【作者】 郑倩；

【导师】 陈武凡； 卢振泰； 冯衍秋；

【作者基本信息】 南方医科大学 ， 生物医学工程， 2014， 博士

【摘要】 随着X线、计算机断层扫描(computed tomography, CT)、磁共振成像(magnetic resonance imaging, MRI)、正电子发射断层成像(positron emission tomography, PET)等成像技术的问世和发展,影像诊断在医学中的作用也日益增大。成像技术使得临床医生可以更加直观和清晰地观察正常和病变组织的解剖结构和功能代谢变化,为临床诊断、制定治疗方案和生物医学研究提供了强有力的科学依据。随着医学图像数据的增多,研究人员和临床医生迫切需要利用计算机自动进行医学图像的处理和分析。医学图像分割,指把图像中的解剖结构或者感兴趣区域的边界勾画出来,是进行图像分类和识别的前提。医学图像分割对三维可视化、三维定位、组织定量分析、制定手术计划和计算机辅助诊断具有重要的意义。医学图像分割方法分为人工分割、半自动分割、自动分割方法这三种。人工分割指临床专家借助特定的软件勾画图像边界,这种方法是十分耗时的,并且依赖于临床专家的知识经验等主观因素,可重复性较差,不能满足临床上的实时需要。半自动分割方法是指将图像分析技术和个人经验结合起来,人机交互提取目标边界,这种方法在一定程度上提高了分割速度,但仍然依赖于观察者,限制了其在临床实践中的应用。自动分割方法是指完全借助计算机对数据进行处理分析,提取出感兴趣区域边缘,这种方法完全避免了观察者主观因素的影响,提高了处理数据的速度,可重复性好。近年来,研究人员一直致力于研究新的半自动或者自动医学图像分割方法,以处理不同的研究对象,满足临床上的不同需求。本文主要针对磁共振(MR)图像,在脊椎MR图像中提取椎体,脑部MR图像中提取肿瘤和心脏黑血MR T2*测量中提取心肌组织方面上做了以下研究：一、本文提出了一种基于邻域信息和高斯加权卡方距离的图割算法,实现了从矢状面脊椎MR图像中分割椎体。面对社会老龄化的趋势和社会工作压力日益增大,脊椎各种疾病如椎骨退化、椎间盘突出、椎管狭窄症等,困扰着各年龄阶段和各种职业的人群,已经成为影响公共健康的几大顽疾之一。脊椎毗邻结构上下交错,因此外科医生需要具有良好的方位感,精确知道椎体的形状,对椎体形状的识别是提高手术精度、降低手术风险的关键。描述人体脊椎生理解剖结构的常见成像方式有CT和MRI。CT图像中分辨率高,骨骼、皮肤等不同类型组织之间的灰度差异较大,用区域生长法可以很好地分割出脊椎CT图像中的椎体。而MRI是一种多参数、多方位成像技术,利用强磁场和射频脉冲来产生图像,可以清晰地观察脊椎的解剖结构。当图像灰度主要由T1弛豫时间决定时,则为T1加权成像；当图像灰度主要由T2弛豫时间决定时,则为T2加权成像。T1加权图像可以很好地辨别脂肪和水,解剖结构清晰,脂肪呈白色高信号,水呈黑色低信号；T2加权成像水信号比较亮,脂肪是灰白的,脑脊液呈白色信号。由于软组织对比度变化和射频场非均匀性等诸多因素影响,从矢状位脊椎MR图像中分割椎体非常具有挑战性,目前关于从MR图像中提取椎体的文献还很少。Gamio等人在2004年将图论的方法应用到脊椎MR图像椎体分割中,图论是一种无监督的图像分割技术,它不需要任何的初始化,将图像分割问题转化为图割问题,该方法没有将图像划分为目标和背景两个区域,而是划分为几个互不相交的区域(类)。基于图论的图像分割方法其基本原理是将图像中的每一个像素都看作无向图中的一个节点(顶点),两个像素之间的连接构成了图中的边,像素之间的相似性就是图中的权重,定义了一个全局优化准则函数,使得各类类内相似度最大,类间相似度最小,利用求解特征值和特征向量的方法完成图像分割。Gamio算法中利用了局部直方图来计算像素之间的权重,没有考虑图像的空间信息,当图像块之间的统计信息相似时,容易造成错误的分割结果。因此,本文提出了一种自动的基于邻域信息和高斯加权卡方距离的脊椎MR图像椎体分割方法。本文的主要贡献是：(1)由于成像过程中存在噪声和各向异性的影响,只考虑单个像素的灰度值特征对噪声敏感,为此本文采用5×5窗口提取以每个像素点为中心的邻域内的空间-灰度特征,该特征对噪声具有较强的鲁棒性,将图像中像素的局部邻域信息以向量的方式排列起来；(2)考虑到邻域内各像素对中心像素的影响不同,权重也应该不同,利用高斯核函数将这些特征向量以卡方距离的方式组合起来；(3)我们引入局部收缩的概念,图像中所有像素不再用一个单一的尺度参数,而是用自适应的局部收缩方式为每一个像素自动赋予一个尺度参数。(4)将构造的相似度矩阵引入基于图论的分割方法中,完成提取椎体的任务。本文所提出的方法不需要反复调节尺度参数以获得令人满意的分割结果,高斯核函数组合的邻域信息可以更好地描述图像,克服了噪声的影响。为了验证本文算法的可行性,利用来自于广州南方医院3.0T磁共振设备扫描的100幅脊椎MR T1加权和T2加权图像进行评估,实验结果表明,本文提出的新算法克服了传统方法中常见的过分割和欠分割现象,分割的正常和退行性改变椎体光滑且清晰,与椎体真实边缘比较接近,准确率比较高；对带有噪声的脊椎图像提取椎体时,结果也令人满意,说明本文方法具有鲁棒性强的优点。准确地椎体分割结果可以用于脊椎变形的图像配准、分析中,同时,分割结果可辅助临床医生完成脊椎矫正手术。作为一种一般性的分割方法,该算法可以拓展到其它器官的分割中。二、本文提出了一种基于熵和局部邻域信息的高斯约束Chan-Vese(CV)模型,实现了从脑部MR图像中提取肿瘤组织。据报道,脑瘤是一种十分常见的脑部疾病,每年都有400000余人被诊断出脑瘤。脑部肿瘤的分割对三维可视化、计算机辅助病理诊断、制定手术计划至关重要,然而,脑部肿瘤在位置、形状和外观上多种多样,个人差异性也比较大,因此从MR图像上分割脑部肿瘤相当困难。活动轮廓模型是一种比较受欢迎的图像分割方法,它利用一条闭合的曲线提取图像中目标的边界,能量函数获得最小值时,曲线演化到感兴趣区域边界。传统的CV模型利用区域同质性划分图像,而MR图像具有灰度不均匀性,CV模型不适合分割MR脑部肿瘤图像,因此本文提出了一种基于熵和局部邻域信息的高斯约束CV模型,完成了脑部肿瘤的分割,其主要贡献如下：(1)针对CV模型中参数不宜设置的问题,我们利用熵构造内部和外部区域能量的权值系数,加强了对演化曲线的控制；(2)同时将曲线上各点的局部邻域信息引入到曲线演化过程中,降低了曲线内外区域灰度不均匀等因素对演化曲线的影响,提高了分割的准确性；(3)最后,高斯约束保证了曲线演化过程中的稳定性、光滑性,同时不需要曲线长度约束项和重复初始化。为了验证本文算法的可行性,我们利用来自天津医科大学附属医院的100幅脑部MR图像进行实验,实验证明,本文所提方法可以准确地提取脑部肿瘤区域。三、本文首次提出了一种全自动的心脏黑血MR T2*测量框架,实现了对心肌组织的提取。地中海贫血病是全球最大的单基因遗传病之一。地中海贫血患者需要长期输血使其生长发育接近正常和防止骨骼病变,改善他们的生活并延长他们的生命,然而长期输血容易造成铁血黄素沉着症,损伤肝脏、心脏和胰腺组织,导致病发率和死亡率的增加。临床上,一般是给予铁螯合剂治疗以便去除身体内多余的铁,但是由于其化学毒性,治疗时会伴随着并发症。准确地测量和控制人体的铁含量对诊断铁过载病人和评估铁螯合剂治疗具有重要的意义。心内膜心肌活检是测量心脏铁含量的一种方法,但是其具有并发症的侵害性及由于心脏内铁分布不均匀所导致的测量不准确性。随着数字图像的发展,基于MR的参数测量技术已经成为一种评估心脏铁含量的方法,其优势是对人体无侵害性损伤。T2*是T2弛豫时间的另一种表达形式,其对铁产生的场强变化更加敏感。铁的存在会破环磁场,缩短弛豫时间,这些量化的改变可以经过校准曲线转变成铁含量,所以组织的铁含量可以直接由心肌细胞中氢原子的弛豫时间来决定。研究人员发现在临床中可以用短轴心室图像的一个感兴趣区域的T2*值作为心脏铁含量的量化指标。