Research Of Segmentation And Bias Field Correction In Infant Brain MR Images

In recent years,with the improvement of medical technology,and the substantial increase in the neonatal intensive care unit(NICU),the survival rate of the high-risk children and low-weight children is increasing greatly,and the incidence of infant brain diseases is on the rise.The early childhood(0-2 years)is a critical period of brain development,and the early childhood is the basis of brain development.The impact of the early childhood on the brain will continue until adulthood,so it's extremely important for us to effectively analyze the brain development of the early childhood.As a result,the research on infant brain segmentation method is important.Brain image segmentation has important medical significance for brain lesion detection and diagnosis.Magnetic resonance imaging(MRI)technology has been universally applied to the detection and analysis of the clinical brain disease diagnosis,because of its high soft tissue contrast ratio,the function of multidirectional layer cutting,multiparameter imaging and non-invasive imaging.etc.At the same time due to the MR device,volume effects and differences in brain tissue between different organizations,the image of the gray uniformity deteriorate,resulting in an additional image bias field.And that is why the bias field of image processing is also important for the accurate of image.This thesis does some improvements for bias field correction segmentation algorithm.Specific work is as follows:(1)In this thesis,the low-pass filtering algorithm is improved by the rough set theory.Since the bias field belongs to the low frequency signal in the image.In the thesis,classical low-pass filter is used to filter out the bias field,The lower approximation set and the boundary set are formed for the division of the filtering window.Different weight ratio are given to diverse element in the window which can effectively improve the accuracy of the bias field in the boundary region and the edge position pixel in the image.Base on the theory that point within a certain gray difference of the center pixel contribute more to the calculation of the bias field than the other elements in the window,lower approximation set and the boundary set are given different weight ratio.This new algorithm can better remove the bias field.(2)In this thesis,to study the Gaussian mixture model,make the maximum a posterior probability(MAP)as the optimal criterion.The GMM prior to the spatial information in Hidden Markov algorithm is introduced to the segement algorithm,calculate the optional mean value and variance information by the EM iteratively algorithm.In this thesis,stomped Gaussian distribution and Combining with Rough Set.Stomped area been known as the lower approximation set and other area been known as the boundary set.The parameters in GMM are also updated.The first step is important,Improved k-means Algorithm is been used in the new Algorithm to get better Initial segmentation parameters.In general,The new algorithm can get better segmentation.(3)In order to work off the impact of the non-typical samples and noise on final estimate when estimating the component mean and covariance matrices by the light tail characteristic of Gaussian distribution.Student t-distribution mixed model is used to replace Gaussian mixture model as Student's distribution has a heavier tail.Which can improve the noise immunity of the algorithm to a certain extent.The bias field is converted to additive by a logarithmic transformation,so the bias field correction and segmentation can be achieved at the same time.New methods can be better segmentation and bias field correction through experiments.