节点文献
基于脑MR图像的三维组织自动分割
Automatic and Three-dimensional Tissue Segmentation of Brain MRI
【作者】 李伟;
【导师】 陈武凡;
【作者基本信息】 南方医科大学 , 生物医学工程, 2009, 博士
【摘要】 近几十年来,医学图像极大地影响了神经科学的许多领域。随着先进的医学成像技术的发展,许多神经科学的研究放在了比较脑内组织解剖结构的差异上,从而寻求与脑疾病有关的解剖结构形态改变的特征,以期提高脑疾病诊断的可靠性和治疗方案的有效性。医学图像分割作为图像分割领域中的一个重要分支,是实现医学图像分析,进而完成医学图像理解的首要、关键性步骤。随着磁共振成像(Magnetic Resonance Imaging,MRI)技术的发展,磁共振(MR)图像可以提供脑内部组织解剖结构的高对比度和高分辨率的三维(3D)医学图像。神经科学研究人员逐渐地对能将脑精确区分为三大主要组织,灰质(GM)、白质(WM)和脑脊液(CSF),进而将脑分割为皮层结构、皮层下结构和病理组织的方法产生浓厚兴趣。这些基于解剖学形态结构改变的研究均依靠对MR图像的分割。而医学图像分割技术就提供了这种从多模式医学图像中分割和提取出各种脑组织结构的自动和半自动的方法。然而,医学图像分割是医学图像分析中最困难和最具有挑战性的问题之一。由于MR成像设备成像能力的限制,临床采集的脑组织的MR图像通常含有噪声、偏场(Bias Field,BF)导致的灰度不均匀(Intensity Non-uniformity,INU)、部分容积效应(Partial Volume Effect,PVE)和运动伪影等不利因素,加之脑组织复杂的形状、边界和拓扑结构,使得快速、准确和鲁棒地分割脑组织MR图像是一件困难的事。此外,二维(2D)图像分割已不能满足临床和研究的需要,脑组织3D图像的分割逐渐成为主流,临床医生和研究人员迫切需要快速、准确和鲁棒的3D分割算法。因为人体组织结构毕竟是三维结构体,而且3D分割充分利用了当今成像设备采集的3D图像数据的信息,其分割结果在空间上更加准确和连续,提供给研究人员更丰富的人体组织的3D形态结构、大小、位置等信息,显示也更加的直观明了。近几十年来,针对图像分割领域的相关算法虽然种类繁多,且仍层出不穷,但依然无法完全满足人们的实际需求。其原因相当复杂,包括:无法完全用数学模型来简单描述人们所面临的实际问题;分割对象结构性质的千差万别;图像退化以及人们对分割结果预期目标互不相同等。这些原因决定了不可能实现一种普适、通用的分割方法。只能针对特定问题和具体的需求给予合理选择,在精度、速度、和鲁棒性等关键性指标上做出均衡或侧重。针对目前医学图像分割领域的发展现状,本文从脑内组织MR图像三维分割的角度出发,分析并回顾当前主要的医学图像分割方法,特别是三维医学图像分割算法。针对这些算法中的不足,提出了一些新的模型和3D分割算法;并利用这些新算法和已有的算法,在三个分割层次上,来实现在临床和神经科学研究中对脑组织的分割。随着研究人员对脑和脑疾病的研究不断深入,脑组织的分割已经被分为三个层次。第一个层次是将脑组织分割为三大脑组织,灰质、白质和脑脊液;第二个层次是将脑组织进一步分割为皮层结构和皮层下结构;第三个层次是病理组织的分割和提取。这三个层次不是互相独立的,而是有机的联系在一起。如第一层次的分割结果有助于其它两个层次的分割。本文包括以下的研究工作和创新:第一、由于脑部MR图像存在图像偏场和噪声,本文提出了一种基于象素灰度值的改进FCM自适应快速自动分割算法,来完成对含有偏场和噪声的脑部MR图像进行快速3D分割。在该算法中,给出了一种新的分割目标函数,采用参数模型来近似偏场和类似马尔可夫随机场先验的模糊隶属度矢量邻域约束来模拟脑组织分布的空间一致性。该算法不需要对MR数据取对数或滤波等预处理,在目标函数递归优化的过程中,利用偏场参数模型和邻域约束来同时完成象素的分割和图像偏场的估计。由于算法利用分割结果估计偏场,使得偏场的估计更加合理和准确。同时参数模型减少了需要估计参数的数目,提高了算法分割结果的准确度和分割的速度。模拟和临床脑部MR图像的分割实验结果表明,我们的算法对初始值不敏感,对噪声有较强的抑制,有效地克服了偏场的影响,分割结果准确度高,速度快。我们的算法在三维MR图像的自动分割实验中取得了满意的分割速度和准确度。第二、在本文中,我们提出了一种基于多约束和动态先验的自动三维分割算法,称之为MCDPMRF-EM算法。该算法将来自MR图像的一种大尺度约束引入MRF模型中,在贝叶斯框架和最大后验准则下,利用改进的EM算法,实现MR脑图像的分割。MCDPMRF-EM算法具有分割准确、鲁棒和分割结果与解剖学一致性高的优点。我们的算法有以下创新点:(1)提出多约束模型和构造方法,并利用构造的多约束来提升分割算法的性能。算法对分类数不敏感,在不同分割数目设置下,算法的分割结果具有解剖学先验知识一致性。(2)提出动态先验的概念来模拟人眼分割图像的自适应特点,使得先验的作用根据具体的待分割图像自适应调整,有效地克服了图像偏场。(3)利用参数模型来模拟由于图像偏场存在而导致的同种组织灰度的变化,避免对图像灰度值的对数变换和由此导致的灰度值统计概率不再符合高斯统计模型假设的缺陷。(4)提出一个新的EM改进优化算法。在高斯马儿可夫随机场模型和偏场模型的假设下,该算法可以快速、准确和鲁棒地求解分割目标函数的最优解。第三、中脑黑质的病变及多巴胺能神经元功能的破坏是帕金森病(PD)的主要原因。通过对黑质精确的三维分割,来获得其位置、体积和3D形状,再对PD病人和正常人、PD病人早期和晚期以及PD病人治疗前后的黑质形态学上的比较,测量黑质形状和体积的统计变化,有利于提高对早期PD的诊断和评价治疗效果。本文提出了一种基于动态曲面模型和解剖先验知识为约束的自动3D分割方法,该方法能够精确地提取黑质的3D形状结构。而且,就我们所知,黑质的3D分割还未见报道过。第四、MRI图像肿瘤的自动分割具有相当大的难度,因为肿瘤及其周围组织的表现和外观比较复杂。在临床上,医生和分割算法间的一些交互信息会极大地提高分割算法的分割结果的准确性和针对性。基于此,我们提出了一种基于图论最大流/最小切准则的交互式半自动3D肿瘤分割算法。该算法通过简单的交互,就能够快速、准确地分割出脑肿瘤。第五、人类大脑的脑室系统是由四个相互连通的脑室构成,脑室内脑脊液体积容量的变化和脑室形态的改变与多种神经性疾病有关联,脑室系统体积形态变化的量化研究对诊断各种脑部疾病,评价治疗效果和对疾病发作和结果的预测都有重要作用。本文提出了一种准确,快速,稳健的自动三维人脑脑室系统的混合分割方法,能够自动地从脑部3D MR图像中提取出整个脑室系统。该方法由以下两个算法串联构成:(1)基于多约束和动态先验马儿可夫随机场模型(MRF)和最大期望(EM)优化的三维自动分割算法(MCDPMRF-EM算法),MCDPMRF-EM算法将3D图像分割成5种组织类型,同时估计图像偏场。由于MCDPMRF-EM算法采用多约束和动态先验MRF模型,同时用参数模型表达MR图像偏场,本文方法具有偏场校正、抑制噪声的能力,并且算法对不同分类数目,其分割结果一致性高。(2)基于高斯马尔可夫随机场模型(GMRF)分割目标函数和s/t最大流—最小切图论优化的快速3D分割算法。前一个算法用来将脑组织分为白质(WM)、白质/灰质(WM/GM)、灰质(GM)、灰质/脑脊液(GM/CSF)和脑脊液(CSF)五种组织类型;后一个算法利用前一个算法的结果来进一步分割出整个脑室系统;两个算法之间通过形态学方法来联系。两个串联算法均在MR图像的三维空间中进行分割,利用三维空间的参数偏场模型和三维空间约束信息。该混合方法利用像素的部分容积效应和图论网络流中最大流—最小切优化分割算法的解偏小的特点,来实现各脑室不同部分的种子像素点提取和带硬约束的最大流—最小切优化方法的脑室自动分割。本混合方法不需要对MR图像进行去噪,偏场校正等预处理,就能获得与解剖学知识一致的优质分割结果。