通常人们选择室间隔作为感兴趣区域来评估心脏铁含量,这主要是因为室间隔区域不容易受到来自于脂肪、肝脏、和肺等组织磁敏感性的影响。该方法会计算出一个衡量全心脏铁沉积的T2*值,正常人的T2*值范围是52±16ms,中值是40ms,不低于20ms。在MR图像中,血液既可以表现为明亮区域也可以表现为灰暗区域,这主要取决于血液的流动性和MR信号的获取方式,一般称为亮血成像技术和黑血成像技术。研究人员发现利用屏气状态下MR亮血成像技术的T2*弛豫时间来分析心脏铁沉积具有很好的可重复性。与亮血成像技术相比,黑血成像技术避免了血液信号的污染、压缩了血液流动信号减少了伪影,具有更加清晰的心肌边缘。为了自动测量心脏铁含量,我们首次提出了一种全自动心脏黑血MR T2*测量框架,实现了心肌组织的提取。本文的主要贡献如下：(1)引入圆形Hough变换和带约束的模糊聚类算法对左心室内外膜进行定位,检测到的圆作为改进的CV模型的初始轮廓,避免了手动设置初始轮廓及初始轮廓位置对分割结果的影响,由于检测到的圆位于左心室附近,因此也减少了演化曲线向目标边界移动的时间；(2)利用本文提出的基于熵和局部邻域信息的高斯约束CV模型,分割左心室内外膜；(3)引入阈值分割和连通区域标记实现右心室血池区域提取。根据已分割的左心室内外膜,得到左心室血池,因为心肌组织与血池对比度明显,利用阈值分割方法,获得包含右心室血池区域的图像,阈值由左心室血池灰度自动确定,利用解剖位置关系,我们用连通区域标记及提取的方法得到右心室血池；(4)利用左心室和右心室血池的解剖位置信息,检测左、右心室相交区域的底端和顶端,提取出室间隔；(5)利用单指数拟合模型进行T2*测量。(6)用Bland-Altman分析法分析两种方法偏差的均值和标准差,验证算法的可重复性,利用变异系数来评估两种方法之间的一致性。我们提出的自动MR T2*测量方法不需要人为参与,不仅避免了观察者的主观差异,并且提高了数据处理效率。为了验证本文方法的可行性,我们利用来自于伦敦皇家布鲁顿医院的1.5T磁共振设备扫描的144个地中海贫血症病人的心脏黑血MR梯度回波序列图像进行评价,算法执行时间仅为1.4s,该算法与人工T2*测量结果的变异系数仅仅为4.15%(临床上,小于10%就认为两种算法一致性较好)、偏差均值是1.71%,实验结果验证了所提算法的可行性,虽然所提算法比人工方法高0.18ms,但是这对于临床专家对地中海贫血症患者铁含量分级并没有影响。因此,我们提出的全自动心脏黑血MRT2*测量方法可重复性好,可以辅助临床医生对地中海贫血症患者进行诊断和治疗。更多还原

【Abstract】 With the development of computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET) and other imaging technologies, diagnostic imaging is a critical component in clinical medicine today. These modalities have greatly improved knowledge of normal and diseased anatomy and functional metabolism of the anatomy for medical clinicians and provide a valuable tool for clinical diagnosis, making aggressive treatment planning, and numerous biomedical imaging researches. With the increasing size and number of medical image data, more and more scientific researchers and clinicians make contribute to facilitating the data processing and analyzing by the usage of computers. Medical image segmentation, the delineation of anatomical structures and other regions of interest, is the main key for patterns analysis and clustering as well as plays an important role in3D visualization, quantitive analysis on tissue characterization,3D localization and making surgery planning.Medical imaging segmentation consists of three categories:manual segmentation, semi-automated segmentation and automated segmentation. The manual segmentation is performed by the clinical experts using the designed tool for delineating the boundaries of the object, and this method has good accuracy but is prohibitively a labor intensive, tedious and time-consuming task, not meets the real-time requirement in clinical practice. Semi-automated segmentation algorithms make full use of the computers to extract the boundaries of the object by taking the knowledge of the experienced experts into account, which limits the application in clinical practice. Researchers and clinicians pay attention on developing automated segmentation methods, which facilitate to process the image data by taking advantage of computers and detect the boundaries of the object automatically. They could avoid the subjectivity from the worldwide operators and improve the efficiency of data processing as well as good reproducibility. The automated or semi-automated algorithms are highly desired and developed to deal with the various objects and solve the specific problem.We have developed the automated segmentation methods of the MR images in this paper. The main contributions of this paper are listed as follows.First, we