2021 年 39 巻 5 号 p. 229-242
We report a chest X-ray (CXR) anomaly detection method based on models of normal anatomical structures (ASs). Conventional computer-aided diagnosis (CAD) methods for CXR involve machine learning of predetermined target lesions; thus, these methods are unable to detect and diagnose unknown lesions. As such, we are building a new CAD system to model normal ASs and detect anomalies based on changes in ASs due to disease. The method consists of AS segmentation by U-Net, index calculation from the segmented AS region, and anomaly detection compared with the distribution of normal indices. The position, size, and continuity of the segmented ASs were used as indices. Six structures near the central shadow, which tend to be easily overlooked on CXR, were used as the target ASs. For the anomaly detection evaluation, 240 normal cases and 255 abnormal cases were used, and an average sensitivity of 0.774 and specificity of 0.827 for each AS were obtained. However, abnormalities that do not affect the currently used indices cannot be detected. In the future, we plan to apply this method to other ASs and the entire lung and adopt other indices, such as features that directly reflect image patterns.
