基于细胞自动机的MRI脑肿瘤分割算法研究

发布时间：2018-06-17 19:21

本文选题：脑肿瘤 + 图像分割　；参考：《南方医科大学》2014年硕士论文

【摘要】：脑肿瘤是指发生于颅腔内的神经系统肿瘤,包括原发性肿瘤和继发性肿瘤两类。原发性脑肿瘤是指发生于颅内脑组织、脑神经、脑膜、垂体以及胚胎残余组织等的肿瘤；继发性脑肿瘤是指颅腔外身体其他部位的恶性肿瘤转移或侵入颅内形成的转移瘤。在人群中,脑肿瘤发病率很高。据调查,原发性脑肿瘤的发病率为7.8/10万人～12.5/10万人。脑肿瘤可发生于任何年龄,以20～50岁年龄组多见。由于颅内肿瘤发生于有限的颅腔容积内,无论良性还是恶性肿瘤,占位效应本身就可以压迫脑组织并造成功能损害,甚至威胁生命。近年来,随着环境污染的加剧、生活压力的增大以及遗传因素的影响,脑肿瘤的发病率呈逐年上升的趋势。最新的肿瘤流行病学调查研究表明,脑肿瘤发病率约占全身肿瘤发病率的1.4%,死亡率超过2.4%。脑肿瘤是一类发病率、死亡率都很高的肿瘤疾病,已经成为危害人类生命健康的杀手。脑肿瘤的临床表现多种多样,早期症状有时不典型,甚至出现“例外”情况,而当脑肿瘤的基本特征均己具备时,病情往往已属晚期。脑肿瘤的临床表现一般为由颅内压增高引起的头痛、恶性呕吐、视乳头水肿和视力障碍、癫痫等,以及定位症状与体征如肌肉力减退、癫痫等。临床上,医生通过详细的病史询问和神经系统检查,可以了解起病方式、首发症状、症状经过以及有无高颅压和局灶性脑症状,根据这些可以推断是否存在脑肿瘤,一般凡有进行性颅内压增高并伴随有局灶脑部症状者,基本可以确定脑肿瘤的存在。进一步参照脑肿瘤的好发年龄、好发部位、症状的发生方式及进展情况以判断肿瘤的部位和性质。医生在诊断脑肿瘤时常辅助使用影像诊断来进行。计算机断层摄影(Computed Tomography, CT)和磁共振成像(Magnetic Resonance Imaging, MRI)己成为目前诊断脑肿瘤的最主要影像学手段。CT和MRI的应用大大提高的脑肿瘤的诊断能力,也使得脑肿瘤的临床治疗效果得到了改善。早期诊断,早期治疗是包括脑肿瘤在内所有疾病的医疗原则,治疗越早,效果越好。脑肿瘤的治疗原则上是以手术治疗为主,辅助以放疗、化疗、生物治疗等方式。手术治疗需要对脑肿瘤进行合理、有效的检测和监测,以尽可能的切除脑肿瘤。但由于恶性肿瘤的侵入性生长,使其在图像上表现为与周围组织边界模糊,这给临床医生的诊断带来了很大困难。同时,医生在诊断脑肿瘤时,存在一定差异。不同医生对同一个病人的脑肿瘤图像,或是同一个医生在不同时期对同一个病人的脑肿瘤图像的分割结果存在一定的差异。利用计算机技术对脑肿瘤进行有效分割,越来越受到人们的重视。在现代医学影像诊断技术中,MRI是一种重要的解剖性影像诊断技术。磁共振成像具有较高的组织对比度和组织分辨率,对组织的形态和病理改变有很高的敏感性,无电离辐射,属于无损伤性检查。同时,它可以进行多参数、多序列,任意方位的成像。目前,MRI已成为诊断脑肿瘤的主要手段。临床上,脑肿瘤的分割一般由经验丰富的医生在MRI图像上,利用计算机辅助软件,手动分割来完成的。手动分割有很强的主观性,可重复操作性差。同时,在磁共振成像的过程中,由于噪声、组织运动和局部体积效应等的影响,获得的图像对比度低,不同病灶与周围组织之间边界模糊,这又给手动分割带来了更大的困难。因此,利用脑肿瘤分割算法对脑肿瘤进行效应分割,成为临床上治疗脑肿瘤的迫切需要。图像分割就是把图像中的感兴趣区域分出来,使这些分开的区域之间相互不交叉,每个区域都满足特定区域的一致性。图像分割是图像分析和图像理解的基础,在医学、军事等领域都有着广泛的应用,吸引了国内外许多专家学者进行研究。随着研究的深入,研究人员提出了很多实用的分割算法,大致可分为基于区域的分割方法、基于边缘的分割方法、结合特定理论工具的分割方法等几类。基于区域的分割方法主要包括阈值法、区域生长和分裂合并法、特征聚类法以及基于马尔科夫随机场的方法。基于边缘的分割方法是通过检测边缘来进行分割的。为此,设计成了各种检测算子,如Sobel算子、LOG算子、Krish算子、Canny算子等。结合特定理论工具的分割方法主要包括基于数学形态学的分割方法、基于神经网络的分割方法、基于模糊理论的分割方法、基于分形理论的分割方法,基于形变模型的分割方法等。