【作者】 陈允杰；

【导师】 夏德深；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2008， 博士

【摘要】 核磁共振成像(Magnetic Resonance Imaging,MRI)以其非介入性、非损伤性、很少受目标物体运动的影响等特点,已被广泛运用于医学图像拍摄,并在临床医学上起着越来越重要的作用。MR图像分割在生物医学研究和临床应用中具有重要的意义。而虚拟人计划得到的图像最真实、直观地反映了人体组织信息,该图像可以构建一个人体信息库,为医学诊断、人体结构学习等提供一个平台。图像分割可用于研究解剖结构、组织定量化测定、病灶确定、病疾诊断等。精确的分割是后继分析的关键和重要基础。根据特定的医学图像分析任务的要求,分割的目的是将原始图像划分为一系列有意义的区域或提取图像中感兴趣的区域(Region of Interesting,ROI)。脑图像分割主要分为如下几个步骤:1)图像预处理,即图像恢复。由于设备等因素的影响,导致图像中含有噪声。将噪声去除而不损失组织的信息有利于后继图像处理;2)去壳。就是去掉脑图像中非脑组织,如脑壳、脂肪等其它组织。由于非脑组织与背景部分在脑图像中含有较大比重。因此,将非脑组织与背景部分都剔除掉可以提高后继的处理精度;3)去偏移场。由于成像机制的影响,使得图像中会含有偏移场,导致图像灰度不均匀,而使得分割的结果不准确。精确的偏移场恢复模型可以很大程度地提高后继图像处理精度;4)分割出感兴趣区域。使用活动轮廓模型等方法将感兴趣区域分割出来,以对图像进行分析。本文试图将各个问题都综合到一个统一的图像分割框架下,使它具有较高的智能化程度。本文所做的主要工作和研究成果如下:(1)提出一种基于非线性扩散方程的图像去噪方法。在讨论了图像去噪的三个基本要求的基础上,总结了调和项模型和彩色总变差去噪模型中的不足,利用图像的局部信息构造函数使得模型在接近图像边缘处各向异性平滑,保持边界;在平坦区域各向同性平滑,防止阶梯效应的产生;并利用角点信息保持了角点形状。实验结果表明,本文所提出的模型能够很好地保持图像中目标的几何结构,同时具有良好的去噪能力。(2)提出一种基于高斯混合模型的几何活动轮廓模型。几何活动轮廓模型分割图像时,必须设置适当的停止项以使演化曲线运动到目标的边界时停止。该模型依赖于初始曲线的选取,而且当图像还有弱边界时曲线容易从弱边界处泄漏。本文利用高斯混合模型以及图像全局信息构造新的速度函数,使曲线在目标内部膨胀,非目标区域收缩,最终较好地达到去壳的目的。(3)提出一种基于信息熵的去偏移场模型。传统的基于信息熵的模型在求解时往往使用梯度下降法,易陷入局部最优。本文引入遗传算法、粒子群算法对其进行改进。理论上遗传算法可以得到全局最优解,但是由于种群中的个体受最优值影响较大,在一定的计算步骤内很难得到最优解。粒子群算法中,使用了全局信息,该方法收敛速度快,但该算法受种群最优值的影响,整个群体易陷入局部最优。本文使用初始信息与突变对粒子群算法进行改进,使其更易于找到全局最优点。实验证明:该方法具有一定的通用性和较好的性能。(4)提出了三种针对脑MR图像的分割方法:◆提出一种基于粒子群算法的高斯混合模型,并利用高斯混合模型可以较好地描述整幅图像性质的特点来改进活动轮廓模型,使改进的模型更适合脑MR图像的分割。高斯混合模型的关键是其参数估计,通常使用Expection-Maximization(EM)算法进行高斯混合模型的参数估计,但该算法是局部优化算法,且对初值依赖性强。为此,将粒子群算法引入到高斯混合模型的参数估计中,利用粒子群算法良好的全局优化特性来提高参数估计精度,并将高斯混合模型与传统的活动轮廓模型相结合,利用粒子群算法估计高斯混合模型的参数以获得图像的统计描述,并以此构造新的活动轮廓约束项,改善了活动轮廓模型的图像分割性能。利用图像多种信息构造新的多元信息场,使得由新的信息场构造的高斯混合模型与活动轮廓模型结合后更具抗干扰的特性。◆提出一种基于Gibbs理论的高斯混合模型。传统的高斯混合模型仅考虑了图像灰度信息,因此对噪声较为敏感。为了克服高斯混合模型的局限性,本文利用Gibbs理论和图像结构信息构造各向异性Gibbs随机场,并将其引入到高斯混合模型框架中,完善高斯混合模型分类效果,使其克服噪声影响的同时还能够保持细长拓扑结构区域信息以及角点区域信息。实验证明本文提出的算法可以得到较好的分。类结果。◆提出了一种变分耦合模型,将配准知识与曲线演化理论融合到一起。通过求解两个耦合的非线性偏微分方程使得模型的总能量达到最小值,实现两者信息的融合,同时达到配准与分割的目的。传统的基于配准模型的分割方法往往仅能对单一模态成像机制得到的图进行配准,而且针对不同人的脑中的脑室、海马体等组织结构差异较大的情况,很难得到较好的分割结果。将基于信息熵的非线性配准与基于图像全局信息与局部信息的曲线演化模型相结合,提出耦合模型,使用变分方法进行求解,实验结果表明本文提出的算法可以得到较好的分类结果。(5)提出了三种虚拟人脑图像的分割方法:◆提出一种基于HSV颜色空间的虚拟人脑图像分割方法。在HSV颜色空间对其进行分析,提出一种改进的各向异性扩散方程,并构造混合信息场,以降低噪声、过渡区域等因素的影响。使用ostu算法与一种新颖的快速符号表算法对饱和度信息场与色度信息场进行分类,从而得到灰质分割结果,并利用解剖学知识、区域信息以及数学形态学知识对亮度场信息进行分析,以修正分割结果,最终将脑组织分离出来,实验结果表明本文算法能较精确地得到分割结果。◆提出一种多项位水平集模型。传统的C-V模型仅能够分割单一目标,本文将传统的C-V模型扩展到多水平集,使其同时可以分割多个目标;使用PCA理论将颜色空间投影到新的空间中,以扩大灰质与下层数据之间的颜色距离,减低下层数据的影响。同时使用局部信息校正颜色强度不均匀。将距离约束项引入到模型中,使模型能够无需重新初始化,提高了演化速度。实验结果表明改进的算法能较精确地得到分割结果。◆提出一种各向异性Mean Shift算法。FCM方法、高斯混合模型等统计方法在分割图像时需要知道图像中含有多少类别,而Mean Shift算法进行分割,该算法不需要先验知识就能够将图像分割好。本文分析了Mean Shift算法本质,发现该模型是一个各向同性模型。使用结构信息构造各向异性高斯核,使Mean Shift算法具有各向异性,从而克服细长目标的影响,并使用矩阵分解方法将高斯核分解,加快了模型计算速度;将颜色空间投影到新的坐标系下,使得相近颜色可以有较大的距离,以增大虚拟人脑图像中灰质与下层数据之间的区别,并从理论上证明了Mean Shift算法本质类似于EM算法,本文算法收敛到精确解。虚拟人脑图像分割结果说明本文算法可以得到较好的分割结果。更多还原

【Abstract】 The application of magnetic resonance imaging, with the characteristics of no intervention、not harmful、seldom effected by the motions of objection, has been used in taking pictures of medical images. Medical image segmentation plays an important role in biomedical research and clinical applications such as study of anatomical structure, quantification of tissue volumes, localization of pathology, diagnosis, treatment planning, and computer aided surgery, etc. As a result, accurate segmentation method is crucial to the follow-up analysis.According to different image analysis task, medical image segmentation aims at partition the original image into several meaningful regions or isolating the region of interesting (ROI). Variational method could naturally convert complex segmentation into a variational functional optimization problem. In this thesis, variarional method-based medical image segmentation for specific tasks is extensively explored, and efficient numerical algorithms are discussed.The application of magnetic resonance imaging, with the characteristics of no intervention、not harmful、seldom effected by the motions of objection, has been used in taking pictures of