Medical Image Segmentation Based On T Mixture Models

The technology of Magnetic Resonance Imaging (MRI) with high soft tissue resolution, non-invasive and rich imaging data, and so on, have been widely used in medical image acquisition, and also has played an increasingly important role in the clinical diagnosis. Segmentation of MRI images has a great significance in biomedical and clinical fields, and so has been extensively applied in the study of anatomical structure, tissue quantification, lesion localization and disease diagnosis. Thus an accurate segmentation is the basis of subsequent image analysis and diagnosis. Because the brain is the most important part of human body, segmentation of MRI brain images has become the research focus of medical image segmentation with its great value.In this thesis, we summarize the purpose of medical image segmentation and practical significance based on the results of previous researches. After deeply understanding the principles of magnetic resonance imaging and the characteristics of MRI images, finite mixture model is choosen as the main framework for segmentation, which is based on theory of Probability and Statistics,.Finite mixture model is a flexible and powerful tool for analyzing complicated dataset, which provides an efficient method of simulating a complicated density function by simple density functions. It has been widely used in pattern recognition, machine vision, machine learning and other fields. Finite mixture model is often used for unsupervised learning (such as clustering) in statistical pattern recognition. Because gaussian density function has a simple expression, easily-estimating parameters and is similar to the shape of the actual data sets, the gaussian mixture model is one of the most popular modeling tools. However, gaussian mixture model can not be applied to the MRI images with high noise in view of the affection by atypical samples or outliers,t-distribution with heavier tails and good properties of anti-noising is a modeling tool more robust than the gaussian mixture model, which can solve the problems of gaussian mixture model to a certain extent. The experiments in the thsis show that the t mixture model can adapt better than the Gaussian mixture model due to the intensity characteristics of MRI imagesMarkov random field (MRF) theory is a branch of probability theories for analyzing the spatial or contextual dependencies of physical phenomena. MRF theory provides a convenient and consistent way of modeling context-dependent entities such as image pixels and correlated features. Although t-distribution can deal with the problems of gaussian mixture model to some extent, finite mixture model does not take spatial information into account, which is a histogram-based model. To overcome this difficulty, we import spatial information constraints by incorporating estimation of mixture model with MRF theory to make up the deficiency of finite mixture model.Since MRI is a multi-parameter, multi-core kind of imaging technology, we can acquire three different kinds of images:T1 weighted images, T2-weighted images and proton density (PD) weighted images, which provide different information of the same human tissue. The information provided by different weightsed images on tissue is redundant and also some are complementary, so it meets the condition of image segmentation based on data fusion for MRI image segmentation. Pixels misclassified concentrate in the transitional region of different tissues owing to the characteristics of MRI images with high noise, fuzzy boundaries and artifacts. Classification of pixels in the transitional region of different tissues has a greater uncertainty and unpredictability, because the gray values of the region between different tissues in brain magnetic resonance images are overlapped, which are known as "artifacts". The phenomenon of artifacts particularly occurs in the transitional region between cerebrospinal fluid (CSF) and gray matter (GM) or Gray matter (GM) and white matter (WM). Due to the unknown and uncertainty of information can be well described by DS evidence theory, we can use Dempster's rule to integrate multi-information in order to improve the segmentation results. Therefore, in this thesis, we firstly use finite mixture model to classify the MRI image, and then employ the DS evidence theory to integrate the classified results, which can finally improve the accuracy of segmentation results.