这些分割方法各有优缺点,基于马尔科夫随机场的分割方法虽然得到了广泛应用,但边缘定位不准确,运算量大,而且优化过程比较复杂。基于边缘检测的分割方法定位比较精确,但受噪声影响大,仅使用该方法很难对医学图像进行有效分割。基于神经网络的分割方法对随机噪声具有很强的鲁棒性,对人工干预要求比较小,但是图像的能量函数容易陷入局部最小。基于形变模型的分割方法对噪声和伪边界具有很强的鲁棒性,并且可以直接产生闭合参数曲线或曲面,但它对轮廓的初始位置比较敏感。随着理论研究的深入,细胞自动机越来越受到研究者的关注。细胞自动机(Cellular Automata,简称CA)是定义在有限状态、离散的细胞空间上,并按照一定的局部规则,在离散的时间维度上进行演化的动力学系统。细胞自动机主要包括细胞节点集合、邻域系统以及状态转移函数。细胞自动机的空间结构与数字图像的网格式存储结构具有一致性,可以把数字图像中的每个像素点一一对应到细胞自动机空间中的每个细胞单位上。依据处理目的的不同来制定不同的状态转移函数,从而得到不同的分割效果。 Grow Cut算法是一种基于细胞自动机的分割算法,主要应用于图像编辑和医学图像处理领域。它是一种交互式分割方法,可以利用一些先验知识进行简单的交互处理,简化分割过程。使用Grow Cut算法对MRI脑肿瘤图像进行分割时,分割结果不理想,主要因为该算法的状态转移函数不适用于复杂的MRI脑肿瘤图像,往往误判脑肿瘤边界附近的像素点,不能准确找到脑肿瘤边界。本文提出了一种新的基于细胞自动机的MRI脑肿瘤分割算法。该分割算法在细胞自动机的基础上,引入了活动轮廓模型来对分割算法进行优化。本文对种子点的选取进行了改良,只需要手动标记前景种子点就可以了,背景种子点通过一定的计算来得到,这就简化了人工交互过程。使用8邻域的Moore邻域来定义每个像素点的邻域空间,针对MRI脑肿瘤图像的特点,构造合适的状态转移函数,并使用一些先验知识对状态转移函数进行约束。当细胞自动机演化结束后,可以得到脑肿瘤图像的标号图,此时得到的分割结果还不够精确,会把脑肿瘤区域附近的一些非肿瘤像素点错分为肿瘤像素点。使用活动轮廓模型对标号图进行优化处理,最终可得到精确的脑肿瘤分割结果。使用我们提出的分割算法对临床脑肿瘤图像进行分割,实验数据是由在线数据库BRATS2012提供的对比增强T1加权MRI图像。把专家手动分割的结果作为分割真值,来对分割结果进行评价。使用Dice系数(DSC)、JM相似性系数(Jaccard's Measure)、Sensitivity (Sens.)和假阳性率(FPR)等技术评价指标对分割结果进行评价。评价结果显示,本文提出的分割算法具有很高的准确性,与真值结果很接近。由此验证了本文分割方法具有很高的可行性和实用性。
[Abstract]:Brain tumors refer to the tumors of the nervous system occurring in the cranial cavity, including two types of primary and secondary tumors. Primary brain tumors refer to tumors that occur in the brain, the brain, the meninges, the pituitary, and the remnant tissues of the embryo; secondary brain tumors refer to the metastasis or invasion of the other parts of the outer body of the skull. The incidence of brain tumors is high in the population. It is investigated that the incidence of primary brain tumors is 7.8/10 million to 12.5/10 million. Brain tumors can occur at any age and are more common in the age group of 20~50 years. It can oppress brain tissue and cause functional damage and even threaten life.