【Abstract】 In the last two decades, the field of medical image analysis has greatly influenced many areas in neuroscience. With the advancement of the medical imaging technologies, neuroscientists have been increasingly interested in methodologies that can identify brain normal tissues, subcortical structures and pathological tissues in anatomical imaging modalities. Many neuroscience studies aim to find new disease related anatomical characteristics in order to increase the reliability of diagnosing the illness or improving the effectiveness of treatment methods against the brain disease. As the one of the most important branches of segmentation of image, the segmentation of medical image is the primary and critical step for the analyzing and understanding the medical images.Today, with the advancement of the Magnetic Resonance Imaging (MRI), the MRI has provided a means for imaging tissues in the brain at very high contrast and resolution in the three dimensional space. Most neuroscientists are keenly interested in outlining the three main brain "tissue" classes - cerebral spinal fluid (CSF), white matter (WM) and gray matter (GM) - in Magnetic Resonance (MR) images and further parcellating these tissue classes into their substructures such as cerebral ventricles, thalamus and so on. These anatomical studies of brain tissues and structures with disease are often based on the analysis of Magnetic Resonance images. The segmentation of medical images provides different kinds of algorithms or tools to segment and extract the tissues and structures in the brain automatically or interactively from MR images for the analysis.The task of automatically segmenting medical images is challenging as the images are corrupted by several artifacts. Because of the limited resolution and imperfection of the medical imaging devices, the sampled MR images from clinic are often degraded by noise, bias field (BF, also known as intensity non-uniformity, INU), partial volume effects (PVE) and motive artifacts. In addition, the complex shape, boundary and topology of brain tissues and structures make the accurate, fast and robust segmentation of brain tissues very difficult. Furthermore, two-dimensional (2D) segmentation of medical images can not meet the demands of clinic and research anymore. Three-dimensional (3D) segmentation and visualization of medical images becomes popular which is because of the 3D nature of the tissues and structures of brain in reality. Moreover, the 3D segmentation offers more accurate and continuous results utilizing the more rich information provided by 3D medical imaging data volumes as much as possible. The visualization of objective structures is more vivid with rich relevant 3D information about shape, size and location.By far, the field of medical image analysis has developed a variety of automatic segmentation methods to fulfill the difficult problem of medical segmentation. However, the current medical segmentation algorithms still can’t satisfy the various demands in clinic and research completely. The reasons causing the above deficient state of the field include poor mathematic model for describing the problem faced in practice, significant difference between different targets to be segmented, degraded medical images