Humans versus machines: The use of AI for chest X-ray interpretation to diagnose laboratory confirmed Tuberculosis among people presenting for primary healthcare in Karnataka, India.

Artificial intelligence (AI) may assist healthcare providers in diagnosing active tuberculosis (TB) by interpretation of chest X-rays, especially in settings with limited access to doctors. <bold>Aim:</bold> To assess if AI radiography is comparable to humans in detecting laboratory confirmed active TB in a high endemic setting. <bold>Methods:</bold> Cross sectional study of people presented to a primary health care center in Bangalore, India. Participants underwent digital or traditional chest-rays interpreted by AI (“<italic>CAD4TB</italic>” score 0-100, <italic>Delft systems</italic>). Two experienced physicians also independently scored the X-rays as probable, possible or unlikely tuberculosis. Predictive scores >60 were considered highly suspicious for tuberculosis. Participants were followed to determine if TB was confirmed microbiologically (culture or PCR). <bold>Results:</bold> 90 participants with X-rays were included and 78 (87%) achieved an interpretable AI score. AI considered 31 of 90 (34%) as high suspicion for TB and humans identified 26 (29%) as either probable (5) or possible (21). Discordant results were found in 5 of the possible cases (24%) and 1 (20%) of probable cases had CAD4TB scores <60. Laboratory confirmation of TB occurred in 3 (60%) probable and 2 (10%) possible cases, however follow up was incomplete. One rescanned image was assigned a CAD4TB score of 86 and subsequently found positive for TB. <bold>Conclusion:</bold> AI enhanced chest X-ray compares well to physician interpretation to identify people with active TB and appears to correlate well to microbiologically-confirmed TB disease