Artificial intelligence-assisted tuberculosis screening in Viet Nam

Background: Tuberculosis (TB) remains one of the leading infectious causes of death worldwide, with delayed diagnosis contributing substantially to ongoing transmission. Chest X-ray (CXR) screening combined with artificial intelligence- powered computer-aided detection (AI-CAD) software has emerged as a promising strategy to expand access to systematic TB screening, particularly in settings with limited radiology services. However, important evidence gaps remain regarding the optimal integration of AI-CAD software into TB screening programmes.

Aims: This thesis aims to evaluate the performance of commercially available AI- CAD software for TB screening and to investigate how this technology can be integrated into screening programmes to improve detection of TB and the efficiency of care.

Methods: Four complementary studies were conducted. Study I used aggregate diagnostic testing data from TB screening programmes in Bangladesh, Nigeria, Viet Nam and Zambia to model and compare diagnostic testing approaches, including CXR screening, indiscriminate pooled testing and AI-CAD guided pooled testing. Studies II, III, and IV used test libraries curated from community-based active case finding activities in Ho Chi Minh City, Viet Nam. Study II evaluated the diagnostic accuracy of successive versions of a commercially available AI-CAD software using a microbiological reference standard. Study III compared AI-CAD software performance when analysing original digital CXR images and photographs of printed CXR films using a novel approach to assign a composite reference standard for TB to test library participants. Study IV compared the diagnostic accuracy and operational performance of different CXR interpretation methods, including AI-CAD only, AI-CAD assisted, double and sequential reading, to determine how AI-CAD software can most effectively complement human readers during TB screening; this study used the same test library as Study III. AI- CAD software performance was primarily evaluated using receiver operating characteristic analyses and the calculation of sensitivity and specificity, while modelling approaches were used to estimate the operational impact of alternative screening and testing approaches.

Results: Study I showed that CXR screening could reduce diagnostic test needs by 16.7-52.5% compared with the baseline testing approach; indiscriminate pooled testing could further reduce test use by 42.5-58.4%, although its efficiency declined in the highest-prevalence setting; the greatest test savings (50.8-61.5%) were achieved by AI-CAD guided pooled testing. Study II showed that the newer AI-CAD software version achieved a modest but statistically significant improvement in overall diagnostic accuracy (area under the curve 0.76 vs 0.78, p=0.029). However, specificity at a 90% sensitivity threshold remained similar (45.8% vs 43.2%, p=0.148). Study III showed that AI-CAD software achieved the minimum performance Target Product Profile criteria for a TB screening test (290% sensitivity and 280% specificity) when analysing both original digital CXR images and photographs of printed CXR films, although performance was superior with the digital CXR images. Study IV showed that no CXR interpretation approach simultaneously maximised accuracy and operational efficiency, highlighting trade-offs between TB detection, diagnostic testing volume and human resource requirements. AI-CAD only reading was the only approach to achieve minimum performance Target Product Profile criteria, while double reading achieved the highest sensitivity, sequential reading substantially reduced human reading workload, and AI-CAD assisted reading provided only modest and inconsistent improvements.

Conclusions: AI-CAD software has applications beyond automated CXR interpretation and can support multiple components of TB screening programmes, including risk-stratified diagnostic testing and collaborative human/AI-CAD screening workflows. However, successful implementation requires careful calibration, ongoing performance monitoring, and deployment strategies tailored to local health system capacity. These findings help build the evidence base for effective AI-CAD software deployments during TB screening, which may contribute to improving early TB detection and optimisation of limited healthcare resources.

List of scientific papers

I. Andrew J. Codlin, Luan N. Q. Vo, Tushar Garg, Sayera Banu, Shahriar Ahmed, Stephen John, Suraj Abdulkarim, Monde Muyoyeta, Nsala Sanjase, Tom Wingfield, Vibol Iem, Bertie Squire, and Jacob Creswell. Expanding molecular diagnostic coverage for tuberculosis by combining computer-aided chest radiography and sputum specimen pooling: a modeling study from four high-burden countries. BMC Global and Public Health. 2024; 2: 52. https://doi.org/10.1186/s44…

II. Andrew J. Codlin, Luan N. Q. Vo, Thang P. Dao, Rachel J. Forse, Ha T. M. Dang, Lan H. Nguyen, Hoa B. Nguyen, Luong V. Dinh, Kristi Sidney Annerstedt, Johan Lundin, and Knut Lönnroth. Comparison of different Lunit INSIGHT CXR software versions when reading chest radiographs for tuberculosis. PLOS Digital Health. 2025; 4: e0000813. https://doi.org/10.1371/jou…

III. Andrew J. Codlin, Thang P. Dao, Binh H. Nguyen, Luan N. Q. Vo, Rachel J. Forse, Ha T. M. Dang, Lan H. Nguyen, Hoa B. Nguyen, Luong V. Dinh, Kristi Sidney Annerstedt, Johan Lundin, and Knut Lönnroth. Bridging the analog divide: a comparison of printed X-ray films and digital images when using computer-aided detection software for tuberculosis screening. BMC Global and Public Health. 2026; 4: 6. https://doi.org/10.1186/s44…

IV. Andrew J. Codlin, Thang P. Dao, Luan N. Q. Vo, Rachel J. Forse, Binh H. Nguyen, Phuong N. Trinh, Ha T. M. Dang, Lan H. Nguyen, Hoa B. Nguyen, Luong V. Dinh, Kristi Sidney Annerstedt, Johan Lundin, and Knut Lönnroth. Comparative evaluation of computer-aided detection software use cases for TB screening. [Submitted]