Computer-aided detection of tuberculosis from chest radiographs in a tuberculosis prevalence survey in South Africa: external validation and modelled impacts of commercially available artificial intelligence software

This study compared the performance of 12 commercially available CAD products in detecting tuberculosis (TB) in a case–control sample of 774 participants from a South African TB prevalence survey, including 516 bacteriologically negative and 258 bacteriologically positive cases. Accuracy, measured by AUC, varied across products, with Lunit and Nexus achieving AUCs near 0.9, followed by qXR, JF CXR-2, InferRead, Xvision, and ChestEye (AUCs 0.8–0.9), while XrayAME, RADIFY, and TiSepX-TB had AUCs below 0.8. Some products (Lunit, Nexus, JF CXR-2, and qXR) maintained high sensitivity (>90%) across a wide threshold range, reducing the need for confirmatory testing. However, performance declined in older individuals, people with prior TB, and those with HIV. The study highlights significant variations in thresholds, sensitivity, and specificity across products and populations, emphasizing the need for a global validation strategy to guide CAD product and threshold selection for TB screening.