present a novel method for the automated segmentation of the vertebral bodies from2D sagittal MR images of the spine based on local spatial information and Gaussian weighted chi-square distance.Due to the aging of the society and the increasing of work pressure, a large population has the spondylopathy, such as pleurapophysis, slipped disc and spinal stenosis, which influence lives of people in various trades. Spinal anatomica structure and circum-structure of the vertebrae are very complicated, thus the clinician must has good localization and good knowledge of the vertebrae. The recognition of the vertebral body is the main key of improving efficiency, accuracy and security of the operation in clinical practice.Physicians often make their decisions in diagnosis and treatment of spine with the help of CT and MR images. CT images show reasonably high resolution and give good visualization of the bone, and the vertebral body segmentation is very simple using the thresholding method. MR makes full use of a strong magnetic field (BO field) and radiofrequency pulses for generating images. If the image intensity is determined by the T1relaxation, it is T1-weighted imaging. If the image intensity is determined by the T1relaxation, it is T2-weighted imaging. Tl-weighted scans work well for differentiating fat from water with water appearing darker and fat brighter. This scans are obtained to observe the anatomical structure of the spine. T2-weighted scans are another basic type. Like the T1-weighted scan, fat is differentiated from water but in this case fat shows darker, and water lighter. In the case of spinal study, the cerebrospinal fluid will be lighter in T2-weighted images. This scans are acquired to examine the pathologic change of the spine. The segmentation of vertebral bodies in MR images is much challenging and complex due to the relatively variations in soft tissue contrast and artifacts like radio-frequency inhomogeneity. There is little work on the vertebral body extraction from the sagittal MR spinal images.Gamio et al. applied graph-cut to segment MR T1-weighted sagittal images of the spine. Graph-cut is an unsupervised method and does not require initialization. Using methods based on graph cut, an image is usually segmented into several distinct regions rather than the target and the background, and the pixels have high similarity within each region. A weighted graph is constructed, where nodes of the graph correspond to image pixels, and the weight of the edge reflects the similarity between two joined nodes. Solving the eigenvectors and eigenvalues of the affinity matrix performs the image segmentation. This method defined a partitioning criterion that maximizes the total similarity within groups and minimizes the total similarity between different groups. Gamio et al. segmented the vertebral body using windowed histograms of intensity as the most promising features. Due to the usage of the simple statistical characteristics of local histogram, Gamio algorithm is not a good choice for segmenting the images with same statistical characteristics of local histogram and low-contrast objects.We present develop a new approach to automatically segment vertebral bodies from spinal MR T1-weighted and T2-weighted sagittal images. Our methodology is novel in the following ways. To build a new affinity matrix for advanced image