International Journal on “Technical and Physical Problems of Engineering” (IJTPE) Published by International Organization of IOTPE ISSN 2077-3528 IJTPE Journal www.iotpe.com ijtpe@iotpe.com March 2013 Issue 14 Volume 5 Number 1 Pages 103-107 103 IMPROVING CLUSTERING RESULTS USING RE-EVALUATION OF BOUNDARY DATA M.A. Balafar Department of Information Technology, Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran, balafarila@tabrizu.ac.ir Abstract- Image segmentation is preliminary stage in diagnosis tools and the accurate segmentation of brain images is crucial for a correct diagnosis by these tools. Due to inhomogeneity, low contrast, noise and inequality of content with semantic; brain MRI image segmentation is a challenging job. A post processing algorithm for improvement of clustering accuracy is proposed. The proposed algorithm re-evaluates boundary data to reduce clustering error. Proposed algorithm quantitatively evaluated by applying the mentioned algorithm on two recently reported clustering algorithms. The proposed algorithm improved clustering results and gives comparable results when user interaction is applied to the clustering algorithms. Keywords: Clustering, Brain, MRI. I. INTRODUCTION The identification of brain structures in Magnetic Resonance Imaging (MRI) is very important in neuroscience and has many applications, such as in the detection of temporary changes in the brain’s electrical function which causes seizure (epilepsy), tumours, Multiple Sclerosis (MS) and Alzheimer’s disease. Brain image segmentation [1, 2] is also crucial for the mapping of brain functional activation onto brain anatomy, the study of brain development, and the analysis of neuro- anatomical variability in normal brains [3]. In addition, it is useful in the clinical diagnosis of psychiatric disorders, treatment evaluation and surgical planning. MRI is an important imaging technique for detecting abnormal changes in different parts of brain at the early stages. It is popular to obtain images of the brain with high contrast. MRI acquisition parameters can be adjusted to give different grey levels for different tissues and various types of neuropathology [4]. MRI images have good contrast compared to computerised tomography (CT). The application of brain MRI image- processing techniques has rapidly increased in recent years. Nowadays, the capturing and storing of these images are done digitally [5]. However, the interpretation of their details is challenging. This matter is especially observed in regions with abnormalities, which should be identified by radiologists for future studies. Brain image segmentation is a key task in many brain image processing and diagnosis tools. Brain image segmentation aims to partition images to different regions based on given criteria for future processing. Brain images usually contain noise [6], inhomogeneity and complicated structures. Therefore, segmentation is a challenging job. However, precise brain segmentation is necessary for detecting tumours, oedema, necrotic tissues and clinical diagnosis [7]. There are different brain MRI image segmentation methods, like thresholding, region growing, statistical models, active control models and clustering. Due to noise [8], inhomogeneity [9] and the complexity of intensity distribution in medical images, the determination of the threshold is difficult. Therefore, usually a combination of the thresholding method with other methods is used for brain MRI segmentation. The region-growing method is an extension for thresholding, which adds connectivity to it. This method needs initialisation for each region, known as the seed, and inherits the problem of thresholding to determine suite threshold for homogeneity. Clustering methods are very common in brain MRI segmentation. Fuzzy c-means (FCM) [10] and statistical methods [11] are popular clustering methods. Brain MRI segmentation is a key task in many medical applications such as surgical planning, post- surgical assessment and abnormality detection. Noise is one of obstacles in brain MRI segmentation. Nowadays, radiologists use fast scanning techniques to reduce scanning times. These techniques raise the scanning noise level in MRI systems. There are different de-noising methods [12, 13] but they cannot totally remove noise Intensity inhomogeneity [14] is another obstacle for brain MRI segmentation which decrease similarity index. Intensity inhomogeneity is the smooth intensity change inside tissues. Inhomogeneity is hardly visible by the user. But even invisible ones are enough to hamper segmentation results. All inhomogeneity correction methods obtain just estimation for an inhomogeneity bias field and could not totally correct inhomogeneity. Sometimes due to the inequality of content with semantics, clustering methods fail to segment images correctly. For these images, it is necessary to post-process the clustering results.