due to the imperfectness of the imaging devices, random and complex change of pathological tissue, diverse expectances of the result, and so on. So, there is no such an algorithm of segmentation which is competent for all kinds of the problems. What we can do is to develop different algorithms of segmentation for different problems.In this thesis, the current algorithms of medical image segmentation are reviewed in detail especially the 3D segmentation algorithms for brain tissues. Some new models and algorithms are proposed to fulfill the accurate and fast 3D segmentation of brain tissues and structures on the following three levels. As researchers have been furthering the studies in brain and the related diseases, the segmentation of brain tissues has been stratified into three levels: the segmentation of three main tissues of brain, the segmentation of the subcortical structures and the segmentation of pathological tissues. We believe that the three levels are interdependent in automatic parcellation of brain. In the thesis, the following creative research works are included:First, an new intensity based improved FCM algorithm is presented to fast segment the MR brain volumes with significant bias field and noise into three main brain tissues. The algorithm is formulated by proposing a new objective function based on standard FCM algorithm with bias field correction and neighborhood constrain. In the algorithm, a parameterized model is adopted to express the bias field and a neighbor constrain on membership vectors similar to Markov random field (MRF) is proposed to express spatial consistency of brain tissue. The proposed algorithm segment MR data volumes and estimate the bias field without need for a logarithmic transformation and preprocessing. Experimental results with both synthetic and real clinic data are included, as well as comparisons of the performance of our algorithm with that of other published methods. The validation of the algorithm shows good accuracy and fast convergence. Second, in the paper, a new multi-constrain and dynamic prior based automatic 3D Segmentation algorithm called MCDPMRF-EM is developed for segmenting MR Brain volumes into three main brain tissues. A novel big scale constrain extracted from MR volumes is introduced into Markov Random Field model (MRF) in the algorithm. The algorithm searches the optimal segmentation configuration using the Maximum a Posteriori (MAP) criteria and a modified expectation-maximization algorithm (EM) in the Bayesian frame. The MCDPMRF-EM algorithm segments MR brain volumes corrupted by bias field and noise accurately and robustly. The results of the algorithm are more consistent with the known anatomical facts. The proposed algorithm incorporates the following novel features.1)The multi-constrain model is proposed as well as its creation to improve the function of statistical segmentation algorithms. By incorporating the new model, our algorithm is insensitive to class number, and the segmentation result of our algorithm is more consistent with the known anatomical facts.2)The dynamic prior concept is also proposed to simulate the self adaptive function of human eyes. The dynamic prior can automatic adjust the constrain to effectively conquer the bias field.3)We propose parametric, smooth models for the intensity of