Medical images have a critical role in diagnosing the diseases and become an important assisting tool for doctors and experts. Despite the aid of medical images, there are still many difficulties in detecting the abnormal tissues from healthy ones. Manual segmentation method that is used by experts can be prone to error, not repeatable and time consuming due to fast growth of images in clinics during these years. Computerized image processing methods represent a good solution to many of the named problems. Segmentation means to divide images into meaningful regions which have common features like image intensities. As a result, in this thesis we focus on Magnetic resonance images of brain to segment tumor from normal tissues. Existing methods for segmenting brain tumors can be categorized into three broad family: Learning based, Generative and Combined methods. Learning based methods obtain information from images to characterize the brain lesions against other tissues. Generative methods make use of prior information about physical structure and anatomy of brain. Combined methods incorporate both the learning and generative methods. Variation of human brain anatomy may lead to decrease accuracy of generative methods. Also the learning based methods need a sufficiently large training data to generalize well into unseen images. In the end, learning and combined based methods show promising results among the recent algorithms. The proposed methods in this thesis gain from learning techniques. The 3D correlation and features of MR images play significant role in our methods. In the first method, supervoxel algorithm creates meaningful groups of voxels. Then, feature extraction and dir=rtl align=right Keywords: 1.Image Segmentation, 2. Magnetic Resonance Imaging, 3. 3D segmentation, 4. ltr"
