A systematic review and meta-analysis of artificial intelligence software for tuberculosis diagnosis using chest X-ray imaging.

BACKGROUND

Pulmonary tuberculosis (PTB) remains a global public health challenge, with 10.8 million new cases reported in 2023. Early diagnosis is crucial for controlling its spread, yet traditional sputum-based tests face limitations in turnaround time and resource availability. Chest X-ray (CXR) is a cost-effective diagnostic tool, but its use in high-tuberculosis (TB) burden regions is restricted by a shortage of radiologists. Artificial intelligence (AI)-based computer-aided detection (CAD) systems, leveraging deep learning, offer a promising solution for automated PTB detection. However, variability in diagnostic performance across AI tools and the need for scenario-specific threshold adjustments remain challenges that need to be addressed. Our meta-analysis evaluated the diagnostic accuracy of five AI-based PTB detection products, aiming to provide insights for advancing AI applications in TB screening and diagnosis.

METHODS

The PubMed, Embase, Web of Science, and Cochrane Library databases were searched for literature related to CXR diagnosis of TB based on AI technology published from the establishment day of the database to December 19, 2024. The keywords were "artificial intelligence", "tuberculosis", "chest X-ray", and "diagnosis". The literature search, screening, data extraction, quality evaluation, and bias risk assessment were conducted independently by two researchers, and Stata 17.0 software (StataCorp) was used to process and analyze the data.

RESULTS

A total of 5,651 references were retrieved, and 21 references were finally selected according to the inclusion and exclusion criteria. The meta-analysis included five software solutions for CXR analysis: JF CXR-1 (JF Healthcare, Nanchang, China), qXR (Qure.ai, Mumbai, India), Lunit INSIGHT CXR (Lunit, Seoul, South Korea), CAD4TB (Delft Imaging, 's-Hertogenbosch, Netherlands), and InferRead DR Chest (Infervision, Beijing, China). Their sensitivity and specificity were as follows: JF CXR-1, 86.0% and 80.0%; qXR, 90.0% and 64.0%; Lunit INSIGHT CXR, 90.0% and 63.0%; CAD4TB, 91.0% and 60.0%; InferRead DR Chest, 89.0% and 59.0%.

CONCLUSIONS

AI software has demonstrated excellent diagnostic performance in assisting the CXR diagnosis of TB and can help clinicians to make rapid and accurate decisions in screening and treating patients with TB.