each class instead of multiplicative bias field that affects tissue intensities. This may be a more realistic model and avoids the need for a logarithmic transformation and, hence, the related nonlinear distortions.4) We propose a novel variant of the EM algorithm which allows for the use of a fast and accurate way to find optimal segmentations, given the intensity models which incorporate MRF spatial coherence assumptions.Third, the MRI-based quantity analysis of substantia nigra (SN) in human brain has more and more value in diagnosis of Parkinson disease in today. We describe an anatomic knowledge-constrained algorithm based on active surface model and adaptive region growth to automatically delineate the SN region from a magnetic resonance image. The result of the algorithm can be used to calculate position, shape and volume and help early clinical diagnosis as well as treating effect of SN. The validation of the algorithm was tested and showed good accuracy and adaptation.Fourth, we propose a graph-based three-dimensional (3D) algorithm to automatically segment brain tumors from magnetic resonance images (MRI). The algorithm uses minimum s/t cut criteria to obtain a global optimal result of objective function formed according to Markov Random Field Model and Maximum a posteriori (MAP-MRF) theory, and by combining the expectation-maximization (EM) algorithm to estimate the parameters of mixed Gaussian model for normal brain and tumor tissues. 3D segmentation results of brain tumors are fast achieved by our algorithm. The validation of the algorithm was tested and showed good accuracy and adaptation under simple interactions with the physicians. Last, the human cerebral ventricular system consists of four inter-communicating chambers. Changes in CSF volume and ventricular shape are associated with several neurological diseases. Quantification of the degree of abnormal enlargement of ventricles is important in diagnosis of various diseases, measuring the response to treatment, and predicting the prognosis of the disease process. In this paper, we present a novel 3D hybrid approach for automatic extraction of human cerebral ventricular system from MR neuroimages. The approach consists of following two algorithms and intermediate step serially. First, an new 3D algorithm of segmentation called MCDPMRF-EM based on multi-constrain and dynamic prior with Markov Random Field model, which optimized by a Maximum A posteriori Probability criteria and the expectation maximization algorithm. The algorithm segment brain tissues into five tissue class types and estimate bias field (BF) accurately and robustly from MR image volumes. Second, an intermediate step of some morphological processes is followed with the above algorithm to extract the seed regions which are the parts of the four ventricles. Thirdly, we use a combinatorial s/t graph cuts algorithm with the hard constraints to segment the ventricular system from MR neuroimages. Our approach can automatically extract the complete ventricular system with no need for denoise and correction of bias field. The test result of the approach is accurate and agrees with the anatomical structure of the ventricular system.
【Key words】 Brain tissues; Magnetic resonance imaging; Three-dimensional segmentation; Bias field; Markov random field; Graph theory;