Review of Computer-Aided Detection (CAD) Software for Tuberculosis on Chest X-Rays : A Systematic Review of Randomized Controlled Trial and Primary Studies

Abstract

Introduction: Tuberculosis (TB) remains a leading global infectious disease and major public health challenge, particularly in low- and middle-income countries. Computer-aided detection (CAD) software utilizing artificial intelligence (AI) and deep learning algorithms has emerged as a promising tool to augment chest radiograph (CXR) interpretation for pulmonary TB screening and triage. This systematic review aims to comprehensively evaluate the diagnostic performance, clinical utility, cost-effectiveness, and implementation characteristics of commercially available CAD software for TB detection on CXRs across diverse clinical and epidemiological settings. Methods: The study strictly adhered to the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) 2020 guidelines. Eligible study designs included randomized controlled trials (RCTs), prospective and retrospective cohort studies, case-control studies, and cross-sectional studies evaluating at least one commercially available CAD software for pulmonary TB detection on digital CXRs with microbiological reference standards. Risk of bias was assessed using QUADAS-2 and Cochrane RoB 2.0. Results: Seventeen primary studies encompassing over 130,000 participants across Africa, Asia, Europe, Oceania, and Latin America were included. Evaluated CAD products included CAD4TB (versions 5-7), qXR (versions 2-3.2), Lunit INSIGHT CXR (versions 3.1-4.9), JF CXR-1/2, and InferRead DR. Area under the receiver operating characteristic curve (AUROC) values ranged from 0.70 (paediatric populations) to 0.92 (unselected adult populations). At 90% sensitivity, specificity ranged from 22.2% (prior TB history populations) to 84% (prison settings). Multiple CAD products met WHO Target Product Profile (TPP) minimum thresholds (≥90% sensitivity and ≥70% specificity) in symptomatic adult populations. CAD outperformed human radiologists in several large-scale studies. Performance was consistently lower in people living with HIV, elderly individuals, persons with previous TB, and paediatric patients. Discussion: Accumulating evidence substantiates that AI-based CAD systems are accurate, scalable, and cost-effective tools for pulmonary TB screening. However, heterogeneity in performance across software versions, geographic populations, and patient subgroups underscores the necessity for local threshold calibration and version-specific validation. Integration of CAD into national TB programs requires addressing economic, regulatory, and data governance challenges. Conclusion: CAD software demonstrates substantial diagnostic accuracy for pulmonary TB on CXRs and holds considerable promise for scaling up TB case finding. Standardized evaluation frameworks, population-specific threshold optimization, and equitable access are prerequisite conditions for maximizing the public health impact of CAD-CXR systems globally.