结合MRI多模态信息与SVM参数优化的脑肿瘤分割研究

发布时间：2018-04-28 04:17

本文选题：多模态 + 混合核函数　；参考：《南方医科大学》2014年硕士论文

【摘要】：脑肿瘤是指生长在颅腔内的癌性物质,包括由脑、脑膜、血管、神经及脑附件等脑实质发生病变引起的原发性肿瘤,和由身体其他部位转移侵入颅内的继发性肿瘤,在人群中发病率很高,可发生于任何年龄,以20～50岁最为多见。而且不论其性质是良性还是恶性,一旦在颅内占据一定空间,势必压迫脑组织,造成颅内压升高、中枢神经损害,危及患者的生命。近年来随着加剧的环境污染、过重的生活压力等因素的影响,脑肿瘤的发病率呈上升趋势,最新的肿瘤流行病学调查研究结果表明,脑肿瘤发病率占全身肿瘤发病率的1.4%,而死亡比例超过2.4%,仍然是一类发病率、死亡率较高的肿瘤性疾病,成为威胁人类生命的重要疾病之一。胶质瘤是最常见的原发性恶性脑肿瘤,居中枢神经系统肿瘤首位。胶质瘤多呈浸润地弥漫性生长,形状多变且与周围组织边界模糊,需要在全面的神经系统检查的基础上,采用适当的成像方式辅助检查,使医生能够准确地诊断并有效地分割肿瘤。脑肿瘤起病缓慢,逐渐进展,病程长短不一,一般表现为由颅内高压引起的头痛、呕吐、视神经乳头水肿等症状。不同的肿瘤发病部位,还可以表现为局灶性症状和体征如偏瘫、失语、精神及意识障碍等麻痹性症状以及癫痫、肌肉抽搐等刺激性症状。临床医生依靠病史和可靠的查体,在神经解剖、生理和各种疾病发展规律的诊断学基础上,进行综合、客观地分析,并进一步选择辅助检查工具,全面分析研究肿瘤的部位、大小、性质、血供及对周围组织的累及程度,对肿瘤做出较为精确的定位与定性鉴别诊断。脑肿瘤的治疗原则上是以手术治疗为主,辅助以放、化疗、生物治疗等方式。手术治疗的前提就是需要合理有效的手段检测和监测脑肿瘤,尽可能地切除肿瘤,由于恶性肿瘤浸润性生长,与周围组织边界模糊,不同医生对同一病人的肿瘤图像,或者同一个医生不同时期对同一病人的肿瘤图像分割结果存在差异,因而利用计算机图像处理技术有效地识别分割肿瘤,是临床应用发展的必然趋势。临床上常用于脑部辅助影像检查的技术包括计算机断层扫描(Computed tomography, CT)和磁共振成像(Magnetic Resonance Imaging, MRI)等。这些影像技术的发展及广泛应用,大大增加了脑肿瘤的检出率,给医生和患者带来很大的帮助。影像学设备为计算机处理提供多类图像,磁共振成像(magnetic resonance imaging, MRI)是一种重要的解剖性影像诊断技术,MRI图像对软组织有极好的分辨力。它作为一种无损伤、无辐射、多参数的成像方式,对组织的形态及病理改变的显示有较高的敏感性,目前已经成为诊断脑部肿瘤的重要工具。不同模态MRI图像侧重表现图像不同的差异信息,比如,FLAIR模态中肿瘤与正常组织灰度差异明显,T1C边界纹理特征区别明显。单一模态MRI图像难以充分提供病变组织的可辨识信息,同时,脑胶质瘤形状多变且与周围水肿区域边界模糊,准确分割图像中的肿瘤十分困难。临床上一般由有经验的医生结合多模态(Multi-modality)MRI图像,利用计算机辅助软件,手动一层一层地勾画肿瘤区域,主观性很强,可重复操作性差。因而利用机器有效地分割肿瘤是临床应用发展的必然趋势。目前常见的肿瘤分割方法主要有基于图像灰度信息的模糊聚类(FCM)方法、水平集方法、神经网络方法、AdaBoost迭代方法、以及支持向量机(Support Vector Machine, SVM)方法等。FCM方法实现简单,运算速度快,但是由于医学图像信息复杂,边缘不清晰,因此,种子点的选取对聚类结果影响很大,并且FCM方法难以利用图像的空域信息。对于水平集方法,其最大优点在于曲线的拓扑变化处理自然,稳定性强,但是曲线初始化要求高,参数选择敏感,分割结果容易陷入局部极值。神经网络方法学习能力强,但神经网络方法对训练样本要求较高,容易出现过拟合及局部最优的问题,使得其泛化性能变差,尤其在小样本的情况下。AdaBoost算法分割精度较高,无过拟合现象,但是当AdaBoost算法用于分割多模态MRI图像时,训练所需样本量大,训练时间长。以上分割方法各有优势,但离临床应用还存在一定差距。基于统计学习理论的SVM方法表现出很多优势,SVM在样本相对较少、特征维数较高的情况下仍能取得很好的推广能力,同时引入核函数的SVM可以有效地处理非线性可分数据。有文献采用单一高斯核函数SVM方法,对多模态MRI图像取得了较好的应用效果。但是高斯核函数善于利用样本的局部信息,仅引入单一高斯核函数可以对组织区别明显的图像,获得良好的分割结果,对于边界模糊、形状多变的胶质瘤,单一高斯核函数SVM的性能有一定的局限性,而具有局部性质和全局性质的混合核函数可以克服此类问题。