Study On Computer-aided Diagnosis Of Mammograms

Breast cancer is one of the most dangerous malignant tumors of middle-aged and older women in the world. In the latest twenty years, the incidence and mortality of breast cancer in China increase rapidly. And mammograhy is the most reliable method to detect breast cancer. In order to improve the detection precision, computer-aided diagnosis (CAD) techniques of the mammograms have been widely applied. The purpose of this dissertation is to find out a set of quick, effective and accurate CAD algorithms, for the detection of the focuses of breast cancer, including the masses and the microcalcifications (MCs). These algorithms could be introduced in five parts, just as follows:The first part is to extract the breast region as the detection area. Iterative thresholding is used to separate the background region, least square estimation (LSE) is utilized to detect the orientation of the horizontal frame, and then elastic thread technique is applied to cut the conglutination of the breast and the frame. At last, watershed method is used to fine the breast boundary.The second part is to segment the pectoral muscle from the breast, reducing the detection area. A series of ROI (region of interest) with different sizes are applied upon the region near the pectoral muscle, and the optimal threshold of the pectoral muscle could be computed with the corresponding optimal threshold curve and mean square deviation (MSD) curve. Then zonal Hough transform is used to approach the edge of the pectoral muscle, and elastic thread and polygon approaching techniques are finally carried out to fine its boundary.The third part is to detect the MCs and extract their regions. In the second and third layers of the high-frequency domain of the mammograms, which is decomposed with discreate wavelet transform (DWT), thresholding with hysteresis is carried out to locate all the suspicious MCs. And then, the region expanding algorithm based on filling dilation is used to extract all the objective regions. During the whole detection process, ANFIS is applied to adjust the location and segmentation criterion according to the background features, making the detection more adaptive.The fourth part is to locate the masses and segment their regions. At first, two mass models are proposed to represent all kinds of masses. Then, thosemasses matching Model II could be located with iterative thresholding, while those masses matching Model I could be located by detecting the regions with low modulus in the high-frequency wavelet domain, and those masses on the edge of the denser tissue could be located with adaptive thresholding based on local gradient enhancement. At last, filling dilation restricted with Canny edge detector is utilized to extract the mass regions. The segmentation process is always controlled with ANFIS, and energy field method is applied to adjust the region extraction when the masses are close to each other.The fifth part is to classify all the suspicious MCs and masses extracted above. The most appropriate features of these two focuses are firstly selected, and the extration methods of them are discussed. And then MLP, CMAC and ANFIS are simultaneously used for classification of MCs and masses. By comparing the classification effects of these three neural networks, a conclusion could be drawn that MLP performs best as the classifier of the focuses in the mammograms. Finally, by combining the segmentation and classification result, higher detection precision and stronger robustness than the conventional methods is gained.