We report a segmentation process for multiple anatomical structures in chest X-ray photographs (CXPs) by deep neural networks and the corresponding evaluation results. The segmentation process is a key element of the computer-aided diagnosis (CAD) system, based on changes in appearance of anatomical structures in CXPs. Mainstream, conventional CXP-CAD technologies detect lesions that are machine-learned in advance. However, in a CXP, multiple anatomical structures are depicted in an overlapping manner. Furthermore, if a lesion overlaps with those anatomical structures, it becomes difficult to detect the lesion using conventional methods. Therefore, a new type of CAD system is needed. We use U-Net for the segmentation process. Segmentation targets comprise nine small regions including anatomical structures and boundary lines between anatomical structures. For experimental data assessment, 684 normal cases and 61 abnormal cases were used. For normal cases, Dice coefficients for various structures ranged from 0.653 to 0.919 when using U-Net. For abnormal cases, qualitative evaluation suggested the possibility of anomaly detection. In the future, we will develop anomaly detection for anatomical structures and estimation of abnormal findings for the entire CXP.
