碩士 / 國立中正大學 / 電機工程研究所 / 93 / Abstract
The objective of this thesis is to study the effects of different texture features, when combined with the Stepwise Totally eighty-five ultrasound images were used in this study. Among them, twenty-seven images contain benign tumors and the other fifty-eight images contain malignant tumors. The largest region inside each tumor was segmented with the assistance of an experienced physician and serves as the region of interest (ROI). Different sets of texture features were calculated from the ROI’s including the Laws textures, co-occurrence features, and wavelet texture energy features. The stepwise logistic regression (SLR) method then follows to find best features to differentiate benign and malignant tumors. To raise the generalization capacity of this method, the leave one out training and testing scheme was adopted in the experiments.
The results show that the co-occurrence features and the wavelet texture energy, when combined with the SLR classifier, provide effective texture features for the diagnosis of breast tumors. Five features were selected by the SLR classifier to achieve admirable results. Three of the features were selected from the Laws texture energy, including the mean of the ROI with L5’XS5 kernel, mean of the ROI with L5’XR5 kernel, variance of the ROI with L5’XE5 kernel, The other two features were co-occurrence features, including the information measure of correlation of the ROI, the angular second momentum of the HL band, Using these features and SLR classifier results in a high accuracy of 95.3 % and an pretty high Az value of 0.97. This result demonstrates the superiority of this system in the differentiation of benign and malignant tumors.
| Identifer | oai:union.ndltd.org:TW/093CCU00442070 |
| Date | January 2005 |
| Creators | Shih, P.-Yu., 施泰宇 |
| Contributors | Yu, S.-N, 余松年 |
| Source Sets | National Digital Library of Theses and Dissertations in Taiwan |
| Language | zh-TW |
| Detected Language | English |
| Type | 學位論文 ; thesis |
| Format | 75 |
Page generated in 0.0013 seconds