A ROBUST FEATURE SELECTION APPROACH USING LOW RANK MATRICES FOR BREAST TUMORS IN ULTRASONIC IMAGES

The primary use of breast ultrasound today is to help diagnose breast abnormalities detected by a physician during a physical exam and to characterize different types of breast conditions, including both benign and malignant lesions. For this purpose, a large number of features are computed to determine the nature of a breast abnormality. This paper aims to focus on the feature selection problem for classifying benign and malignant breast tumors to assist the clinical diagnosis. We formulate the problem of choosing discriminative features as a decomposition of the computerized feature matrix into a lowrank principal matrix and a sparse error matrix. The low-rank principal matrix contains the best distinctive features for determining the benign and malignant cases whereas the sparse error matrix has the features with a less identification capability. By identifying and selecting essential features, the lowrank matrix based feature selection method can improve the classification outcomes.