Abstract
Tuberculosis (TB) is a global health problem. It not only causes disease burden measured by combining the years of life lost due to premature death and years of life lost due to time lived in state of less than full health, but also results in economic burden. In 2015, the World Health Organization classified Thailand as one of 14 countries on all three “high burden country” (HBC) lists: tuberculosis (TB), multidrug-resistant tuberculosis (MDR-TB), and TB/HIV. Based on recent estimates made by the WHO, Thailand had 119,000 new TB patients in 2016, equivalent to an incidence rate of 172 per 100,000 population. This rate is more than 50 times higher than those in developed countries and is the cause of death for up to 12,500 people per year (1). Artificial intelligence (AI) algorithms can be trained to recognize tuberculosis-related abnormalities on chest radiographs. This study aims to develop deep learning model to classify tuberculosis related chest radiographs using data from the national tuberculosis program. Method: We collected chest radiographs from the active TB screening project of the national tuberculosis program. Then, expert chest radiologists reviewed and classified images into one of six categories according to the standard recommendation for reporting chest radiograph for TB screening from the royal college of radiology of Thailand. Next, we developed deep learning model from this dataset. Conclusion: In the first part, we collected and classified 80,086 chest radiographs according to the study protocol. In the second part, we successfully developed deep learning model to classify TB related or non-TB related chest radiographs with AUC of 0.94.
