Evaluation Method for MRI Brain Tissue Abnormalities Segmentation Study SHAFAF IBRAHIM, NOOR ELAIZA ABDUL KHALID, MAZANI MANAF Faculty of Computer Science and Mathematics University Technology MARA Shah Alam, Selangor MALAYSIA elaiza@tmsk.uitm.edu.my Abstract: - Segmentation poses one of the most challenging problems in medical imaging. Segmentation of Magnetic Resonance Imaging (MRI) images is an important part of brain imaging research as it can facilitates the neurological diseases diagnosis. However, there are few limitations in evaluating the segmentation accuracy due to difficulties in obtaining the ground truth. This research proposes an evaluation method for brain tissue abnormalities segmentation study. Controlled experimental data called mosaic images are used as the testing data. The data is designed which that prior knowledge of the size of the abnormalities is known. It is done by cutting various shapes and sizes of various abnormalities and pasting it onto normal brain tissues, where the tissues and the background are divided into three different intensities. The knowledge of the size of abnormalities by number of pixels are then used as the ground truth to compare with the various segmentation results. The validation of segmentation was done with fifty data of each category using methods of Particle Swarm Optimization (PSO), Adaptive Network-based Fuzzy Inference System (ANFIS) and Fuzzy c-Means (FCM), where the evaluation for each technique exhibits some variation of results. Therefore, the proposed evaluation method of ground truth formation called image mosaicing is found to be reasonable and acceptable to use as it produces potential solutions to the current difficulties in evaluating the brain tissue abnormalities segmentation outcome. Key-Words: - Mosaicing,Evaluation method, Medical imaging, Segmentation, Magnetic Resonance Imaging (MRI). 1 Introduction Segmentation is the labeling of objects in image data and has been a crucial stage in many medical imaging processing tasks for operation planning, radio therapy or diagnostics, and studying the differences of healthy subjects and subjects with tumor. Its purpose is to subdivide an image into meaningful non-overlapping regions which analysis, interpretation or quantification can be performed [1]. In the past years, a large number of researches have focus on the development medical image segmentation methods as accurate segmentation of biomedical images can contribute to improve diagnosis, surgical planning and prognosis [2]. These leads to the increasing number in investigations of applications and considerable effort is needed to find reliable and accurate algorithms to solve the difficulties in evaluating the segmentation accuracy. In past several years, medical image segmentation problems has been approached with several solution methods by different levels of automation and range of applicability such as Particle Swarm Optimization [3], Genetic Algorithm [4], Region Growing [5], [7], Adaptive Network-based Fuzzy Inference System (ANFIS) [6], Self Organizing Map (SOM) [8] and Fuzzy c-Mean (FCM) [9]. However, segmenting brain internal structures remains a challenging task due to their small size, partial volume effects, anatomical variability, and the lack of clearly defined edges [10]. Thus, a thorough evaluation of its performance is necessary with some quantifiable measurement of its accuracy and variability [11]. Evaluation is not only used in evaluating the performance of segmentation algorithms. It could also be used in combining the results of several segmentation results [12], and acted as a guide in selecting appropriate segmentation algorithms [13]. Nevertheless, evaluation of segmentation performance has been very subjective that leaves the researcher in tricky situation [14]. Therefore, it may leads to difficulties in judging the effectiveness of the techniques implemented. Chabrier et al. [15] found that it is difficult to evaluate the segmentation Recent Researches in Computer Science ISBN: 978-1-61804-019-0 297