White Matter Lesion Segmentation For Brain MRI

Medical image segmentation can date from1980s. But with the development of medical imagingtechnology, and appearing of some novel imaging techniques, many brand new techniques continuallychallenge researchers’ ability in the field of image processing, especially when the magnetic resonanceimaging appears later. Magnetic Resonance Imaging (MRI) processing has become one of the mostimportant research topics. Meanwhile, a lot of issues come up with this kind of new imagings. Manytraditional methods can not deal with this high dimension image data directly. On the other hand,medical image processing hlep a lot with diseases diagnosis and predict. For exmple, White MatterLesion has a close relationship with multiple sclerosis and Alzheimer’s disease. It is very importantfor Alzheimer’s desease diagnosis and predict. So we need to redesign those traditional and more newmethods need to be presented.Base on this background, we have a deeply reserch on MRI imagesegmentation of human brain White Matter Lesion. WML analysis begins at the early ninety of lastcentury. It has made a great progress after more than10years of effort. But it is not easy for real-timesegmentation task, the main reasons include:(1)3D MRI images have a big feature space and thesample amount is too large for rapidly computing.(2) As the imaging technology limited, most ofMRI images have lower resolution compared to the traditional image.and less availablefeatures.Based on the above two aspects of the problems, this thesis has made the related study.wepropose a fast clustering method to achieve real-time application requirements.the proposed algorithmonly requires the user to provide a few regions of interest (ROI’s) containing lesions.A k-meansclustering algorithm is applied to segment these ROI’s into areas.The segmentation is done by usingthe probability distributions to generate a confidence map of lesion and applying random walks tolabel lesion voxels. This method improves the speed of segmentation while ensuring theperformance.For the second problem; we focused on improving the overall performance of automaticsegmentation algorithm. We bring the context features into the WML segmentation, and design newalgorithm to make full use of limited information of the WM image. Experimental results show thatboth proposed algorithms agrees well with expert labels and achieve good perfomances.