Development of an artificial intelligence software to detect interstitial pneumonias

Abstract

Background: Interstitial pneumonias (IPs) are commonly intractable diseases with poor prognosis. Recently, antifibrotic agents are developed against idiopathic pulmonary fibrosis (IPF) which is the most common phenotype of IPs and early use is recommended to restrain deterioration of the respiratory function and prevent the acute exacerbation. However, it is challenging to find patients with IPs in their early stage with the chest X-ray examination. In this study, we tried developing an artificial intelligence (AI) software to help doctors detect interstitial shadows. Method: We used 1159 chest X-rays of patients who visited Sapporo Medical University Hospital from Jan 1, 2003 to Nov 31, 2018 in this study. These included 653 X-rays of 263 patients who had IPs and 506 X-rays which did not show any interstitial shadows. We let AI learn 920 X-rays and other 239 X-rays were used to valid the detection capability of the AI software. AI expresses the “likelihood of interstitial shadow” with numerical values. We also compared this capability of the software with that of human physicians. Result: The AUC of ROC curve for the detection capability of AI was 0.979. After setting the cutoff of the likelihood value to 0.267, the sensitivity and specificity were 0.896 and 1.000, respectively, which exceeded those made by human doctors (specialists of the pulmonary medicine, non-pulmonologist doctors and radiologists). Conclusion: We developed the AI software to detect interstitial shadows in the chest X-ray. The detection capability of this software exceeds that of human doctors.