A new medical image segmentation algorithm based on Gaussian-Mixture model

摘要

In this paper, we propose a probability model based method where the image pixels' features are modeled as Gaussian-Mixture distribution. Then the segmentation problem can be reduced to the estimation of the parameters of the Gaussian-Mixture model. Traditional method of estimating the parameters is EM (expctation maximization). But it has the drawbacks of heavy computational load and sensitvity to initialization. In this paper, we get the initial parameters for EM by two steps: (1) Anisotropic diffusion is applied to original image. The histogram of the iamge after anisotropic diffusion is expected to have distinct peaks and valleys to detect, while in original image the modes may be overlapped to detect accurately. (2) A histogram analysis method is presented to deal with parameter initialization. Then the EM algorithm is applied to estimate the parameters iteratively. Due to the good initialization, the heavy computational load and instability of EM are overcome.