人脸识别及掌纹识别中已经广泛地证实,参数最优组合的混合核函数性能优于单一核函数。人脸、掌纹结构简单、相对固定,而含肿瘤组织的脑部图像,尤其是胶质瘤中的低级胶质瘤,呈弥漫的浸润性生长,信号强度介于正常组织之间,肿瘤形状、位置、大小多变,与周围组织边界模糊,组织纹理结构复杂,如何充分使用图像多模态信息,并寻找SVM模型参数的最优组合,这是目前SVM方法应用于肿瘤图像分割的难点。本文改进现有的多模态MRI脑肿瘤分割方法,充分利用MRI图像的多模态信息,同时结合支持向量机(Support Vector Machine, SVM)统计学习方法的优势,提出一种基于SVM模型参数优化的多模态MRI图像肿瘤分割方法。该方法首先分析MRI所成的多模态图像,不同模态的图像突出的肿瘤组织与正常组织的差异信息不同,有效区分肿瘤组织与周围组织的支持向量位置有差异。其次优化核函数支持向量机分类器,支持向量机分类器引入核函数,巧妙地解决非线性可分问题。核函数包括局部性核函数和全局性核函数,不同类型的核函数侧重的信息不同,性能有差异,结合局部性高斯核函数和全局性Sigmoid核函数的性能优势。然后,对单一模态训练最优混合核函数SVM子分类器,仅需要小样本的训练集,且性能优于单一高斯核函数。由于不同模态图像选择的支持向量各有侧重,分割结果存在差异。通过迭代修改分割错误数据点的权值,优化选择SVM模型子分类器权重系数,得到多模态加权组合的SVM分类器模型,增强分割性能并应用于多模态MRI图像分割。实验表明本文方法泛化性能良好,可行性和实用性强,可以实现对脑肿瘤的精确分割。因此本文引入并提出的关键技术包括：(1)图像去噪算法；(2)核函数混合方法；(3)SVM分类器组合方法。 (1)MRI图像中的噪声会降低图像质量,影响图像的视觉观察效果,使用图像过程中,机器从图像中获取的信息减少,甚至是得到错误信息,使得图像处理的算法结果准确度受到影响,因而需要首先对图像进行滤波处理。针对MRI图像中的加性噪声,同质区像素只差仅与噪声有关,引入一种增维型双边滤波的快速算法,在保证滤波性能的前提下,使双边滤波的快速实现,既可以有效防止去噪过程破坏图像的重要信息,又加快了整体方法的实现。 (2)优化混合核函数的组合系数。通过自适应调节新映射空间中各个样本点的距离,削弱分类器惩罚因子对分类结果的影响,使得参数寻优过程中可以固定惩罚因子,而不影响分割精度；同时,权重系数的优化能改变序列最小优化(Sequential Minimal Optimization, SMO)算法中的修正因子,从而影响支持向量的选取,以得到更优的分类间隔,最终大大提高脑肿瘤的分割精度。 (3)多模态分类函数加权组合,充分利用不同模态突出差异信息的不同。支持向量机方法是一个很有优势的学习方法,但是医学图像信息复杂,有限样本训练的最优分类器并不能满足高精度的要求,因而利用集成学习理论,通过构造多个差异性大、性能较好而又独立的子分类器,将其组合来提高最终分类器的泛化性能。利用MRI图像的多模态信息,每种模态都对应一个新的样本集,分别训练子分类器。然后将子分类器组合,优化分类结果。本文以在线图像库MICCAI2012中34例脑胶质瘤病人图像数据为实验样本,采用本文算法对病人脑部图像中的肿瘤进行分割。通过临床医生判断和定量分析,本文对脑肿瘤的分割准确率达到92.50%,与真值结果非常接近。由此验证了本文方法的可行性和实用性。
[Abstract]:Brain tumor is a carcinomatous substance that grows in the cranial cavity, including the primary tumor caused by brain parenchyma, such as brain, meninges, blood vessels, nerves and brain appendages, and secondary tumors that are transferred from other parts of the body, with high incidence in the population, at any age, most common at the age of 20~50, and no matter what it is. The nature is benign or malignant. Once occupying a certain space in the brain, it is bound to constriction brain tissue, causing intracranial pressure to increase, central nervous system damage, and endanger the lives of patients.
