Research On The Location And Classification Of Breast Tumor In Ultrasound Images

In recent years, breast cancer has become the most prevalent carcinoma in women, and it affected women's health seriously. Early detection is the most important therapy for curing breast cancer. Due to its effectively detective, radiation-free and cost-effective, breast ultrasound (BUS) imaging has become one of the most popular approach for the detection of breast cancer.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. Especially, automatic classification of breast tumor is very beneficial to clinical diagnosis due to its objective and operation-effective. The implementation of automatic classification of breast tumor has become a pressing need in clinical pratice. However, in published studies, automatic classification of breast tumor has still not been accomplished.This paper focused on two critical problems in automatic classification of breast tumor: the location of tumor feature region and the classification of tumor under the location difference between tumor feature region and actual tumor region. In this paper, three tumor location and classification approaches are proposed.1. Local texture-based location of approximate tumor regionThe 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, in this paper, an approximate tumor region location method is proposed. Utilizing this method, an approximate tumor region close to actual tumor region can be generated. Based on this result, good inputs can be produced for following steps. In this method, utilizing local texture and statistical learning theory, the BUS segmentation is transformed as a classification problem. In this method, BUS image is enhanced based on fuzzy logic at first. And then, the image is divided into lattices with same size, and local texture in each lattice is extracted as feature. A support vector machine (SVM) classifier is established for classifying the lattices into"normal tissue"or"tumor"class. Finally, medical background knowledge is utilized for selecting a suitable approximate tumor region in the binarized image. The result of the proposed method can be utilized for producing three kinds of inputs for following steps: (1) region of interest for locating the precise tumor region; (2) initial condition of other tumor feture region location approaches; (3) tumor feature region for tumor classification. The experimental results demonstrate that the approximate tumor regions generated by the proposed method are close to the corresponding actual tumor regions, and the inputs generated based on this result can be benefical to following steps.2. Probability density difference and local boundary information-based precise tumor region locationIn this paper, an active contour model based on probability density difference and local boundary information is proposed for locating the precise position of tumor region. This method has two properties. The first one is that the proposed model utilized the differences between actual probability densities and estimated probability densities of intensities for establishing the segmentation model. In this method, the estimated probability densities are modeled by utilizing the background knowledge of ultrasound imaging, and they can model the distributions of the intensities of different regions effectively. With this property, by minimizing the differences between actual and estimated probability densities, the generated regions will have suitable distributions of intensities, and good segmentation results can be obtained. Secondly, the proposed method utilizes the local boundary response in the de-nosied image for modeling the local information while utilizes the global information in the original BUS image. This approach considers both of the ideas of utilizing or removing speckle noise in published studies. In this method, the evolution of the curve is implemented by level set method and finite difference method. 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.3. Local texture-based classification of breast tumorIn published breast tumor classification methods, there is always difference between the location of tumor feature region and actual tumor region. This kind of difference seriously affects the classification of breast tumor. In this paper, a tumor classification approach robust to the difference is proposed. The main idea of the proposed method is selecting a set of classification checkpoints in tumor feature region. For each classification checkpoint, local texture information is extracted as features and a SVM-based classifier classifies the classification checkpoints as malignant or benign respectively. Finally, the class of the tumor is determined by voting. The expiermental results demonstrate that the proposed method can effectively model and classify breast tumors and it is robust to the location difference of tumor feature region.