A Superpixel-Classification-Based Method for Breast Ultrasound Images

As a special disease, breast cancer is a serious threat to females. Besides, medical image segmentation influences greatly on the computer-aided diagnosis system. In this paper, a novel method for breast ultrasound (BUS) images segmentation was introduced. The proposed segmentation algorithm based on superpixel classification, can automatically extract the part of breast tumor from the collected image. The algorithm consists of six steps, i.e., crop tumor centered image, histogram equalization, bilateral filter and pyramid mean shift filter for image preprocessing, SLIC for superpixels generation, feature extraction for each superpixel and bag-of-words model for representation, classification for initial segmentation, k-nearest neighbor (KNN) for reclassitication and postprocessing. A classification scheme based on multilayer perceptron (MLP) is used to predict superpixels. Experiments are carried out and the results have indicated that our method is efficient to segment the BUS images. It brings a better performance in some aspects compared with four methods.