Font Size: a A A

Research On Key Tchnologies For Computer-Aided Diagnosis Of Breast Cancer Based On Ultrasound Images

Posted on:2009-07-27Degree:DoctorType:Dissertation
Country:ChinaCandidate:X F LiFull Text:PDF
GTID:1118360278961980Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
Currently, the most effective method for early detection of breast cancer is mammography. But the high number of false positives in mammography causes a lot of unnecessary biopsies, at the same time use of mammography for high-density breast examination will lead to serious missing diagnosis phenomenon. Sonography is an important adjunct to mammography in breast cancer detection, and using ultrasound imaging to detect breast cancer has many superiorities that mammographic imaging doesn't has, such as differentiation of cysts from solid tumors or detecting lesions in dense breast. However, reading breast ultrasound image is a demanding job for radiologists. The readings depend on abundant experience, and the diagnosis is often a certain degree of subjectivity. Therefore, development of a reliable computer-aided diagnosis system to reduce the subjectivity of the diagnosis, depress medical waste and reduce the missing diagnosis rate is of great significance. In order to study an effective computer-aided diagnosis system of breast ultrasound images, this dissertation carries out the following research subjects on breast ultrasound image pre-processing, tumor contour extraction, tumor features extraction, breast tumors classification and medical ultrasound image lossless compression:In order to reduce the interferences of speckle noise and tissue texture on tumor contour extracting, a mode filter based anisotropic ultrasound images diffusion algorithm is proposed. In this algorithm, a mode filter is embedded into the function which is used to control the speed of the mean curvature motion (to control the smoothing degree of image) to eliminate the influence on image gradient computing by speckle noise. Mode filter has the properties of being effectively denoising speckle noise and being able to enhance image edge, the diffusion algorithm proposed in this dissertation can denoise the speckle and smooth redundant texture effectively, while keep tumor'edge.Image denoising is an important subject in the field of image processing, and also a necessary part of interest region extraction and target recognition. This dissertation proposes an adaptive total variation denoising algorithm. The algorithm can decide the smoothing degree of each pixel in the image based on the edge confidence adaptively. For the regions around the edges, this algorithm has a smaller smoothing effect on it, otherwise bigger. Mathematical analysis and experimental results show that this adaptive total variation denoising algorithm can solve the problem that Euler-Lagrange equation of the traditional total variation is not well posed, improve the deficiencies of the slow denosing speed, and reduce the step effect and noise leaving out resulting from using the traditional total variation.The geometric shapes of the tumors are important information for diagnosis, however, their features are calculated based on tumor's contours. To extract tumor contour, a contour extraction method with geometric deformable contour model based on edge and region statistical features is proposed. When deciding pixel's speed (this pixel sits on evolution front), its edge confidence is calculated, and the texture features are extracted at the same time. Then the texture features are inputted into the trained SVM. The SVM will produce the signed distance between the texture features and the optimal separating hyperplane. Then a mapping function is employed to map the signed distance into [0, 1], which represents the probability of the pixel belonging to normal tissue. Finally, the segmentation method decides the evolution speed according to the pixel's edge confidence, the probability belonging to normal tissue and the pixel's curvature. In order to extract the tumor's contour automatically, a rough lesion contour extraction method is designed based on tumor ultrasound images'features. The extracted rough contour will be used to initialize the level set function proposed above.Breast ultrasound image tumor features extraction and classification are studied, and following the results of the studies, a computer aided diagnosis (CAD) system based on SVM is constructed. In this CAD system, 34 typical features belonging to texture features, tumor morphologic features, echogenicity, edge sharpness respectively are extracted as the tumor classification criterions. To select the parameters in the classifier of Gauss-SVM, the grid-search using cross-validation is adopted. The indexes such as classification accuracy, sensitivity, specificity, positive predictive, negative predictive, ROC curve are used to evaluate the breast CAD system proposed in this dissertation.To meet the need of images transfer for remote CAD, according to the characteristics of the ultrasound image's weak edge, an ultrasound image lossless compression scheme based on least-square (LS) and Huffman code of most-likely magnitude (LS-MLMH) is proposed. When a pixel is predicted, a set of prestored coefficients are used to predict the pixel and generate the prediction error. If the prediction error is larger than a pre-selected threshold, the pixel is considered near the edge, and the pixel will be repredicted by least-square, at the same time the prestored prediction coefficients will be updated; if the prediction error is less than the threshold, the pixel is considered far away from the edge, the prestored prediction coefficients can be used to predict the pixel. This switching strategy utilizes the LS's inherent edge-directed property to overcome the shortcoming that can not give accurate prediction caused by ultrasound image weak edge, at the same time, by the premise of guarantee of prediction quality, reduces the amount of computation. To improve the compression performance furtherly, Huffman of most-likely magnitudes is used to code prediction residual image in the LS-MLMH. The results show that LS-MLMH compression scheme is superior to the lossless JPEG, JPEG-LS and CALIC compression method.
Keywords/Search Tags:Computer-aided diagnosis, Breast ultrasound image, Ultrasound image denosing, Tumor contour extraction, Support vector machine
PDF Full Text Request
Related items