segmentation, we first use a cut-off window (5×5) around each pixel and stack the gray values inside the window into a vector, which local intensity is introduced to depict the image exactly and help to distinguish different tissues and suppress the effect of noise. Second, considering the contribution of the nearby pixels to the centered pixel, we adopt the Gaussian kernel function to incorporate local spatial information, thus allowing the suppression of noise and improving the accuracy of the segmentation. Third, an adaptive local scaling parameter is used to refine the segmentation rather than a fixed scaling parameter to avoid the manually tuned parameter. Finally, the built affinity is introduced into the segmentation process by using a graph-based method to achieve the complete target. Sagittal MR images of the spine were performed with a3.0T scanner at Nanfang Hospital, Guangzhou, China. The final data set comprised of100images (34healthy and66unhealthy). Extensively experiments show that the present method can segment the vertebral bodies smoothly and clearly, and it has stronger anti-noise property and higher segmentation precision than the conventional methods. The robust and accurate result of segmentation should serve image registration and the analysis of spinal deformities. It can also be used in organ location and image-guided vertebra operation, with presumed significant clinical impact. It is a general method for segmenting object that can develop to segment other tissues and organs.Second, we develop a Gaussian Chan-Vese(CV) model based on entropy and local neighborhood information for the tumor extraction from brain MR image.Tumor is one of the most common brain diseases, so its diagnosis and treatment is valuable for more than4000000persons per year in the world. The brain tumors have a great diversity in shape and appearance with intensities among individuals. It is still a challenge for automatically extracting the brain tumor, which helps contribute to3D visualization, computer aided pathologies diagnosis and surgical guidance for the clinicians in practice.Active contour model is one of the most popular image segmentation methods. The basic idea of the active contour model is to evolve a closed curve to extract the object. The energy function gets to the minimum and the closed curve gets to the boundary of the region of interest. The conventional CV model assumed the homogeneity of image intensities, and it cannot extract the tumor from the brain MR images with inhomogeneities of image intensities in the region of interest. We develop an adaptive Gaussian Chan-Vese(CV) model based on entropy and local neighborhood information for the tumor extraction from brain MR image. The main contributions are as follows:(1) In the cost function of this model, the interior and exterior energies are weighted by the entropy, which improves the robust of the evolving curve;(2) The local information of the curve is considered rather than global image statistics, which reduces the impact of the heterogeneous grays inside of regions and improves the segmentation results.