Machine Learning Aided Screening of Tuberculosis in Chest X-Rays

Abstract

Efforts to eliminate TB in low and middle income countries are mainly challenged by the limited number of skilled radiologists to reliably interpret chest x-rays (CXRs) of potential patients. The use of deep learning presents a feasible approach for automatic detection of TB in CXRs which can be a valuable aid to the screening process with the potential to provide faster, more accurate results. In this work, application of deep learning towards automatic screening of TB was investigated, not only distinguishing TB-positive from healthy cases but also from Pneumonia and Covid-19, thus giving rise to a four-class classification problem. A dataset comprising 700 TB-positive images and 720 images in each of the other classes was accrued from two publicly available datasets and transfer learning was performed on five models: the EfficientNet-B4, ResNet50 and Xception based on Convolutional Neural Network (CNN) architectures and two Vision Transformer models, the ViT-B/16 and ViT-B/32. The EfficientNet-B4 emerged the best overall model with a test accuracy of 97.9%. The best performing model was deployed in a web application allowing the user to upload a chest x-ray image and get results in seconds. The results obtained indicate the potential of the model to be used in the field as a clinical decision support tool for screening of TB.