Diagnostic accuracy of three computer-aided detection systems for detecting pulmonary tuberculosis on chest radiography when used for screening: Analysis of an international, multicenter migrants screening study

This retrospective diagnostic accuracy study aimed to independently evaluate three AI-based computer-aided detection (CAD) systems—CAD4TB v6, Lunit INSIGHT v4.9.0, and qXR v2—for detecting pulmonary tuberculosis (TB) on chest X-rays (CXR) of global migrants screened through the International Organization for Migration (IOM) database for US-bound migrants. The rationale behind the study was to assess the effectiveness of CAD systems as potential tools for TB screening, particularly in settings with limited access to expert radiologists. A total of 2,812 participants were analyzed against the radiological reference standard (RadRS), of which 1,769 (62.9%) had microbiological results for comparison with the microbiological reference standard (MRS). The results showed that Lunit outperformed the other CAD systems against MRS, with an area under the curve (AUC) of 0.85, followed by qXR (0.75) and CAD4TB (0.71). At a set specificity of 70%, Lunit achieved the highest sensitivity of 81.4%, and at a set sensitivity of 90%, it also had the highest specificity of 54.5%. Against RadRS, both CAD4TB and Lunit showed the highest AUC of 0.87, outperforming qXR (0.81). The study concluded that all three CAD systems had comparable diagnostic accuracy to expert radiologists and could be valuable tools for TB screening in resource-limited, high-prevalence settings. However, it highlighted the need for further large-scale prospective studies to explore their operational and technical effectiveness, as the retrospective nature of the study and limited MRS testing in low-suspicion cases were noted as key limitations.