Segmentation-Robust Classification Of Breast Ultrasound Images

In recent years, breast cancer has become one of the most prevalent cancers, and it affects women's health seriously. Early detection is the most effective therapy for curing breast cancer. Breast ultrasound (BUS) imaging has become one of the most prevalent and popular approaches for the early detection of breast cancer due to it's radiation-free, non-invasive, painless and cost-effective. For improving the quality and objectiveness of the diagnosis, BUS imaging-based classification of breast tumor is applied more and more as a kind of computer-aided diagnosis technologies in clinical practice. Mainly, it processes BUS images to classify breast tumors and produce assistant information for radiologist.This paper focuses on two critical problems in classification of breast tumor: the segmentation of tumor region and the classification of tumor under the segmentation difference between tumor region and actual tumor region. All works are described as bellow:1. An improved BUS image segmentation method based on Grow Cut The complicated structure and low image quality of BUS image is the main problem that affects the location of tumor feature region. For solving this problem, an improved BUS image segmentation method based on Grow Cut is proposed. Grow Cut method is based on cellular automata (CA) approach. At first, user specifies some pixels as"seed cells", the labels of seed cells are fixed. During the procedure of segmentation, the labels of other cells will be determined in the evolution. In improved method, not only considering the gray level differences of pixels in the neighborhood, the local texture information is utilized as the restriction, and the similarity between seed pixel and current pixel is compared. In the experiments, comparing to other published approaches, it can demonstrate that the proposed method can effectively segment BUS image and find the precise position of tumor region.2. A tumor classification approach robust to the segmentation difference In published breast tumor classification methods, difference between the location of tumor feature region and actual tumor region seriously affects the classification of breast tumor. In this paper, a tumor classification approach robust to the segmentation difference is proposed. The main idea of the proposed method is that a set of classification checkpoints in tumor region is selected. Then we extract local texture information of each classification checkpoint as features and utilize a SVM-based classifier to classify the checkpoints as malignant or benign respectively. Finally, the class of the tumor is determined by voting. The experiment results demonstrate that the proposed method can effectively model and classify breast tumors and it is robust to the segmentation difference of tumor region.