Pulmonary tuberculosis prediction using CAD4TB artificial intelligence (computer-aided detection for tuberculosis) based on thoracic x-ray photos among Indonesian subjects in hospital

Tuberculosis remains a major global health concern, particularly in high-burden countries where early detection is essential but often limited by insufficient radiological expertise. This study evaluated the diagnostic performance of a computer-aided detection system, CAD4TB, in interpreting chest X-ray images of suspected tuberculosis cases in a hospital setting in Indonesia. Using a retrospective, cross-sectional design, we analyzed chest radiographs from over 1,100 adult patients drawn from the national tuberculosis database. Images were processed using the CAD4TB system and independently reviewed by two experienced radiologists. Bacteriological test results were used as the diagnostic reference standard. At a CAD4TB index cutoff of 60, the tool achieved a sensitivity of 81.04% and a specificity of 63.80%. In comparison, radiologist interpretations achieved a sensitivity of 88.20% and a specificity of 58.18%. Subgroup analyses revealed improved diagnostic performance in individuals without pleural effusion, with CAD4TB sensitivity rising to 83.65% and radiologist sensitivity to 90.35%. CAD4TB also showed consistent specificity advantages across clinical subgroups, including those with prior tuberculosis history and HIV-negative status. These findings support the potential role of CAD4TB in assisting radiologists within hospital settings, especially in high-burden areas. Its time-efficiency and ease of use make it a valuable tool for integration into tuberculosis triage systems, particularly where patient complexity varies and access to expert readers is limited. Author SummaryTuberculosis is still one of the worlds leading infectious diseases, especially in countries like Indonesia where many cases go undetected due to limited access to expert medical imaging professionals. In this study, we explored whether a computer system called CAD4TB, which uses artificial intelligence to read chest X-rays, could help identify people with tuberculosis in a hospital setting. We compared its performance to experienced radiologists and used laboratory tests to confirm the results. We found that CAD4TB was able to detect tuberculosis with similar accuracy to human experts in many cases, especially among patients without complications like fluid in the lungs. Although expert radiologists remain slightly more accurate overall, CAD4TB performed well enough to suggest that it could be used to support hospital teams--especially where there are not enough trained readers. Because this system can analyze X-rays quickly and consistently, we believe it could be useful in busy hospitals and in regions with high numbers of tuberculosis cases. Our findings may help health programs consider how digital tools like CAD4TB can be integrated into screening and diagnosis strategies to improve early detection and treatment.