medical images. Chinese Visible Human images can give the true information about human tissues. These images segmentation plays an important role in biomedical research and clinical applications such as study of anatomical structure, quantification of tissue volumes, localization of pathology, diagnosis, treatment planning, and computer aided surgery, etc. As a result, accurate segmentation method is crucial to the follow-up analysis.According to different image analysis task, medical image segmentation aims at partition the original image into several meaningful regions or isolating the region of interesting (ROI). Variational method could naturally convert complex segmentation into a variational functional optimization problem. In this thesis, variarional method-based medical image segmentation for specific tasks is extensively explored, and efficient numerical algorithms are discussed.The segmentation of the brain images can be separated into a few steps: 1) image denoising. Because of the effect of ficilities, the images have noise in images. Denoise the image can make the segmentation more exactly; 2) Skull stripping. In the brain images the skull and other tissues take a large part, so a exact skull stripping method can make the segmentation more exact; 3) de-bias. The bias field in images can make images inhomogeneous and hard to segment; 4) segment the images. Use methods such as active counter models to find region of interesting. This paper wants to form a union image segmentation framework, which has a high intelligent ability in order to solve some complex image segmentation problems. The primary work and remarks of this paper are as follows:(1) A new anisotropic diffusion based image denoising method is proposed by analyzing the traditional important denoising models: harmonical model, CTV model. At first, three requirements of image denoising are proposed. Using the structure information, the new model can anisotropic diffuse the image in the edge region and isotropic diffuse the flat region. In order to contain the corn region, the corn information is added to the new model. Experimental results show that the new denoising method is capable of sufficiently preserving geometric information such as edges and corners in addition to its effectiveness for image denoising(2) Traditional Level Set method can not get prefer results for it only depend on the gradient information. In this paper a new speed function is presented which is based on anisotropic diffusion function, Gaussian mixture model and global information. With the new speed function the adapted level set model can get better results. The experiments to skull stripping the brain MR images show that the method of this paper can get better results in an accuracy way.(3) We propose a new de-bias model based on entropy method. The best bias will make the image has the smallest entropy, but it need to find a lot of parameter to compute the bias. The traditional method uses the gradient descending method to find the parameters. The method plunges in local best easily. In order to deal with this problem, genetics algorithm (GA) method, Particle swarm optimization (PSO) method are analyzed and an adapt method is present to get the global best result. The experiments show that the new method can get accurate result robustly.(4) This paper presents three brain MR images segmentation models:◆Gaussian mixture model can approach the probability of the image’s histogram. Active contour models can be improved with the Gaussian mixture model to be more fit with the segmentation of medical images. But the estimation of the parameters of the model is hard and usually based on the Expection-Maximization (EM) method. Btu the method is local best and sensitivity to its initial parameters. In order to get better results, we use PSO model to compute the parameters. With this parameter, we construct a new constrain force and with this new force the model can get better results. After some experiments, we found that the Gaussian mixture model only