Study Of Medical Image Segmentation Algorithm Based On Finite Mixture Model

Image segmentation is an important branch in the field of image processing, whose target is divided images into several parts based on gray or texture information. Firstly, we introduce some current medical image segmentation algorithm in detail, and study the principles of image segmentation based on finite mixture model and the process of parameter estimation using the EM algorithm, elaborate existing problems and some solutions when applying EM algorithm to estimate parameters.Gaussian distribution is relatively simple and close to the actual distribution shape of a data set, which is a common modeling tool. This thesis first studies the medical images segmentation based on Gaussian mixture model (GMM). GMM is a histogram statistical model that pixels in a image are independent, without considering the prior knowledge that neighboring pixels are likely to belong to the same category, resulting in the traditional GMM can achieve good segmentation result without noise, while segmentation get worse with the noise increasing. Based on this shortcoming, we propose a method to add adaptive mean filter into priori probability and conditional probability of GMM, that is, Gaussian mixture model with adaptive mean filter (2D GMM-AM). Misclassification rate MCR and segmented regions’Dice similarity coefficient are used in the experiment to evaluate the merits of the segmentation algorithm. Experimental results show that 2D GMM-AM algorithm not only can preserve image details, but also tolerance to noise.Student’s-t distribution has heavier tail compared to Gaussian distribution, which can suppress the atypical samples fundamentally and is more robust than Gaussian distribution. On the basis of the GMM, we study the medical image segmentation algorithm based on Student’s-t mixture model (SMM). Since there is no consideration of the spatial relationship in image pixels when using SMM to segment image, imitate the idea of 2D GMM-AM, we propose Student’s-t mixture model with adaptive mean filter (2D SMM-AM), and verify the result that 2D SMM-AM segmentation algorithm is superior to traditional SMM on accuracy and robustness by experiment.With the development of image processing technology, people are no longer satisfied with the information provided by the two-dimensional image. In order to analysis human tissue from stereo and multi-angle, we extend two-dimensional segmentation algorithm to three-dimensional, taking into account the spatial relationship among pixels in a slice, and also considering the impact of the different slices. Like the idea of 2D GMM-AM and 2D SMM-AM, we propose 3D GMM-AM and 3D SMM-AM methods by adding three-dimensional adaptive mean filter. Experimental results show that the 3D GMM-AM segmentation algorithm has a lower MCR value than 2D GMM-AM, and 3D SMM-AM segmentation algorithm has a lower MCR value than 2D SMM-AM, with a improvement in segmentation accuracy and robustness. Meanwhile 3D SMM-AM, which is more robustness, gets a better segmentation result than 3D GMM-AM segmentation.