Application of Artificial Intelligence (AI) in Chest Radiography Interpretation to Enhance the Efficiency of Active Tuberculosis Screening in Community Settings

Abstract

Active Tuberculosis (TB) Screening is crucial for the early detection of TB cases. Chest radiography is the primary tool used to screen individuals suspected of TB and to identify those requiring sputum testing for diagnosis. However, there are limitations to community-based active screening due to the limited availability of physicians to interpret chest radiographs in the field. This limitation can lead to delays and may result in individuals with chest X-rays suspected of TB not receiving sputum tests, which could lead to the spread of TB within families and communities. Therefore, applying artificial intelligence (AI) to interpret chest radiographs may help enhance the efficiency of active TB screening in community settings. The objective of this study is to evaluate the outcome of applying Artificial Intelligence in interpreting chest radiography to enhance the efficiency of active tuberculosis screening in community settings. This retrospective study involved data from 1,968 volunteers aged 15 years and older in three communities. Chest radiography interpretations by AI and physicians were compared based on sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), and the area under the ROC curve (AUC). The Composite Reference Standard (CRS) was used as the diagnostic reference, and McNemar’s test was used to compare the results between AI and physicians. The result showed that AI had a sensitivity of 71.4%, specificity of 96.0%, accuracy of 95.8%, PPV of 11.2%, and NPV of 99.8%, with an AUC of 0.9 (95% CI: 0.80-0.99). Physicians showed a sensitivity of 71.4%, specificity of 90.0%, accuracy of 89.8%, PPV of 4.8%, and NPV of 99.8%. Statistical comparison showed no significant difference in sensitivity between AI and physicians (p = 1.00), but AI had significantly higher specificity than physicians (p < 0.0001). In conclusion, AI improves the efficiency of TB screening in communities by quickly and accurately identifying normal chest radiographs, thereby reducing the workload for physicians. This enables physicians to focus on reviewing abnormal or high-risk cases more effectively. Furthermore, in areas without physician availability, AI can serve as a decision-support tool for timely sputum collection and diagnostic testing. As a result, TB cases can be detected and treated more quickly, reducing transmission within the community.