Mammographic Mass Detection by Different Sized Windows in Optical Density Images

• Main Authors: Chia-JuiChuang, 莊佳叡
• Other Authors: Shen-Chuan Tai
• Format: Others
• Language: en_US
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/40334801045797861541

碩士 === 國立成功大學 === 電腦與通信工程研究所 === 101 === The mass detection of mammograms is a difficult and exhausting work owing to complicated morphological characteristics and ambiguous margins. Therefore, the Computer-Aided Detection (CAD) systems are developed to resolve the situation, which provide opinions to radiologists. The system acts as a second reader while the radiologists make the final judgment. In the thesis, some features with different sized windows are applied to improve the performance of mass detection. The proposed algorithm consists of four steps. First, the pectoral muscle is removed to focus on the breast region. Because texture characteristics of masses are similar to the pectoral muscle, false positives can be reduced when the pectoral muscle is eliminated. Second, a hierarchical template matching method is used to find the suspicious areas. Adaptive square regions of interest (ROIs) are extracted according to sizes of suspicious areas. Then, in total of 86 descriptors including Weber local descriptors, local binary patterns and Haralick's texture features are introduced to describe the characteristics of each ROI after optical density transformation. Finally, two kinds of classifiers, SVM and linear discriminant analysis (LDA) based on stepwise method, are utilized to determine the model and discriminant function to distinguish between masses and normal regions. In the proposed experiment, 92 cases from the Digital Database for Screening Mammography are conducted. The sensitivity is $96.2\%$ with 4.4 false positives per image and the area under ROC curve is $0.966pm0.006$. The results show that the proposed algorithm of mass detection achieves satisfactory performance and preferable compromises between sensitivity and false positives per image.