In recent years, with the intensification of environmental pollution, the increase of life pressure and the influence of genetic factors, the incidence of brain tumors is increasing year by year. The latest epidemiological investigation of tumor shows that the incidence of brain tumors is about 1.4% of the incidence of whole body tumors. The mortality rate over 2.4%. is a kind of incidence and the mortality is very high. Tumor diseases have become the killer of human life and health.
The clinical manifestations of brain tumors are varied. The early symptoms are sometimes untypical and even "exceptions". When the basic features of the brain tumors are all possessed, the condition is often late. The clinical manifestations of brain tumors are usually caused by increased intracranial pressure, emetic vomiting, papillatous papillae, visual impairment, epilepsy and so on. Symptoms and signs such as muscle degeneration, epilepsy, and so on.
Clinically, doctors can understand the mode of onset, first symptoms, symptoms, and whether there are high intracranial pressure and focal brain symptoms by detailed medical history inquiry and nervous system examination. According to these, it is possible to deduce whether there is brain tumor. In general, patients with progressive intracranial pressure and accompanied with focal brain symptoms can basically determine the brain swelling. The presence of the tumor. Further reference to the good onset age of the brain tumor, the location of the good hair, the way of the symptoms and the progress to determine the location and nature of the tumor. The doctor is often assisted by imaging diagnosis in the diagnosis of brain tumors. Computed tomography (Computed Tomography, CT) and magnetic resonance imaging (Magnetic Resonance Imaging, MRI) It has become the most important imaging method for the diagnosis of brain tumors,.CT and MRI, which have greatly improved the diagnostic ability of brain tumors, and have improved the clinical therapeutic effect of brain tumors.
Early diagnosis, early treatment is the medical principle of all diseases including brain tumors, the earlier the treatment, the better the effect. The treatment of brain tumors is mainly based on surgical treatment, assisted by radiotherapy, chemotherapy, biological treatment and so on. Surgical treatment requires rational, effective detection and monitoring of brain tumors, so as to remove brain tumors as much as possible. But because of the invasive growth of the malignant tumor, it shows the blurring of the boundary of the surrounding tissue on the image, which is very difficult for the diagnosis of the clinician. At the same time, there are certain differences between the doctors and the brain tumor in the diagnosis of the same patient. There are some differences in the segmentation results of human brain tumor images. More and more attention has been paid to the effective segmentation of brain tumors using computer technology.
In modern medical imaging diagnosis technology, MRI is an important anatomical imaging diagnosis technology. Magnetic resonance imaging has high tissue contrast and tissue resolution, has high sensitivity to the morphological and pathological changes of tissue, without ionizing radiation, and is a noninvasive examination. At the same time, it can carry on multi parameter, multi sequence and arbitrary side. Imaging. At present, MRI has become a major means of diagnosis of brain tumors. Clinical segmentation of brain tumors is usually done by experienced doctors on MRI images, using computer aided software and manual segmentation. Manual segmentation has strong subjectivity and poor repeatability. In the process of magnetic resonance imaging, noise, With the influence of tissue movement and local volume effect, the image contrast is low and the boundary between the different focus and the surrounding tissue is blurred. This brings more difficulty to the manual segmentation. Therefore, it is an urgent need to use the brain tumor segmentation algorithm to segment the brain tumor and to treat the brain tumor in clinical.
