Research Of Brain MR Image Segmentation Algorithm Based On FCM And SVM

Brain disease is a threat to human health of common diseases, become one of the problems attracting the wide attention from society, the medical profession. The diagnosis of brain disease mainly relies on medical imaging. Through medical imaging qualitative and quantitative analysis of brain tissue, and then analyze the relationship between brain tissue and brain diseases has become a research hotspot. Nuclear magnetic resonance imaging (MRI) due to its especially effective on soft tissue imaging, so brain MR images in clinical medicine to get a number of applications. Precise segmentation on brain structure can improve the efficiency and reliability of brain disease diagnosis.In this thesis, a variety of image segmentation method at home and abroad are reviewed, in view of the problems that exist in the brain MR image segmentation, found that based on fuzzy c-means clustering (FCM) in combination with support vector machine (SVM) method has many advantages and has great application prospect, as a result, this article will focus on brain MR image segmentation based on FCM and SVM on the research of the main work and research results are as follows:First based on the advantages and disadvantages of FCM algorithm and SVM algorithm, build a combination of unsupervised classification algorithm of FCM to fuzzy support vector machine (FSVM) model. The FCM algorithm provides the training sample and fuzzy membership degree for FSVM, to solve the disadvantages of the FSVM algorithms which need to manually select sample, and make full use of good generalization ability of FSVM.Aiming at the phenomenon which FCM algorithm whose sample points is so serious aliasing that membership of meaning is not clear, this thesis adopts the improved fuzzy segmentation of FCM algorithm, and on this basis to add the space constraints and bias field estimation, enhances the processing capacity of noise and bias field, providing support vector machine segmentation a better membership degree matrix of training samples and more clear meaning.At last, Nearest Neighbor Figured method is proposed to filter the training samples in high dimensional feature space, fuzzy membership degrees of the abnormal points which the clustering algorithm can’t distinguish are reduced to a number close to zero, minimizing the effects of these abnormal points to FSVM classification hyper plane and obtaining better segmentation precision and generalization ability. And the segmentation accuracy of this algorithm is verified through experiments in the image with high noise and high bias fields.