Research On Segmentation Algorithm Of Brain Magnetic Resonance Image

The physiological and functional mechanisms of human brain are not well known now, magnetic resonance imaging, which is thought to be non-radiant, plays an important role in disease diagnosis and adjuvant therapy, and also provides a new way to help explore the mysterious organ - brain. The goal of this thesis is to develop a segmentation method to separate the gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF). The ratio alterations between GM and WM in human brain take important effects in disease diagnosis, where an accurate segmentation may do a great favor. Although a number of image segmentation approaches have been proposed, the processing speed and accuracy are still challenges to the algorithms.With respect to the features of brain MRI: (1) There is no obvious difference among the texture characteristics of gray matter, white matter and cerebrospinal fluid, especially for the intensity-invariant texture feature. (2) The differences among the gray matter, white matter and cerebrospinal fluid were mainly on the intensity value. (3) Although classical clustering algorithms are easy to realize with reasonable processing speed, yet the robustness remains a problem. (4) The graph cut algorithm showed advantage with good robustness, nevertheless it is time-consuming and computational expensive.This paper has focused on two points. The first is feature extraction. The gray scale gradient and texture information were measured. Finally, a new feature was defined which helps improve the segmentation accuracy. The second is segmentation methods. This thesis concentrates on the Ncut algorithm, analyzing both advantages and disadvantages of it. Then a modified segmentation algorithm was proposed, which preserved the merit of the robustness of noise resistance and meanwhile improved the computing speed.The new segmentation method proposed in this paper. Firstly, preprocessing the image, then extract the new features; combine the Ncut and clustering algorithm to finish the segmentation. The proposed method was applied to both synthesized image and real image, using MATLAB 2010. All experiment results demonstrated a better segmentation performance on computation speed and noise resistance.