Image segmentation is to divide the region of interest in the image, which makes the separate areas not cross each other. Each region meets the consistency of the specific region. Image segmentation is the basis of image analysis and image understanding. It has a wide application in medical and military fields, which attracts many experts and scholars at home and abroad to study. With the deepening of the research, researchers have proposed many practical segmentation algorithms, which can be roughly divided into region based segmentation methods, edge based segmentation methods and the segmentation methods of specific theoretical tools. The region based segmentation methods mainly include threshold method, regional growth and split merge method, feature clustering method and base. In Markov random field, the edge based segmentation method is segmented by detecting edges. Therefore, various detection operators, such as Sobel operator, LOG operator, Krish operator and Canny operator, are designed. The segmentation method combined with specific theoretical tools mainly includes the segmentation method based on mathematical morphology, based on neural network. The segmentation method, the segmentation method based on the fuzzy theory, the segmentation method based on the fractal theory, the segmentation method based on the deformation model, and so on. These segmentation methods have the advantages and disadvantages. Although the segmentation method based on Markov random field has been widely used, the edge location is inaccurate, the computation is large and the optimization process is complex. The segmentation method of edge detection is more accurate, but it is greatly affected by noise. It is difficult to effectively segment medical images by using this method only. The segmentation method based on neural network has strong robustness to random noise and small requirement for artificial intervention, but the energy function of the image is easy to fall into the local minimum. The segmentation method of the model has strong robustness to the noise and the pseudo boundary, and can directly produce the closed parameter curve or surface, but it is more sensitive to the initial position of the contour.
Cellular automata (Cellular Automata) is a dynamic system defined in a finite state, discrete cellular space, and evolves in discrete time dimensions. Cellular automata mainly include cell nodes, which are defined in a finite state, discrete cellular space, and according to certain local rules. Set, neighborhood system and state transfer function. The spatial structure of cellular automata is consistent with the network format storage structure of digital images. Each pixel in the digital image can be corresponded to each cell unit in the cellular automaton space. Different state transfer functions are formulated according to the difference of the processing order. Different segmentation results are obtained.
Grow Cut algorithm is a kind of segmentation algorithm based on cellular automata, which is mainly used in image editing and medical image processing. It is an interactive segmentation method. It can use some prior knowledge to do simple interactive processing and simplify the segmentation process. Using Grow Cut algorithm to segment the image of MRI brain tumor, the segmentation results are not. Ideal, mainly because the state transfer function of the algorithm does not apply to the complex MRI brain tumor image, often misjudges the pixels near the brain tumor boundary, and can not accurately find the brain tumor boundary.
In this paper, a new MRI brain tumor segmentation algorithm based on cellular automata is proposed. Based on the cellular automata, the active contour model is introduced to optimize the segmentation algorithm. In this paper, the selection of seed points is improved, only the foreground seed points need to be labelled manually, and the background seed points have been passed through certain points. This simplifies the process of artificial interaction. We use the Moore neighborhood of the 8 neighborhood to define the neighborhood of each pixel, construct a suitable state transfer function for the features of the MRI brain tumor image, and use some prior knowledge to deal with the state transfer function. When the cellular automaton is over, the brain can be obtained. The segmentation results of the tumor image are not accurate enough, and some non tumor pixels near the brain tumor area will be misclassified as the tumor pixels. The active contour model is used to optimize the label map, and the accurate segmentation results of the brain tumor can be obtained.
We use the segmentation algorithm we proposed to segment the clinical brain tumor images. The experimental data is the contrast enhanced T1 weighted MRI image provided by the online database BRATS2012. The result of the segmentation is evaluated by the result of the expert manual segmentation as the segmentation true value. The Dice coefficient (DSC), the JM similarity coefficient (Jaccard's Measure) and Sensitivi are used. Ty (Sens.) and false positive rate (FPR) evaluation indexes are used to evaluate the segmentation results. The results show that the segmentation algorithm proposed in this paper is very accurate and close to the true value results. Thus, it is proved that the segmentation method is highly feasible and practical.
【学位授予单位】：南方医科大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：R739.41;R445.2

【参考文献】