In recent years, the incidence of brain tumors is on the rise with the influence of aggravated environmental pollution, heavy life pressure and other factors. The latest oncology epidemiological investigation results show that the incidence of brain tumors is 1.4% of the incidence of whole body tumor, and the proportion of death is over 2.4%, which is still a kind of incidence and high mortality of tumor. Disease is one of the most important diseases that threaten human life. Glioma is the most common primary malignant brain tumor. It is the primary tumor in the middle and central nervous system. Gliomas are mostly infiltrating and diffuse growth, and the shape is changeable with the boundary of the surrounding tissue. It is necessary to use appropriate imaging methods on the basis of comprehensive neural examination. Examination enables doctors to accurately diagnose and effectively segment tumors.
The brain tumor begins slowly, progresses gradually, and the course of the disease is different. It is usually manifested by the symptoms of headache caused by intracranial hypertension, vomiting, and papillary edema of the optic nerve. Different parts of the tumor can also be shown as focal symptoms and signs, such as hemiplegia, aphasia, mental and cognitive disorders, as well as epilepsy, muscle twitching and so on. Sex symptoms. Clinicians rely on medical history and reliable physical examination, based on the diagnostics of neuroanatomy, physiology and the development of various diseases, and make a comprehensive, objective analysis, and further choose an auxiliary examination tool to comprehensively analyze the site, size, nature, blood supply and involvement of the surrounding tissue, and to make a more specific tumor to the tumor. Accurate localization and qualitative differential diagnosis.
The treatment of brain tumors is mainly based on surgical treatment, assisted by radiotherapy, chemotherapy, and biological treatment. The premise of surgical treatment is to detect and monitor brain tumors with reasonable and effective methods and to remove the tumor as much as possible. Because of the invasive growth of malignant tumor, the boundary of the surrounding tissue is blurred, and the tumor map of the same patient is different from the doctor. There is a difference between the image segmentation results of the same patient in different periods of the same doctor, so it is an inevitable trend for the development of clinical application to identify the tumor effectively by using computer image processing technology.
The techniques commonly used in brain assisted imaging include computed tomography (Computed tomography, CT) and magnetic resonance imaging (Magnetic Resonance Imaging, MRI). The development and extensive application of these imaging techniques greatly increase the detection rate of brain tumors and bring great help to doctors and patients.
Imaging equipment provides multi class images for computer processing. Magnetic resonance imaging (MRI) is an important anatomical imaging diagnosis technique. MRI images have excellent resolution to soft tissues. It is a imaging modality without damage, radiation and multiple parameters. It shows the morphological and pathological changes of tissue. High sensitivity has become an important tool for the diagnosis of brain tumors. Different modal MRI images focus on different information of different images. For example, the difference of the gray level between the tumor and the normal tissue in the FLAIR mode is obvious, and the difference of the texture features of the T1C boundary is obvious. The single modal MRI image is difficult to provide the identifiable information of the pathological tissue. At the time, the glioma has a changeable shape and blurred boundary with the surrounding area of edema. It is very difficult to accurately segment the tumor in the image. In general, it is commonly used by experienced doctors to combine multimodal (Multi-modality) MRI images and use computer aided software to manually delineate the swelling area with one layer and one layer. Therefore, the subjectivity is very strong, and the repeatability is poor. It is an inexorable trend to use machine to segment tumor effectively.
At present, the main methods of tumor segmentation are fuzzy clustering (FCM) based on image gray information. The level set method, neural network method, AdaBoost iterative method, and support vector machine (Support Vector Machine, SVM) method are simple and fast, but the edge is not clear because of the complicated medical image information. Therefore, the selection of seed points has great influence on the clustering results, and the FCM method is difficult to make use of the spatial information of the image. For the level set method, its biggest advantage is that the topological change of the curve is natural and strong, but the requirement of the curve initialization is high, the parameter is sensitive to the selection, and the segmentation result is easy to fall into the local extremum. Neural network method is easy to get. The learning ability is strong, but the neural network method requires higher training samples, easy to appear over fitting and local optimal problem, which makes its generalization performance worse. Especially in small sample cases,.AdaBoost algorithm has higher segmentation precision and no overfitting phenomenon, but when AdaBoost algorithm is used to segment multimodal MRI images, the training sample is trained. The above segmentation methods have their own advantages, but there is still a certain gap between them.