(3) The Gaussian kernel is utilized to regularize the level set function, which not only keeps the level set function smooth and stable, but also removes the traditional Euclidean length term and re-initialization. The final data set comprised of100barin images, which were obtained using a3.0T scanner at Tianjin Medical Universiy General Hospital, Tianjin, China, for validating the present method. Extensively experiments show that the present method can segment the tumor from the brain MR images in terms of high accuracy.Third, we develop and validate an automated method of interventricular septum segmentation (AISS) from myocardial black-blood images for the T2*measurement in thalassemia patients.The thalassemia is the most common monogenic inheritance disorders in the world. The thalassemia major often requires the chronic blood transfusions, which can prevent bone lesions and has greatly prolonged survival and maintained a reasonable quality of life. However, long-term transfusion therapy can result in progressive accumulation of iron, which can cause damage in many organs, particularly the liver, heart, and endocrine organs, ultimately leading to increased morbidity and mortality. Chelation therapy can remove excessive tissue iron from the body and reduce the risk of organ failure in these patients. Thus accurate and robust measurement of myocardial iron concentration is clinically vital for tailoring appropriate iron-chelating treatment and assessing the prognosis in thalassemia patients.Myocardial iron deposits can be measured by using endomyocardial biopsy, but which is invasive with the risk of complications and uncertainty owing to the inhomogeneous myocardial iron distribution and quantification errors caused by very small biopsy samples. MR has been established as a non-invasive approach for detecting myocardial iron content (MIC) in various patients with iron overload. T2*is a form of T2with higher sensitivity to field changes caused by iron. Tissue iron can be measured indirectly from the effect of particulate iron within myocytes on relaxation times of hydrogen nuclei. The presence of iron causes a local disruption of the magnetic field, thereby shortening the relaxation times and these quantifiable effects can be calibrated for absolute iron concentration.T2*is quicker and easier to measure in the heart than T2, and is more sensitive to iron and less sensitive to cardiac motion. A mid-ventricular short-axis slice was chosen with a region of interest in the septum to give a uniform myocardial segment free of artefact for the measurement of T2*decay. In routine clinical practice, a representative T2*value of a region-of interest is generally reported as a quantitative marker of MIC. T2*measurements are preferably performed in the region of interventricular septum, as this region is least affected by severe susceptibility effects from epicardial fat, liver and lung tissues. The normal range for T2*(derived from a group of normal individuals with no history of transfusion or cardiac disease) is52±16ms, with a median normal value of40ms. The lower limit of the normal range is20ms.The myocardial MR images consits of conventional bright-blood imaging and black-blood imaging, and both of them depend on the blood and MR acquirements. The black-blood imaging uses the pulse sequence to detect the slowly moving hydrogen nuclei in the myocardium. Due to the flowing blood, the blood has low signal, nearby tissue show high signal and generate an enhanced image. The bright-blood technique could image rapid flow of blood in the cardiac chambers and vessel. A single-slice