[1H4-J-13] AI application: medicine

[1H4-J-13-01] Chest X-ray anomaly detection based on normal models of anatomical structures segmented by U-Net

〇Kenji Kondo^1,2, Jun Ozawa¹, Masaki Kiyono^2,3, Shinichi Fujimoto³, Masato Tanaka³, Toshiki Adachi³, Harumi Ito³, Hirohiko Kimura³ (1. Advanced Industrial Science and Technology, 2. Panasonic Corporation, 3. University of Fukui)

Keywords:anomaly detection, chest X-ray pictures, anatomical structure, normal model

We report a chest X-ray anomaly detection method based on normal models of anatomical structures, and the corresponding evaluation results. The method consists of segmentation process for anatomical structures and anomaly detection process for the segmented regions. We use U-Net for segmentation and Hotelling’s theory for anomaly detection. Targets for segmentation and anomaly detection are nine structures including anatomical structures and boundary lines between anatomical structures. For experimental data assessment, 684 normal cases and 13 abnormal cases were used. Positions and sizes of segmented regions were used as indices for anomaly detection. When cutoff values for anomaly detection are decided by maximizing Youden indices, the sensitivities were all 1.0 and specificities ranged from 0.80 to 1.0 for anatomical structures.