Deformable Model To Medical Image Segmentation Using Statistics From Multiple Sources

In recent years,medical image technology has attracted great attention in the field of clinical medicine because of its good characteristics,such as non-invasiveness and high specificity.It has been widely recognized in the fields of medical diagnosis and interventional therapy,and has now become an important part of clinical medicine.This makes the medical image technology develop rapidly and the medical image data increase drastically on behalf of Ultrasound,CT,MRI,etc.However,the processing,analysis and diagnosis of medical images are cumbersome and repetitive.Meanwhile,the medical images present different appearance because the anatomy of organs is complex and the imaging conditions are variable.These make doctors owning high level of experience in the diagnosis process.The continuous and rapid development of image processing and analysis based on algorithms and models brings a dawn to the pain of medical image research.However,the image segmentation is still one of the most fundamental and difficult problems in automatic image processing,analysis and recognition because of the noise,weak contrast,and inhomogeneity in medical images.Therefore,how to accurately segment the object from these complex images is an important problem in the current research of precision medical and assistance diagnostics.Inspired by problem-and algorithm-driven based research means,we propose three deformable models for medical image segmentation using the statistical information of images or objects to tackle the problems of noise,weak contrast,and inhomogeneity in the images.The related theories and methods of the models and their application in medical image segmentation are introduced in detail in this thesis.This thesis mainly has the following three works:(1)To tackle the noise and low contrast in CT liver tumor segmentation,3D deformable method guided by the image statistics of intensity information is proposed to automatic liver tumor segmentation.In order to obtain the closed,smooth and accurate tumor boundaries,the deformable segmentation model is constructed based on Non-negative Matrix Factorization(NMF)and 3D locally cumulative spectrum histogram(LCSH).For the automatic segmentation of the tumors,a Fully Convolutional Networks(FCN)is used to semantically train and segment the liver and tumor in the abdominal CT image.The results of FCN are automatically used to initialize the deformation model.All the training data and the test data for the FCN are all preprocessed using the BM3 D in order to reduce the signal-to-noise ratio of the CT image and improve the contrast between the object and the background.Experimental results on clinical data show that the proposed segmentation method is robust to noise,low contrast,and heterogeneous tumors,and achieves comopetitive segmentation performance with state-of-the-art techniques.(2)To tackle the noise,low contrast,abnormal morphology and adhesion in liver segmentation of CT images,the statistical shape model(SSM)guided liver segmentation method is proposed using statistics from shape priori,global and local image intensity.The method includes two stages: shape construction and deformation segmentation.In the statistical shape model construction,the shape of the liver is represented by the signed distance function,and the shapes are aligned based on the maximum degree of overlap.Finally,based on the PCA method,the varied pattern in the shape space is represented by the unconstrained shape coefficients.In the deformation segmentation model,the enforced locally statistical feature(ELSF)is introduced by using the locally statistical feature of the image region and the mean of the cumulative distribution function(MCDF)from a user-specified liver region.Then,the liver segmentation method combining the statistical shape prior,Gaussian global intensity analysis,and enforced local statistical feature is proposed under the level set framework.Finally,the improved CV model is used to refine the PCA based shape representation,so as to capture the long and narrow regions of the liver.Experiments on two publicly available databases showed that the utilization of multiple items of statistical information can effectively improve the accuracy of liver segmentation.(3)To tackle the noise,inhomogeneity,and texture in medical image segmentation,the shape prior and image statistics of intensity information guided global B-spline optimization segmentation method using multi-scale Gaussian statistical analysis.This multi-scale fusion approach accords with the process of visual cognition of human from the overall outline to local details.The Gaussian distribution is used to analyze the global statistical information of image grayscale,and multiple Gaussian kernel gray equalization is used to obtain the high-level shape information of the object.Under the framework of level set,the low-level and high-level information are fused and embedded into the segmentation energy functional.The functional considers the prior contour achieved at coarse scale as high-level information for both contours initialization and constraint of finer scale.Finally,cubic B-Spline basis functions are used to explicitly represent the relaxation characteristic function which contributes to fast convergence and intrinsic smooth globally optimal segmentation results.Numerical experiments on synthesized images and real images demonstrate the advantage of the proposed method.And the multi-scale fusion approach is robust to weak contrast,noise and texture.The multiple statistics in this paper mainly refer to the statistical information of the shape prior of the target,the global or local statistical information of the image grayscale,and the edge information of the image.Experiments show that the proposed methods can achieve good segmentation results in most of complex cases and demonstrate that using multiple statistical information can improve the performance of the segmentation method to some extent.