breath-hold bright-blood T2*method has demonstrated good inter-study and inter-centre reproducibilities in MIC assessment and is widely employed due to its speed, sensitivity, and wide availability. Compared with the bright-blood technique, the black-blood T2*technique has superior reproducibility because this technique avoids signal contamination from blood and shows a clearer definition of the myocardial boundary due to suppressed blood signal and reduced imaging artifacts. In this paper, we aimed to develop a fully automated method to extract IS from black-blood myocardial T2*images for assessing MIC. We make five contriutions as follows:(1) The circular Hough transformation and constrinted fuzzy C-means were introduced to locate the contours of the epicardium and endocardium of the left ventricular as the initialization for the improved CV model automatically, and this avioded to the initial curve placement manually and reduced the time of the curve moving toward the boudary of the object;(2) We utilize the present Gaussian Chan-Vese(CV) model based on entropy and local neighborhood information to extract the endocatdium and epicardium;(3) We use the thresholding method and label the connect component to exract the blood pool of the right ventricle. Due to the high contras between myocardium and blood pool, we utilize the segmented left ventricle to determine an threshold value automatically for extracting the blood pool of the right ventricle;(4) We utilize the anatomical information to detect the base and apex of the interventricular septum and extract the interventricular septum;(5) The T2*values were measured by fitting the average signal of the interventricular septum to the monoexponential model.(6) Reproducibility analyses were assessed using the mean and standard deviation (SD) of the differences as described by the Bland-Altman analysis. For accessing the intra-and inter-observer variability, the coefficient of variation (CoV) was defined as the SD of the differences between the two independent measurements divide by their means and presented as a percentage.A comprehensive framework for automated IS extraction from black-blood myocardial images is proposed and the myocardial T2*values measured using the proposed method avoided the subjectivity from the worldwide operators and improved the efficiency of data processing. A total of144thalassemia major patients (age range11-51years,73males) were scanned with a black-blood multi-echo gradient-echo sequence using a1.5T Siemens Sonata system at Royal Brompton Hospital, London, UK. The proposed AISS method is fully automated, and it takes approximately1.4s to fulfill the T2*analysis of one dataset, and the T2*measurements using the AISS method were in good agreement with those manually measured by experienced observers with a mean difference of1.71%and a CoV of 4.15%. In general, a CoV less than10%is considered high reproducibility, while a CoV less than5%indicates very high reproducibility. The small T2*overestimation (0.18ms) of the AISS method over the manual method will not produce significant differences in the clinical MIC grading. The T2*measurements using the AISS method were in good agreement with those manually measured by experienced observers, and the automated T2*measurements method could assist the clinician to manage the iron content in thalassemia major.更多还原