摘要： 传统高斯混合模型分割核磁共振图像时严重依赖初值，且易受图像中偏移场与噪声的影响。为此，提出一种基于片信息的改进高斯混合模型。采用模糊C均值聚类方法优化初始值，以减小初值对分割结果的影响，加快算法的收敛速度。使用Legendre多项式对偏移场进行拟合，并融入EM框架中，得到光滑的偏移场。利用邻域信息降低噪声的影响，使模型在降低噪声影响的同时，保留细长拓扑结构信息。实验结果表明，该模型能恢复出偏移场，分割结果较好。

关键词: 核磁共振成像, 图像分割, 高斯混合模型, 偏移场, 区域信息

Abstract: Traditional Gaussian Mixed Model(GMM) seriously depended on the initial value and it is easily affected by the bias field and noise when segmenting Magnatic Resonance(MR) image. Aiming at this problem, this paper proposes a kind of improved Gaussian Mixed Model(GMM) based on patch information, which can manage the bias field and noise while segmenting the image. It utilizes the Fuzzy C-Means(FCM) method to optimize the initial value and accelerates the convergence. In order to obtain a smooth bias field, it employs the Legendre Polynomials to fit it and merges it to the EM framework. This paper introduces neighbor information of each point to reduce the effect of noise so that slender topological objects can be reserved. Experiments results show that the model can bring out bias field and has good segmentation results.

Key words: Magnatic Resonance Imaging(MRI), image segmentation, Gaussian Mixed Model(GMM), bias field, patch information

中图分类号: 

• TP391.41