脑MR图像互信息最大的凸优化分割模型  

A Convex Model for Segmentation of Tissues for Brain MR Images Based on Mutual Information Maximization

作　　者：潘晓花[1] 孙文杰[1] 韦志辉[1] 王平安[2] 孙权森[1] 夏德深[1] 

机构地区：[1]南京理工大学计算机科学与技术学院,南京210094 [2]香港中文大学计算机科学与工程学系

出　　处：《计算机辅助设计与图形学学报》2012年第8期1082-1089,1107,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(60773172);江苏省自然科学基金(BK2006704-2);江苏省博士后基金(AD41158);香港特区政府研究资助局资助项目(CUHK/4433/06M);香港中文大学研究项目基金(2050345)

摘　　要：脑MR图像中普遍存在灰度不均匀性,传统的分割方法无法得到理想的脑组织分割结果.为此提出一种基于互信息最大化准则的变分水平集凸优化分割模型.首先建立最大化图像灰度与标记之间互信息能量的分割模型,并融入偏移场信息;对模型进行水平集表示和凸优化后,再引入边缘指示函数加权的总变差范数;最后采用SplitBregman方法快速求解.实验结果表明,该模型可以得到较准确的脑组织分割和偏移场矫正结果,对噪声和灰度不均匀性有很好的鲁棒性.In segmentation of tissues for brain MR images, major difficulties arise from intensity inhomogeneities. Unfortunately, satisfactory segmentation results are hard-to-get using conventional segmentation techniques on MR images perturbed by intensity inhomogeneities. This paper presents a convex variational level-set segmentation model based upon the maximization of mutual information for brain MR images. First, an image segmentation model is developed aiming at maximizing the energy of mutual information between image intensity and label, in which the information of bias field is incorporated. The model is then equivalently related as a convex model with a level set method formulation. After the convexification, the model is modified by replacing the TV-norm by a weighted TV-norm given by the edge indicator function. Then the convex energy is easily minimized utilizing a Split Bregman iteration scheme. The proposed approach has capability of carrying out the segmentation of brain tissues and the estimation of bias field simultaneously. Experimental results demonstrate that our model is efficient and robust against the noise and the intensity non-uniformity.

关 键 词：脑MR图像 图像分割 互信息 变分水平集 凸优化 偏移场 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]