The study assembled 1,097 chest radiographs from five institutions and 3,117 interpretations by six NIOSH-certified B readers. Participants were at least 15 years old, while people with HIV or opportunistic infection were excluded. Readings were compared with sputum smear, culture or molecular reference tests. Sixty-nine per cent of interpretations were abnormal; 87% of these were TB cases, while 83% of unremarkable readings were non-TB.
Key findings
- Agreement was κ=0.83 for all findings, κ=0.67 for TB-consistent findings and κ=0.76 for active TB. Across B readers, sensitivity ranged from 77.2% to 91.1%, specificity from 87.4% to 98.6% and accuracy from 84.1% to 90.1%.
Why this matters globally
Radiology AI often loses performance across countries and populations. A Thai expert-labelled dataset anchored to microbiology can test transportability and bias in global models. The dataset itself does not prove that any AI system is ready for clinical use; each model requires separate validation.
Thai researcher contribution
Thai teams across Prince of Songkla University, Mahasarakham University, regional hospitals, the Ministry of Public Health, the Central Chest Institute, Rangsit University and Mahidol University curated images and produced a multicentre expert reference.
Limitations to consider
Children, people with HIV and opportunistic infections were excluded, and data came from high-incidence Thai settings. B readers were not compared with general readers. Case composition was designed for evaluation, so prevalence and real-world predictive values cannot be inferred directly.