Brain MRI Automatic Segmentatin Based On Improved MRF Parameter Estimation

The current research on the human brain physiology and function is in its infancy. Magnetic resonance (MR) technology is particularly effective for the soft tissue of the brain imaging, and it has no radioactive harm to the human body, so the MR technology has become an important adjunct to the clinical diagnosis of brain diseases. Accurate segmentation of brain MR image has important guiding significance for clinical treatment and scientific research about the brain anatomy, detection of different pathological conditions which affect the brain parenchyma, radiotherapy planning and the study of brain functional. However, due to the inhomogeneity in the resonance field (RF), the MR device itself, the differences and partial volume effect of brain tissue, the image uniformity is poor. What’s more, the low contrast ratio, the vague border between the different soft tissues as well as the complex shape of the structure lead difficulties to segmentation. In addition, the clinical use of hand-painting on image segmentation at present, which makes heavy workload, may tends to be error-prone due to personal subjective reasons. Therefore, the accurate automatic computer-aiding segmentation of brain MR image seems to be very necessary. Based on the reasons above, the thesis will study how to use Markov random field (MRF) to achieve automatic segmentation of brain MR image.The thesis found that:(1) the differences among gray matter, white matter and cerebrospinal fluid is mainly in gray-scale;(2) classical clustering algorithm is easy to implement, and faster, but less robust;(3) MRF model can deal with the degradation of brain MR image caused by partial volume effect and artifacts, but the accuracy has to be further enhanced. In response to these problems, the paper use the C-V model preprocessing to remove the brain tissue first of all, and then the segmentation is achieved by the MRF model according to the guidelines of MRF-MAP.Main work of the thesis was focused on the MRF parameter estimates and optimization, which is to select the appropriate parameters in order to make use of MRF modeling the extracted MR image gray-scale information. The thesis described several existing methods about the gray field parameters and penalty factor estimation, and takes advantage of the improved particle swarm optimization (PSO) algorithm and local entropy method to solve the problems brought by these methods, which did not only ensure the robustness of segmentation, but also improve the segmentation accuracy. The thesis experiments used real brain MR images from IBSR database and clinical brain MRI images from Huashan Hospital of Fudan University. The results showed that the proposed method realized fully automatic segmentation and had greatly improved the accuracy of segmentation.