Abstract: Objective: To evaluate the clinical performance of using artificial intelligence (AI) based computer aided diagnostic (CAD) system on detecting pulmonary tuberculosis (TB) in primary tuberculosis control and prevention facilities.Methods: A retrospective study enrolling 396 untreated presumptive tuberculosis cases was conducted from November 2020 to April 2021 in 8 TB dispensaries in Shandong province, and the clinical performance of the AI system was analyzed by making a comparison between the results from the AI system and from local radiologists.Results: The TB detection rate of AI system on 396 presumptive cases was higher than that of local radiologists (97.8% (131/134)) vs (75.4% (101/134))(χ²=28.88, P<0.05). The consistency between the AI system and local radiologists was 86.1% (341/396), and the false positive rates of the AI system and local radiologists were 0.8% (2/260) and 6.5% (17/260), respectively. In general, the sensitivity, specificity, positive predictive value, negative predictive value, and the accuracy of the AI system were 97.8% (95%CI: 93.3%-99.5%), 99.2% (95%CI: 97.1%-99.9%), 98.5% (95%CI: 94.3%-99.9%), 98.6% (95%CI: 96.5%-99.8%) and 98.7% (95%CI: 97.0%-99.6%). The sensitivity, specificity, positive predictive value, negative predictive value, and accuracy of local radiologists were 75.4% (95%CI: 67.4%-81.9%), 93.5% (95%CI: 89.8%-96.0%), 85.6% (95%CI: 78.0%-90.9%), 88.1% (95%CI: 83.8%-91.5%) and 87.4% (95%CI: 83.7%-90.3%).Conclusion: The AI system could help local radiologists in primary facilities improve diagnostic efficiency and accuracy.

Key words: Tuberculosis, Pulmonary, Radiography, Artificial intelligence, Program evaluation