부산대학교 도서관

Objectives: Tuberculosis (TB) is one of the ten leading causes of death worldwide. Microscopic examination of sputum smear is the most common detection method for the infection. However, the process of detecting acid-fast TB bacteria under a microscope often encounters low efficiency and unsatisfactory accuracy. In this study, we used the deep learning method to perform acid-fast bacteria auto-identification and explored the accuracy of the AI system and its real-world performance. Methods: Acid-fast stained smears of sputum specimens collected from the Taipei City Hospital were converted to digitized images and then used to train and test the identification system. Results: The validation of the AI system demonstrated a sensitivity and specificity of 0.91 and 0.99, respectively, while the actual sample tests yielded a sensitivity and specificity of 0.91 and 0.97. The AI system outperformed manual examination of the culture samples by 9% in sensitivity. Conclusions: Our data demonstrated that AI-based automatic identification of acid-fast bacilli is accurate and reliable. It helps improve the accuracy of the identification, reduce the burden on the medical technicians, and in tuberculosis control.