Medical Image Processing Based On Support Vector Machines

SVM (Support Vector Machines) is a new pattern recognition technology which was proposed by Vapnik and his research team. Its theory is based on VC Dimension (Vapnik-Chervonenkis Dimension) theory and SRM (Structural Risk Minimization), and have better generation capacity in the small sample, nonlinear and high-dimensional characteristic space. Because of its favorable theories and perfect experiment results, SVM attracts more and more attention of researchers.Medical image segmentation and registration are two important and difficult problems in medical image processing. Using SVM for medical image processing has great significance both in theory and practical application. Considering the characteristics of medical images, this paper does several researches as follows:(1) Medical Image Segmentation based on Least Square Support Vector Machines Considering the advantages of the good generalization for Least Square Support VectorMachines (LS-SVM) in the small-sample, and the disperse feature of the segmented objects in medical images, LS-SVM is used to perform medical image segmentation. Taking the magnetic resonance images (MRI) for experiment, several key problems such as training set selection, texture extraction, the influence of Kernel function and its parameters on the segmentation performance are discussed, and compared with Fuzzy C-Means (FCM) methods. The experimental results show that, for segment targets with blurry edges, intensity non-uniformity and discontinuity (such as medical images), LS-SVM approach is a good choice, which has better classification accuracy and less running time than FCM.(2) Medical Image Registration based on Least Square Support Vector MachinesThis paper, proposes an image registration method based on LS-SVM. Firstly, select control points manually, secondly, estimate transformation model using the regression performance of LS-SVM, thirdly, use the model register the images which have geometric deformation, finally, consider interpolation problems. Taking the anamorphic magnetic resonance images for experiment, and comparing with the maximization of mutual information methods. The experimental results show that, the LS-SVM method can remove geometric distortion and has less running time than the maximization of mutual information method, LS-SVM is an effective image registration method.