Machine Learning Applications in Tuberculosis Detection Using Chest X-Ray Imaging

Abstract

Tuberculosis (TB) is still a primary cause of mortality worldwide and chest X-ray (CXR) imaging is still a first-line, low cost test in high burden areas. Interpretation of CXRs, however, is time consuming, needs radiologists with the expertise to interpret CXRs and is subject to inter observer variability, making interpretation a diagnostic bottleneck in resource limited settings. New techniques in the field of AI like machine learning (ML) and deep learning (DL), notably convolutional neural networks (CNNs) provide the possibility of automated accurate and scalable TB screening directly from the CXR images. This article aims to review and analytically discuss the various ML applications for TB detection on chest X-rays. We explore the history of computer-aided detection (CAD) systems, review publicly available CXR datasets for TB research and outline the typical machine learning system, ranging from image acquisition to preprocessing to lung segmentation to feature extraction to classification. We also compare classical machine learning classifiers with the performance of deep transfer-learning architectures like VGG16, ResNet50, InceptionV3, DenseNet121 and EfficientNet, reporting accuracy, sensitivity, and specificity from the literature. The challenges of limited data, imbalanced data, domain shift across scanners and populations, interpretability, and regulatory issues are explored as well as some new solutions that are starting to be developed in the community, including federated learning, explainable AI, and lightweight models that can be deployed on the edge. The results indicate that deep learning models can perform at the level of radiologists on curated benchmark datasets, but there are still challenges to investigate for real-world clinical use, such as explainability, generalizability, and integration with current health infrastructure. This review is designed for the use as reference for researchers and practitioners developing next generation of AI-assisted TB screening tools.

Keywords: Tuberculosis; Chest X-Ray; Machine Learning; Deep Learning; Convolutional Neural Networks; Computer-Aided Diagnosis; Medical Imaging; Transfer Learning