Validation of several TB-CAD chest-X-ray applications in individuals with presumptive TB visiting peripheral health institutes in Delhi State

Before deploying digital X-ray with Computer Aided Detection (CAD) as a triage tool for tuberculosis (TB), selecting an appropriate product and threshold score is essential to identify patients requiring confirmatory TB testing. We conducted a combined retrospective case-control and prospective cross-sectional study to evaluated the performance and optimal threshold for qXR v3 (Qure.ai, India) as well as six additional CAD products in individuals with presumptive TB attending peripheral health institutes (PHIs) in India. Accuracy was assessed against a microbiological reference standard separately for adults (≥16 year) and children (6-16 year). Among 1245 adults (315 TB-positive, 930 TB-negative) and 159 children (39 TB-positive, 120 TB-negative), qXR demonstrated high accuracy both in adults (AUCROC: 0.88 [95% CI of 0.85-0.90], and children (AUCROC: 0.95 [95% CI 0.89-1.0]). and performed as good as the radiologist in both groups. Sensitivity increased with minimal loss in specificity when using the vendor recommended threshold. CAD4TB, Insight CXR, DrAid, and Genki also demonstrated high accuracy (AUCROC: adults ≥0.80, AUCROCs children ≥0.90), while InferRead DR Chest and Radify Chest performed less well. Local validation confirmed high accuracy for qXR and several other products, in identifying TB in adults and children in India, supporting their potential implementation in similar settings.