Creating and Verifying a Deep Learning-Powered Automatic Algorithm for Chest Radiographs to Identify Active Pulmonary Tuberculosis

Abstract

Chest radiographs (CRs) that demonstrate active pulmonary tuberculosis are required to screen and diagnose conditions related to tuberculosis. An automated system has the potential to improve diagnostic performance while also accelerating the process of TB screening, which is a source of optimism. To build a deep learning-based automatic detection (DLAD) approach, thirteen radiologists who are board-certified went through 54,211 normal CRs and 6,778 CRs with active pulmonary tuberculosis. To ensure that DLAD was effective, six external, multinational datasets were used. To compare the efficacy of DLAD with that of physicians, fifteen doctors including thoracic radiotherapists, board-certified radiotherapists, and non-radiology physicians took part in an observer performance exam. To measure the effectiveness of lesion-wise localization along with image-wise classification, respectively, the area over the ROC (receiver operating characteristic) curves and the area under a separate free-response ROC curve were used. Determined were the DLAD's sensitivity and specificities using two cutoffs: high sensitivity [98%] and high specificity [98%] established by internal validation. It was shown that DLAD achieved a classification performance of 0.977-1.000 and a localisation performance of 0.973-1.000. In contrast to the high-specificity cutoff, which generated sensitivity and specificities ranging from 84.1% to 99.0% and 99.1% to 100%, correspondingly, the high-sensitivity cutoff produced 94.3% to 100% sensitivities and 91.1% to 100% specificities. Both localization (0.993 vs. 0.664-0.925) and classifying (0.993 vs. 0.746-0.971) were areas where DLAD outperformed all physician groups. When it came to reliably identifying active pulmonary TB upon CR, the DLAD fared better than clinicians, particularly thoracic radiologists.