Machine learning applied to chest x-rays to support the diagnosis of pulmonary tuberculosis

Abstract

In 2019, around 10 million people were diagnosed with tuberculosis worldwide, resulting in 1.2 million deaths (WHO, 2020). Among imaging methods, chest X-ray (CXR) is the choice for the initial assessment of pulmonary tuberculosis (PTB). Recent advancement in the field of artificial intelligence has stimulated studies evaluating the performance of machine learning (ML) for medical diagnosis. This study aimed to validate a new original Brazilian tool, titled xmarTB, applied to CXR images to support the diagnosis of pulmonary tuberculosis (PTB). The model was trained on 3800 normal images, 3800 altered without PTB and 1376 with manifestations of PTB, from the publicly available TBX11K database. The binary classification model could distinguish between normal and abnormal CXR with a sensitivity of 99.42% and specificity of 99.40%. To detect cases of tuberculosis among CXR with alterations, the xmarTB tool obtained a sensitivity of 98.11% and a specificity of 99.74%. Therefore, applying this diagnostic tool to CXR images can accurately and automatically detect abnormal radiographs and differentiate pulmonary tuberculosis from other pulmonary diseases satisfactorily. This tool offers great potential to assist the diagnosis made by the radiologist, offering more certainty and agility, thus increasing the excellence of the professional9s performance. As emphasized by the European Society of Radiology (ESR. Insights Imaging 2022; 13:43), diagnostic ML should not replace radiologists, since beyond the diagnosis there is still the need for patient communication and interaction, human judgment for intervention and treatment, quality control, continued education and policy formulation.