uses the information of the histogram and not uses the information of the location of the pixel. So it is sensitive to the noise. In this paper, we give a method to make a new information field, which combines the information of the region, texture and region simulation. With the new information field the Gaussian mixture model can reduce the effect of the noise. In this paper the Gaussian mixture model be introduced to the Level set model and reduce the effect of the noise and prevent the curve over the weak edges. After get the inner edge of the left ventricle, this paper uses the region and shape information to segment the out edge. Experiments on the segmentation of brain magnetic resonance images show this model has better effect in image segmentation.◆In order to overcome the limitation of Gauss mixture model, this article uses the Gibbs theory and the image structure information to construct anisotropic Gibbs random field and to incorporate it into the Gauss mixture model. The new Gauss mixture model can reduce the effect of the noise and contain the information of beam structure regions and corner regions. Experiments on the segmentation of brain magnetic resonance images show this model can attain better effect in image segmentation.◆We introduce a new coupled variational model, which can registration and segmentation simultaneous. In the model a couple function is constructed to fuse the non-rigid registration information and the active contour model, which based on the region information. Using this information an energy function is constructed. Through finding the extremum of the energy function the model can realize registration and segmentation simultaneously. The model can be applied to analyze the images which from different modals. The results of experiments show that the model can get better results robustly.(5) This paper presents three Chinese Visible Humane brain images segmentation models:◆We analyze the images in HSV color space, which can distinguish different areas in brain clearly. In this paper a new fuzzy anisotropic diffusion functions is presented, which can diffuse the images meanwhile preserve the curves. During eliminating other apparatus from brain, the saturation information, hue information, value information, anatomy information, and region information are fused to confirm the results correctly. The experiments to segment the digital human brain images show that the method of this paper can get well results in an accuracy way.◆C-V model is one of the best segment method, but the classical C-V model only segment the image into object and background; only use the intensity information when segment color images; must re-initial the distance function during evolve the curves. In the CVH images, there are many fake grey matters and with the effect of these fake matters the C-V model can hardly separate grey matters with fake grey matters. To deal with the problem the C-V model is adapted to be able to segment more objects; PCA model is presented to large the difference of grey matters and fake grey matters. With the effect of tissues themselves, there are many in-homogenous phenomenons in the CVH images; the local information is added to model to reduce these effects. Use the distance resistance energy, the model can evolve curves without re-initialization. The Chinese visual human brain images segmentation experimental results show that the method of this paper can get right results in an accuracy way.◆In order to overcome the limitation of the Mean Shift method, this paper presents a new anisotropic Gauss kernel, which based on structure information, and with the new Gauss kernel the new model can reduce the effect of gracile topological structure. This paper projects the color space to a new space, based on PCA model, to expand the distance of similar color and make more difference between grey matters and grey matters belonging to next picture. The results of the segmentation of the digital brain image show that adapt Mean Shift method can get better results.更多还原