Population-scale cross-sectional observational study for AI-powered TB screening on one million CXRs

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Public Health and Infectious Disease Control

Background:

• Traditional tuberculosis (TB) screening relies on manual chest X-ray (CXR) interpretation by radiologists, a process that is inefficient, prone to errors, and strained by workforce limitations.
• The need for faster, more accurate, and scalable TB screening methods is critical for global health initiatives.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) model, AIRIS-TB (AI Radiology In Screening TB), for automating the reporting of normal chest X-rays (CXRs) in tuberculosis screening.
• To assess the performance of AIRIS-TB compared to human radiologists in identifying TB and reducing workload in high-volume screening settings.

Main Methods:

• The AIRIS-TB AI model was developed to automate the reporting of CXRs without any findings.
• The model's performance was evaluated on a dataset exceeding one million CXRs.
• Key performance metrics included Area Under the Curve (AUC) and False Negative Rate (FNR), with specific attention to TB-related FNR (TB-FNR).

Main Results:

• AIRIS-TB achieved a high AUC of 98.51% and an overall FNR of 1.57%, outperforming the average radiologist's FNR of 1.85%.
• Crucially, the AI model maintained a 0% TB-FNR, ensuring no TB cases were missed.
• The model demonstrated the potential to automate up to 80% of routine CXR reporting by deferring only cases with findings to radiologists.

Conclusions:

• The large-scale validation confirms AIRIS-TB's safety and efficiency for high-volume tuberculosis screening programs.
• AIRIS-TB effectively reduces radiologist workload without compromising diagnostic accuracy for TB detection.
• Performance remained consistent across diverse demographic subgroups (age, sex, HIV status, region), indicating broad applicability.