Abstract: For the disadvantage of anti noise performance and robustness of traditional Gaussian mixture model, this paper presents a brain MR images segmentation method based on the hidden Gaussian mixture model. Due to neglect of the spatial information and the segmentation results distribution, the traditional Gaussian mixture model is imcomplete. In response to these shortcomings, in this paper, the probability density function of the segmentation results which is regarded as the hidden data is introduced into the Gaussian mixture model and a nonlinear weighted hidden Gaussian mixture model is established. Meanwhile, the Gaussian weighted exponent which contains spatial information and the smoothing factor is introduced. And Expectation-Maximization（EM） algorithm and Newton iteration method are used to calculate the class mean and variance, and the smoothing factor. Finally, the segmentation results are obtained according to the maximum a posteriori criterion. Experimental results show that the method proposed in this paper has good robustness and anti-noise performance to the human brain MRI.

Key words: human brain MRI, spatial information, Hidden Gaussian Mixture Model, Newton iteration method, Expectation-Maximization（EM） algorithm

摘要： 针对传统的高斯混合模型的抗噪性能和鲁棒性较差的缺点，提出一种基于隐高斯混合模型的人脑MRI分割方法。传统的高斯混合模型由于忽略了空间信息和未考虑分割结果的分布情况导致模型不完整。针对这些缺点，把分割结果的概率密度函数作为隐含数据引入到高斯混合模型，建立了非线性加权的隐高斯混合模型；同时引入了含空间信息与平滑系数的高斯权重置指数；运用期望最大化算法与牛顿迭代法对类均值，类方差以及平滑系数进行求解，最后根据最大后验概率准则得到人脑MRI的最终分割结果。经实验表明，提出的方法对人脑MRI具有很好的鲁棒性与抗噪性能。

关键词: 人脑MRI, 空间信息, 隐高斯混合模型, 牛顿迭代法, 期望最大化（EM）算法