Texture Feature Analysis of Breast Lesions in Automated 3D Breast Ultrasound

Abstract: This thesis investigated a variety of texture features performances  on classifying and detecting breast lesions in automated  3D breast ultrasound (ABUS) images with computer-aided diagnosis and detection  system. Regions detected  by the computer-aided detection  system could be categorized into benign and malignant classes, which are supposed to have different texture features. After normalization and segmentation on the original 3D ultrasound breast images automatically, we implemented four texture feature extraction  algorithms on the detected  targets. The proposed  four algorithms are based on 3-dimensional gray level co-occurrence matrix (3-D GLCM), local binary pattern  (LBP), Haar-Like and regional zernike moment (RZM) separately. Three major experiments  were carried out on a set of ABUS images. In experiment  one, we focused on distinguishing malignant lesions (165 samples) from benign lesions (258 samples). In experiment  two, we added a number of normal cases (150 samples) to the dataset, by grouping them with benign lesions against malignant lesions and by isolating them from benign and malignant lesions. In experiment  three, we tested  texture features ability on reducing false positives in the existing computer-aided detection  system. In this step, only normal cases (5263 samples) and malignant lesions (165 samples)  were examined. To estimate the discrimination power of different texture features, Support VectorMachine (SVM) and AdaBoost classifiers were adopted  in corporation withleave-one-patient-out and 10-fold cross validation schemes respectively. The areaunder the receiver operator characteristic (ROC) curve (AUC, also known as Az)values were analyzed corresponding  to each texture feature extraction  method. TheAz values computed  in experiment  one are compared  as follows: Haar-Like feature'sperformance  outweighs others'  with the Az value of 0.86, followed by LBP (0.84),RZM(0.81) and 3-D GLCM (0.75). With respect  to the results from experiment  two,the Az value of grouping normal cases with benign lesions against malignant lesions isbetter  than separating them from benign and malignant lesions, in general. Regardingthe outcome  from experiment  three, the Az value was increased from 0.79 to 0.82after adding LBP and Haralick features to the existing computer-aided detectionsystem. Based on the overall results, we concluded that texture features are useful on classifying benign and malignant lesions in ABUS images and they can improve the performance  of the existing computer-aided detection  system on detecting breast cancers.