På svenska

Presentation Information     2013-06-11 (09:15)   •  The seminar room at Vi2

Speaker Haixia Liu
Type Master thesis presentation
Title Texture Feature Analysis of Breast Lesions in Automated 3D Breast Ultrasound
Abstract This thesis investigated a variety of texture features performances on classifying breast lesions in automated 3D breast ultrasound (ABUS) images in computer-aided detection system. 3D breast ultrasound volumes could be categorized into normal, benign and malignant classes, which are supposed to have different texture features. Step one, we focused on distinguishing malignant cases (165 cases) from benign cases (258 cases). Step two, we added a number of normal cases (150 cases) to the experiments, by grouping them with benign cases against malignant cases and by isolating them from benign and malignant cases. Step three, we tested texture features ability on reducing false positives in the existing detection system. In this stage, only normal cases (5263 cases and malignant (165 cases) cases were examined. After normalization and segmentation on the original 3D ultrasound breast lesion images, we implemented four texture feature extraction algorithms, which are gray level co-occurrence matirx (GLCM), local binary pattern (LBP), Haar-Like and regional zernike moment (RZM). To evaluate the discrimination power of different features, Support Vector Machine (SVM) and AdaBoost classifiers were adopted. Both Leave-one-patient-out cross validation and 10-fold cross validation schemes were tried during the training and testing processes. The area under the receiver operator characteristic (ROC) curve (Az) was analyzed corresponding to each texture feature extraction method. The Az values computed at step one are compared as follows: Haar-Like feature's performance outweights others' with the Az value of 0.86, followed by LBP (0.84), RZM(0.81) and GLCM (0.75). With respect to the results of step two, the Az value of grouping normal cases with benign cases against malignant cases is better than separating normal cases from benign and malignant cases, in general. Regarding the experiment outcome from step three, the Az value was increased from 0.58 to 0.82 after adding LBP features and Haralick features to the non-texture feature based CAD system. Based on the overall results, we concluded that texture features can improve the ABUS performance on detecting breast cancer.