With the improvement of computer imaging technology, medical images play anincreasingly important role in disease discrimination. Recent years, research onmedical images especially on brain medical image has become a new discipline. Theidentification and extraction of brain tumor in MR images is very difficulty in themedical image processing. In this paper, in order to improve the efficiency andaccuracy of disease diagnosis, we use improved SVM method to detect and extracttumor tissue automatically.SVM method was one of the machine learning methods developed on statisticallearning and structural risk minimization theory proposed by Vapnik etal.It has beenwidely used in space issues of nonlinear, high-dimensional, small sample size andlocal minimum point data. CV model segmentation method is a simplified model ofMS method proposed by Tony Chan and Luminita Vese.It was one of the activecontour models using segmentation technique but did not depend on the imagegradient edge detection. Subject to the restrictions of the complexity of brain images,use only SVM or CV model can not complete brain tumor segmentation accuratelyand effectively, so we proposed the CV-SVM method. CV-SVM method combinedCV model and SVM effectively and could classify a brain image to four parts(background, cerebrospinal fluid, white matter and gray matter). It could also extracttumor Organization and image features in similar gray areas. This all make thediagnosis of brain tumor easier and have medical application value.Thesis research mainly contains:(1) To counteract the problems like the eigenvectors complexity and redundancyin using SVM method, the PCA method was used to reduce the dimension of featurevectors,so that to decrease the total amount of data and to improve the image segmentation efficiency.(2)Using Gaussian radial basis kernel function to deal with classified data. To getmore appropriate parameters of the kernel, a large number of brain MR images wereused to get better classification results.(3)Proposed a new segmentation method called CV-SVM, it combined CVmodels and SVM method. CV-SVM overcame the problem of CV model that thecurve can not reach the border in bias field. It could also solve the defects of SVM insimilar grayscale segmentation and vulnerable to noise. After the experiment,CV-SVM achieved good results in brain image classification of four parts and fivecategories of tumor extract. |