OPTIMIZATION OF FEATURES AND CLASSIFICATION OF BRAIN MRI IMAGES USING MAJORITY VOTING

Abstract

Magnetic Resonance Imaging is a very popular method for abnormalities detection in the human body. Computer vision systems make this process automatic and more accurate detection so that diagnosis can be done properly. In this thesis a majority voting based system is designed based on the different intensities of texture features in different part of images. Every image is divided into four parts and all the classifiers are applied on every part of the image. The best classifier is found for every part based on the performance parameters and then majority voting concept is applied to get final class results. One another method that is used in this thesis is feature selection which is used to found the best feature subset among the search space. Total ten texture based features are extracted from the images and represented by string of 0s and 1s. Those which are included are represented by 1s and those which are not included by 0s. The same process is repeated but this time with a combination of feature selection and majority voting. A method for the classification of MS lesion brain MRI images is proposed. The dataset consists of 35 MRI images is taken from which 20 images are used for training (8 normal images and 12 abnormal images) and rest 15 images are used for testing of the classifier (6 normal and 9 abnormal images). Ten texture based features energy, entropy, variance, correlation, inertia, cluster shade, cluster prominence, IDM function, angular second moment, are extracted from the images, on which feature selection is applied to obtain the best feature subset which will give highest fitness function value. Total five classifiers SVM, LDA, KNN, Decision tree and Naïve Bayesian classifier are used whose performance parameters are compared to get the best classifier for a set of images.