Segmentation and classification of benign and malignant breast tumors via texture characterization from ultrasound images

Abstract

The present paper deals with breast tumors classification from ultrasound images. The proposed procedure consists of four steps, namely preprocessing, segmentation, feature extraction and classification. To improve the quality of ultrasound images, the preprocessing step consists of anisotropic filtering and histogram equalization that are performed on the original images. The segmentation is performed on the preprocessed images using the Level Set method that allows to extract the region of interest (ROI) and to reduce its size at the same time. Two feature extraction methods are used in this work namely, the local binary pattern (LBP) method and the histogram of oriented gradients (HOG) method. The two methods (LBP and HOG) are techniques of textures analysis and allow to characterize the ROI. The extracted feature sets constitute the inputs for three classifiers namely, support vector machines (SVM), k-nearest neighbors (KNN) and decision trees (DT). In this work, the best results are obtained by the concatenation of the two feature vectors namely LBP and HOG associated to the SVM classifier. This allows to achieve an accuracy of 96%, a sensitivity of 97% and a specificity of 94%.