A Noval Method For Liver Image Segmentation Based On Sparse Regression Model For Local Region

With the rapid development and wide application of computer technology in medical imaging,medical image segmentation also gets a unparalleled space for development.Liver segmentation is the separation of the liver from the abdominal CT image according to anatomical features of the liver and information about intensity,texture and statistical structure of the liver image.Therefore,the achievement of automatic,efficient and accurate liver segmentation is of considerable significance in surgery planning,disease diagnosis and computer-aided intervention therapy.This paper uses an active shape model-based liver segmentation framework.This framework not only exploits image features of liver CT images,but also considers specific shape prior model of liver,which is able to reduce the over-segmentation and the under-segmentation on liver CT image segmentation.In this paper,a normalized criterion based on the standard deviation of spatial position and the standard deviation of intensity values is proposed,which considers the intensity information of liver CT images as well as the shape variation of liver.Thus,this criterion can carry out effective and accuracy region division for the normal liver shape and the liver shape with complex structure and large lesions.Inspired by hyperspectral unmixing algorithm based on sparse regression,the sparse regression model is introduced into liver segmentation in this paper.The paper assumes that the input liver shape is approximately refined as a sparse linear combination of training shapes in a liver shape repository,and proposes a sparse regression model based on local region.In order to work out the sparse regression model based on local region proposed in this paper,the sparse regression optimization problem will be reformulated as the Basis Pursuit Denoising(BPDN)problem which only deals with one variable.Then the reformulated BPDN problem will be solved by the alternating direction algorithm of multipliers.The active shape model-based liver segmentation framework adopted in this paper is very sensitive to initial position of the liver region.In order to obtain an initial shape of liver to be segmented,an atlas-based automatic liver segmentation method is exploited in this paper for finishing the initialization of the liver shape.In order to make the active shape model-based liver segmentation framework adopted in this paper more effective and more accuracy,the multiscale adaptive liver segmentation strategy is proposed.This paper first handles all liver images of training data set by the multiscale adaptive Gaussion filter,and then carries out liver segmentation in combination with the active shape model-based liver segmentationframework.Finally,the paper has carried out the test experiment of liver segmentation on ten CT images of the abdomen.In order to evaluate the performance of the method for liver segmentation proposed in this paper,the proposed method for liver segmentation is compared with ten methods of liver segmentation competition.Experimental results show that the method for liver segmentation proposed is of better generalization ability and higher performance of segmentation.