “active trachoma” has received still little attention.MethodIn this study, we adapted several pre-trained state-of-the-art deep neural network models like ResNet, Xception from image classification on “active classification” task. Further experiments were also conducted in three cases: training from scratch, training from pretrained models on raw images and on region-of-interest (ROI) focused images.Results and discussionThe results indicate that these models outperformed the previous studies using the same dataset, achieving an improvement of 6\% on detection of follicular trachomatous inflammation and 12\% for detection of intense trachomatous inflammation. Furthermore, we employed the eXplainable Artificial Intelligence tool Grad-CAM, which revealed a significant discrepancy between eyelid's geometric centroid and attention centroid from models with high classification accuracy. This finding suggests that the conventional method of selecting a region of interest based on the geometric centroid may need to be adjusted. Using XAI can offer valuable insights into understanding the classification and progression of active trachoma.