A Novel Lesion Segmentation Method based on Breast Ultrasound Images
Xiaoyan Shen, Jiaxin Liu, Hong Li, Hang Sun, He Ma · 2019
Lesion segmentation is a critical step in computer-aided diagnostic(CAD) systems based on breast cancer imaging. Accurate segmentation directly affects the final determination of the nature of lesion. However, due to the low quality of ultrasound (US) images, lesion segmentation based on US images of breast is challenging. This paper presents an improved marker watershed algorithm for lesion segmentation of breast US images. It uses the efficient and fast curvature filtering(CF) and Gaussian enhancement method to pre-process the image, then creates the segmentation function through Newton filter based on computing gradient of the image, then compares it with the labeling function obtained by binarizing the image to get the most similar parts of them.The intersection is used as the input of the marked watershed(mw) algorithm, and then the candidate boundary is obtained. Finally, the final boundary is determined by maximizing the average radial derivative(ARD) function. The novel method was tested with 400 sets of US images and quantified by using both area and contour error metrics. The result shows that our method can extract the relatively accurate lesion region effectively and efficiently and solves the problem of over-segmentation of the watershed algorithm to some extent. Especially for the segmentation of tumors with internal calcification points or blurred boundaries, it shows better performance.