Automated segmentation of breast lesions in ultrasound images

Breast cancer is one of the leading causes of death in women. As a convenient and safe diagnosis method, ultrasound is most commonly used second to mammography for early detection and diagnosis of breast cancer. Here we proposed an automatic method to segment lesions in ultrasound images. The images are first filtered with anisotropic diffusion algorithm to remove speckle noise. The edge is enhanced to emphasize the lesion regions. Normalized cut is a graph theoretic that admits combination of different features for image segmentation, and has been successfully used in object parsing and grouping. In this paper we combine normalized cut with region merging method for the segmentation. The merging criteria are derived from the empirical rules used by radiologists when they interpret breast images. In the performance evaluation, we compared the computer-detected lesion boundaries with manually delineated borders. The experimental results show that the algorithm has efficient and robust performance for different kinds of lesions.