International Journal of Computer Applications (0975 – 8887) Volume 37– No.12, January 2012 1 A Robust Brain MRI Classification with GLCM Features Sahar Jafarpour Electrical Engineering Department Urmia University, Iran Zahra Sedghi Electrical Engineering Department Urmia University, Iran Mehdi Chehel Amirani Electrical Engineering Department Urmia University, Iran ABSTRACT Automated and accurate classification of brain MRI is such important that leads us to present a new robust classification technique for analyzing magnetic response images. The proposed method consists of three stages, namely, feature extraction, dimensionality reduction, and classification. We use gray level co-occurrence matrix (GLCM) to extract features from brain MRI and for selecting the best features, PCA+LDA is implemented. The classifiers goal is to classify subjects as normal and abnormal brain MRI. A classification with a success of 100% for two normal and abnormal classes is obtained by the both classifiers based on artificial neural network (ANN) and k-nearest neighbor (k-NN). The proposed method leads to a robust and effective technique, which reduces the computational complexity, and the operational time compared with other recent works. General Terms Pattern Recognition, Classification. Keywords Brain MRI, Feature extraction, GLCM, ANN, KNN. 1. INTRODUCTION Nowadays widespread and universal use of computer technology in medical decision support covers a wide range of medical area, such as cancer research, hart diseases, gastroenterology, brain diseases etc. In the recent century, the Computer-Aided Diagnosis (CAD) [1] is progressively becoming an essential area for intelligent systems. Magnetic resonance imaging (MRI) is a valuable diagnostic study, which is a non-aggressive, nonradioactive and pain-free method of assessing the human body, especially the brain. The importance and necessity in accurate brain pathology diagnosis and treatment requires more accuracy in automatically classifying MRI images in distinguishing disease without human interference. In CDA there is a challenging process for automatically classifying MRI in normal and abnormal classes. For this goal researchers have proposed a lot of approaches which fall into two categories. One category contains supervised classification techniques such as artificial neural networks (ANN) [2, 3] and support vector machine (SVM) [4]. The other category has unsupervised classification techniques such as self-organization map (SOM) [4] and fuzzy c-means [5]. Since the goal of this study is to design a more efficient and accurate classifier and on the other hand, supervised classifiers in the term of classification accuracy has a better performance than unsupervised classifiers, we use supervised machine learning algorithms (ANN and K-NN) in our proposed method. Most of works for classifying MRI are based on pattern recognition methods that their main issue is to extract effective features, often by utilizing Digital Wavelet Transformation (DWT) [2, 4, 6] or Co-occurrences Matrix [7]. Gray Level Co-occurrences Matrix (GLCM) introduced in [7,8,9] is used to extract features. GLCM has less computational complexity in comparison to other methods like wavelet transform. The main objective of this paper is to extract effective features using GLCM [7, 10, 11, 16]. These features are completely explained in [8, 9] and [12]. Projection of original feature space, through a transformation, into a smaller subspace is what feature reduction methods do. Linear Discriminant Analysis (LDA) [13, 14] and Principal Component Analysis (PCA) [13, 14] are two major methods which extract new features in different areas. The features extracted by feature extraction methods might have correlations. In the proposed method we use a process of PCA+LDA [13] that leads to the best uncorrelated effective features. In addition to decreasing dimensionality in this method, complexity and time cost are decreased in a satisfied range. In the next step to perform the classification on the input data an artificial neural network (ANN) and a K-nearest neighbor (K-NN) [15] classifiers are used. The proposed method deals with an efficient feature extraction tool and a robust classifier which results in a more robust and accurate automated MRI normal/abnormal brain images classification. The structure of this paper is organized as following: Section 2 has a short description of our method, which consists of database, feature extraction and feature reduction. Classification methods are presented in section 3. Discussion and comparison with previous works are presented in section 4. Section 5 concludes this paper. 2. METHODOLOGY The proposed method as illustrated in Fig. 1 is based on the following techniques: Gray Level Co-occurrence Matrix (GLCM), the principle components analysis (PCA), Liner discriminant analysis (LDA), artificial neural network (ANN) and K-nearest neighbor (K-NN). It consists of three stages: feature extraction stage, feature reduction stage and classification stage. KNN and ANN classifiers with three classes as normal, tumoral and MS are used in classification stage. 2.1 Data Base Two different databases are used in this paper. The first database covers 120 real human brain MRIs with 41 normal and 79 abnormal images, which 43 of them are MS and 36 are tumoral. This data base is collected from the Harvard Medical School website [17]. A sample of each set is illustrated in figure 2. The second used data base is a collection from