A graph-based segmentation method for breast tumors in ultrasound images

This paper introduces a graph-based image segmentation method for detecting breast tumors in ultrasound images. The proposed segmentation algorithm based on the minimum spanning trees in a graph generated from an image, can automatically detect tumor regions and segment lesions in ultrasound images. The algorithm for segmenting breast ultrasound images consists of 3 steps, i.e. the nonlinear coherent diffusion model for speckle reduction, the graph construction for mapping the image to a graph, and the mergence of smaller regions. A pairwise region comparison predicate comparing the inter-component differences with the within component differences, is used to determine whether or not two regions should be merged. Experimental results have shown that the proposed segmentation algorithm is simply structured, robust to noises, highly efficient and much flexible in comparison with Fuzzy C means clustering. It can successfully detect tumors and extract lesions in ultrasound images more accurately. We hope that our method could be useful in various medical practices, providing an alternative way for ultrasound image analysis.