Classification Of Magnetic Resonance Brain Images Based On Kernel Support Vector Machine

This thesis introduces the history and the current advances of the development of MR brain images diagnosis. The new idea of combining support vector machine(SVM) with early diagnosis of abnormal MR brain images is proposed, which realizes the application of pattern recognition to the classification of real data. SVM has many advantages which get global optimum solution and excellent generalization ability, to solve the problems of image classification and pattern recognition. But in the application of SVM, the selection of kernel parameter and punish coefficient has an important influence on the result. Only select appropriate parameters, can have good generalization ability of SVM classifier. In this thesis, the features of MR brain images are extracted, and some of the features are combined, which makes it possible to classify the MR images using a small number of features. Firstly, compared the difference of MR brain images in the function of polynomial function, kernel function, linear function and radial basis function. Then use the meshing method to the best parameters c and σ. In order to improve the classification effect, the genetic algorithm and particle swarm optimization is used to optimize the parameters in the thesis. The optimized SVM algorithm is used in the classification of the MR images data. At last, the results are compared with the current classification methods, including the KNN algorithm, the KM algorithm and fuzzy neural network algorithm.(1) Method:we classify the abnormal or normal images based on the wavelet energy and wavelet entropy (WEWE) features which extracted from MR brain images, combining some features to find the best classification accuracy and the least number of features. There are 125 cases of samples, which are randomly divided into training set and test set. Then train the SVM network, choose the appropriate kernel parameter σ and punish coefficient c. The use of GA and PSO to optimize the parameters, and repeat the process before. Every method of parameter selection is conduced in the sense of cross validation. Comparing the classification model of best parameters with random selected parameters. Find compare the results of various methods, including KNN algorithm and fuzzy neural network algorithm which have been tried in the course of algorithm and exploring. Find the advantages of various algorithms and choose the most appropriate.(2) Results:we find the high classification accuracy of MR images by using the features of WEWE and particle swarm optimization algorithm. The SVM classification has the best effect of 99.68%. Then, comparing PSO-SVM with BP neural network, KM algorithm, linear discriminant method, KNN classification algorithm, PSO-SVM classification effect is the best. What is more, the PSO optimization method has a higher classification accuracy, which has a better generalization ability and a faster operation speed. As a result, the PSO optimization method to the SVM network is more suitable for the abnormal MR images diagnosis and worthy of further research.