Support Vector Machine And Its Application In MRI Brain Image Segmentation

Segmentation of brain tissues is very important in medical image analysis. Many existing image segmentation methods are based on the traditional statistical theory, which is based on the asymptotic theory when the number of the sample tends to infinity. But for high-dimensional features, it is difficult to obtain good results, so the segmentation results are not very satisfactory. Support Vector Machine(SVM) is considered as a good candidate because of its good generalization performance, especially for dataset with small number of samples in high dimensional feature space. This thesis investigates the segmentation of magnetic resonance brain tissues image based on SVM.The main work in the dissertation can be summarized as follow:In this paper, the segmentation algorithm based on support vector machine is studied, and some experiment in different kernel functions and parameters are done. The comparative experiments are made using the different number of training samples, and it confirms that SVM method holds the better classification ability in the small-sample.Because Feature extracting of image has an direct effect on the accuracy of the segmentation result, two groups of image features, the combined features of textures and gray features and gray level features based on window region, are studied.6textures statistics features based on the co-occurrence matrix of gray level(namely, energy, contrast, correlation, entropy, local calm, and variance) and3gray features(namely, the pixel intensity, the median filter intensity and average intensity of window size of each pixel) are chosen as the image combined features both texture and gray in the brain tissues classification. The gray level values of the pixels and pixel intensities in the neighborhood are used for gray level features extraction.When SVM(Support Vector Machine) is used to perform segmentation of brain image in MRI in practice, it needs manual work to select training sample according to reference image segmentation. Regarding to this issue, a new segmentation method that combines the fuzzy C-means clustering method(FCM)with the SVM brain image is proposed in this essay. Firstly, the way of FCM segments the original graph, and selects the wider range of pixel of the membership degree as training sample in the results. Moreover, using the model of training sample predicts all of pixels in the image. The experiment turns out that this method has a higher accuracy and a faster speed of segmentation when compared with the selection of artificial training sample.