Deep learning for detecting pulmonary tuberculosis via chest\n radiography: an international study across 10 countries

Tuberculosis (TB) is a top-10 cause of death worldwide. Though the WHO\nrecommends chest radiographs (CXRs) for TB screening, the limited availability\nof CXR interpretation is a barrier. We trained a deep learning system (DLS) to\ndetect active pulmonary TB using CXRs from 9 countries across Africa, Asia, and\nEurope, and utilized large-scale CXR pretraining, attention pooling, and noisy\nstudent semi-supervised learning. Evaluation was on (1) a combined test set\nspanning China, India, US, and Zambia, and (2) an independent mining population\nin South Africa. Given WHO targets of 90% sensitivity and 70% specificity, the\nDLS's operating point was prespecified to favor sensitivity over specificity.\nOn the combined test set, the DLS's ROC curve was above all 9 India-based\nradiologists, with an AUC of 0.90 (95%CI 0.87-0.92). The DLS's sensitivity\n(88%) was higher than the India-based radiologists (75% mean sensitivity),\np<0.001 for superiority; and its specificity (79%) was non-inferior to the\nradiologists (84% mean specificity), p=0.004. Similar trends were observed\nwithin HIV positive and sputum smear positive sub-groups, and in the South\nAfrica test set. We found that 5 US-based radiologists (where TB isn't endemic)\nwere more sensitive and less specific than the India-based radiologists (where\nTB is endemic). The DLS also remained non-inferior to the US-based\nradiologists. In simulations, using the DLS as a prioritization tool for\nconfirmatory testing reduced the cost per positive case detected by 40-80%\ncompared to using confirmatory testing alone. To conclude, our DLS generalized\nto 5 countries, and merits prospective evaluation to assist cost-effective\nscreening efforts in radiologist-limited settings. Operating point flexibility\nmay permit customization of the DLS to account for site-specific factors such\nas TB prevalence, demographics, clinical resources, and customary practice\npatterns.\n