セマンティックセグメンテーションを用いた超音波画像からの骨表面抽出

抄録

The purpose of this study was to extract the bone surface from ultrasound images by a technique called semantic segmentation, which is a type of deep learning, and to verify the effectiveness of its accuracy. A neural network model called U-Net was used for learning, and the data used for learning was an echo movie of the bone surface divided into 300 images. These 300 ultrasound images were divided into training data and test data. Although the accuracy can be expressed numerically as a similarity rate when learning, we cannot achieve 100% accuracy by machine learning at present. Therefore, how accurate is sufficient depends on the field and application. Therefore, in this study, we investigated how the learning accuracy can be evaluated after extracting the bone surface by semantic segmentation. As a result of learning, extraction was successful with an accuracy of 67 %. In addition, in order to confirm its effectiveness, it was verified whether or not the center point at a certain X coordinate of the bone surface portion in the original ultrasound image was output in the predicted image. As a result, we succeeded in confirming the effectiveness of the accuracy by confirming that it was output.