Fast medical image mixture density clustering segmentation using stratification sampling and kernel density estimation

Abstract

The Gaussian mixture models (GMMs) is a flexible and powerful density clustering tool. However, the application of it to medical image segmentation faces some difficulties. First, estimation of the number of components is still an open question. Second, the speed of it for large medical image is slow. Moreover, GMMs has the problem of noise sensitivity. In this paper, the kernel density estimation method is used to estimate the number of components K, and three strategies are proposed to improve the segmentation speed of GMMs. First, a histogram stratification sampling strategy is proposed to reduce the size of the training data. Second, a binning strategy is proposed to search the neighbor points of each center data to compute the approximate density function of the samples. Third, a hill-climbing algorithm with the dynamic step size is designed to find the local maxima of the density function. The kernel density estimation method and sampling technology reduce the effect of noise. Experimental results with the simulated brain images and real CT images show that the proposed algorithm has better performance in generating explainable segmentations with faster speed than the common GMMs algorithm.