Application Of Support Vector Machines In Tiny Multi-Target Segmentation Of MRI Brain Tissues

The boundary of encephalic tissue is highly complicated and irregular in head magnetic resonance image, and the number of samples is limited. It's the reason that the traditional segmentation methods based on the empirical risk minimization is not suitable. Support vector machine based on statistical learning theory is a supervised classification method, which follows the structural risk minimization principle, shows many special advantages in resolving the small sample set, nonlinear and high dimensional pattern recognition problems. Therefore, this paper carries out the research of support vector machine to segment caudatum, putamen and pallidum region in brain magnetic resonance imaging(MRI).Support vector machine was originally used for two-classification, a multi-classification classifier can be constructed by a few two-classification classifiers using directed acyclic graph. And the multi-classification classifiers can well segment caudatum, putamen, pallidum and background region of the MRI image.In addition to the classifier, the final classification effect also has an important relationship with the feature vector extracted from the brain MRI images. The texture features and gray features are extracted as the feature vectors in the experiments, Texture features are extracted from gray co-occurrence matrix. The total dimensional number of each sample point is 58.Since the high dimensional feature vectors seriously impact the calculation speed, and reduce the segmentation speed, the principal component analysis and the rough sets are adopted respectively to reduce the dimension of feature vectors. A great deal of experiments shows that rough set is better than principal component analysis in the speed and the result of the segmentation.In order to analyze and verify the actual effect of the proposed segmentation algorithm based on SVM, k-means clustering, fuzzy c-mean segmentation, k-nearest neighbor, Bayes classifier, and radial basis function neural network are respectively adopted. The false alarm probability, false dismissal probability and the segmentation accuracy are used as objective indicators. The comparison analysis can objectively indicate the validity of the proposed segmentation algorithm. Experimental results show that whether or not to adopt dimensional reduction processing, the proposed segmentation algorithm is better than the five methods above.