The SVM method based on the statistical learning theory shows many advantages. SVM can still obtain good generalization ability when the sample is relatively small and the feature dimension is high. At the same time, the SVM of the kernel function can effectively deal with the nonlinear separable data. A single Gauss kernel function SVM method is used in the literature to obtain a better multimodal MRI image. But the Gauss kernel function is good at using the local information of the sample, only introducing a single Gauss kernel function to distinguish the obvious image of the organization and obtain good segmentation results. It has some limitations for the performance of the single Gauss kernel function SVM, which has a certain local and global properties. The hybrid kernel function can overcome such problems. Face recognition and palmprint recognition have widely confirmed that the performance of the optimal combination of parameters is superior to that of a single kernel. Face, the palmprint structure is simple and relatively fixed, and the brain images containing tumor tissue, especially the low glioma in glioma, are diffuse infiltrating. Long, the signal intensity is between normal tissues, the shape, position and size of tumor are changeable, and the boundary of the surrounding tissue is fuzzy, the texture structure is complex. How to use the multi-modal information of the image fully and find the optimal combination of SVM model parameters is the difficulty of the SVM method applied to the segmentation of the tumor image.
In this paper, we improve the existing multimodal MRI brain tumor segmentation method, make full use of the multi-modal information of MRI image, and combine the advantages of Support Vector Machine (SVM) statistical learning method, and propose a multimode MRI image segmentation method based on the parameter optimization of SVM model. This method first analyzes the multimode of MRI. The difference in the difference between the tumor tissues and the normal tissues of the images of different modes is different, and the difference between the support vector positions of the tumor tissue and the surrounding tissue is distinguished. Secondly, the kernel function support vector machine classifier is optimized, and the support vector machine classifier is introduced into the kernel function, and the nonlinear separable problem is solved skillfully.
The kernel function includes the local kernel function and the global kernel function. The different types of kernel functions are different in information, the performance is different, and the performance advantages of the local Gauss kernel function and the global Sigmoid kernel are combined. Then, the optimal mixed kernel function SVM Subclassifier for single modal training only needs a small sample training set, and the performance is excellent. In the single Gauss kernel function. Because the support vectors selected from different modal images have each particular emphasis, the segmentation results are different. The weight coefficients of the SVM model sub classifier are optimized by iteratively modifying the weights of the error data points, and the SVM classifier model of multimodal weighted combination is obtained, and the segmentation performance is enhanced and applied to the multimodal MRI graph. Experiments show that the proposed method has good generalization performance, feasibility and practicability, and can accurately segment brain tumors.
Therefore, the key technologies introduced and put forward in this paper include: (1) image denoising algorithm; (2) kernel function hybrid method; (3) SVM classifier combination method.
(1) the noise in the MRI image can reduce the image quality and affect the visual observation effect. In the process of using the image, the information obtained by the machine is reduced, and even the error information is obtained. The accuracy of the image processing algorithm is affected by the image processing. Therefore, the image is filtered first. The addition of the image to the MRI image is added. Noise is only related to the noise in the homogeneity region, and a fast algorithm is introduced. The fast realization of the bilateral filtering can not only effectively prevent the de-noising process to destroy the important information of the image, but also accelerate the realization of the whole method.
(2) optimize the combination coefficient of the mixed kernel function. By adjusting the distance of each sample point in the new mapping space adaptively, the influence of the classifier penalty factor on the classification results is weakened, and the penalty factor can be fixed in the process of parameter optimization without affecting the segmentation precision. At the same time, the optimization of the weight coefficient can change the minimum sequence optimization (Sequenti The correction factor in the Al Minimal Optimization (SMO) algorithm affects the selection of support vectors to get better classification intervals, and ultimately greatly improves the segmentation accuracy of brain tumors.
(3) weighted combination of multi-modal classification functions, making full use of different modes to highlight different information. Support vector machine method is a very advantageous learning method, but medical image information is complex and the optimal classifier trained by limited samples can not satisfy the requirement of high precision. Therefore, the integrated learning theory is used to construct a number of different methods. A better and independent Subclassifier is used to improve the generalization performance of the final classifier. Using the multi-modal information of the MRI image, each mode corresponds to a new sample set, and the sub classifier is trained respectively. Then the Subclassifier is combined to optimize the classification results.
In this paper, the image data of 34 patients with glioma in the online image library MICCAI2012 are taken as experimental samples. The algorithm is used to segment the tumor in the brain image of the patient. The segmentation accuracy of the brain tumor is up to 92.50% through the clinician and the quantitative analysis, which is very close to the true value results. This method is verified by this method. Feasibility and practicability.

【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

【参考文献】