Research Of Statistical Learning Based MR Brain Image Segmentation Algorithms

MRI (Magnetic Resonance Imaging, MRI) technology has become a importanttool for clinical brain tissue structure analysis, disease diagnosis, surgery, andreconstruction of three-dimensional visualization. Because of its no radiation damage,high accuracy, high resolution, motion artifacts and less shadow, etc.. It has a veryimportant significance.Brain MRI image segmentation analysis is the basis of brain image analysis andresearch. It has a guiding role for brain tissue study and disease research, besides, it isalso a hot topic in the field of image processing. The essence of MR brain imagesegmentation is to divide the brain image into a number of regions with differentattributes, and to obtain the region of interest. It includes normal brain tissuesegmentation and abnormality segmentation.This paper describes the characteristics of MRI imaging technology, basis ofimage segmentation, common segmentation methods and existing segmentationproblems. Then, two brain MRI image segmentation methods have been proposedbased on statistical learning theory:1) The first segmentation method is for brain images of healthy subjects.Because of brain tissue image with noise, uneven distribution of gray leveland other characteristics, the gray level spatial distance weighted K-meansclustering and fuzzy confidence brain tissue segmentation algorithms havebeen proposed in third chapter. The algorithm has good stability with highaccuracy, clarifying the role of edge and noise suppression, which improvedthe limitations of traditional K-means clustering. It can provide a referencefor MR brain image anomaly detection and segmentation.2) The second method is to segment tumor from patient brain images. Braintumor has irregular borders, uneven grayscale and differences of the imagingenvironment, etc.. A multi-modal tumor segmentation method integratingSupport Vector Machine and multi-Gaussian distribution Markov RandomField model has been proposed in the fourth chapter. The algorithm utilizedmultiple modes of MR image information, including intensity, texture and others. The introduction of priori knowledge for the model parameterinitialization has improved the segmentation accuracy. It can accomplishautomatic detection and segmentation of brain tumor.