结构磁共振影像特征信息提取方法研究
发布时间：2018-11-23 10:12
【摘要】：阿尔兹海默症(Alzheimer's Disease,AD)是一种最常见的痴呆症,发病者大多为老年人。AD是一种神经变性紊乱疾病,关于其发病机制目前仍未有非常明确的说法。AD病情的发展是一个渐进、平缓的过程,患者的记忆、注意力、语言等认知能力将逐渐减退或者受损,最终患者会出现昏迷状况,通常死于感染等并发症。随着现在国家人口老龄化日益严重,AD的发病率会越来越高,由其引发的开销和损失也是不可估量的。对于AD患者以及他们的家人来说,除了巨大的经济负担以外,更难以承受的则是精神和情感上的巨大压力与折磨。 临床和神经病理学研究已经极大地推进了人们对AD病理生理和疾病发展的认识,但是目前还没有任何诊断方法可以对活体个体进行AD的确诊,只有在个体死亡后对其进行尸检才能确诊。所以,寻找一种无创伤性的AD临床诊断方法具有重要的理论和实际意义。 随着医学影像技术的发展,基于医学影像的临床评估方法已经逐渐成为AD临床诊断中一个重要组成部分,其中又以结构磁共振成像(structural magnetic resonance imaging, sMRI)的使用最为广泛。sMRI图像能够客观记录下从疾病潜伏期到发作期整个过程中AD患者脑结构生物标记的变化,这些数据能够从根本上改变人们对这种疾病的认识,并且能够影响和引导疾病的后续诊断和治疗。 传统的sMRI图像处理方法工作量大、耗时长、过程复杂,对使用者的先验知识要求较高,现在需要一种全自动的sMRI数据分析方法来识别和检测受试者脑结构中潜在的AD生物标记。盲源分离、子空间学习和机器学习等信号处理方法的进展为全自动的sMRI数据分析方法提供了可能的技术手段,利用这些信号处理方法可以从受试者的sMRI数据中提取出关键信息并对其是否患有AD进行诊断。 独立成分分析(independent component analysis,ICA)作为一种盲源分离方法,近几年有不少学者将其用于AD的sMRI数据分析。在这些研究中,研究者们假设各受试者的sMRI影像之间是相互独立的,然后用ICA对向量化后的sMRI影像做特征提取,提取出的特征用于后续对受试者的分类诊断。但这种模型在用ICA做特征提取时需要用到一组受试者的sMRI影像,当有新增单个受试者时无法立刻对其sMRI数据进行特征提取进,进而导致无法诊断。这一缺点使得这种诊断方法不符合临床诊断的需求,即在临床上希望每有一个新增受试者都可以即刻对其进行诊断。针对这一问题本文提出一种新的ICA特征提取模型,基于该特征提取模型的诊断方法可以对新增单个受试者进行诊断。该模型假设每个sMRI影像的各体素之间是相互独立的,然后先利用一组sMRI训练数据训练出解混矩阵,这样当有新增sMRI数据时可以利用训练好的解混矩阵即刻对其进行特征提取,进而进行后续的诊断,满足临床诊断的需求。仿真实验证明,本文提出的新的基于ICA的诊断方法可以达到与原ICA诊断方法相当的诊断准确率,且更符合实际诊断时的需求。 包括ICA在内的很多线性特征提取方法在AD的sMRI数据分析上表现出了良好的性能,但是这些线性特征提取方法都需要将原始的三维sMRI影像向量化之后才能对数据进行分析。这样处理带来的后果是原始三维图像数据中的空间信息会遭到破坏,造成了大量有效信息的丢失；同时,sMRI影像的数据量非常大,因此向量化之后得到的向量维数非常高,而受试者的数量是有限的,这样就可能会导致小样本问题(under sample problem)。针对这些问题,本文主要提出了一种基于非相关多线性主成分分析(uncorrelated multilinear principal component analysis, UMPCA)和拉普拉斯分值(laplacian score, LS)的新分类诊断方法。UMPCA是一种多线性子空间学习方法,用其对sMRI数据进行特征提取可以用直接张量模型来表示三维影像数据并对其进行处理,而不需要将原始的三维sMRI数据向量化,保留了原始数据的空间结构信息,避免了前面所提到的向量化带来的问题；另外,特征提取后的信息仍有可能有一定的冗余度,在提取sMRI数据特征信息后加入了LS特征选择的过程,可以进一步减少冗余信息,降低计算复杂度,选取出区别度高的特征,有效地提高了之后诊断过程的准确率。仿真实验表明,同现有的诊断方法相比,本文提出的UMPCA-LS分类诊断方法准确率更高。
[Abstract]:Alzheimer's Disease (AD) is one of the most common forms of dementia. AD is a neurodegenerative disorder, and its pathogenesis is still not well-defined. The development of AD is a gradual and gradual process, and the cognitive ability of the patient's memory, attention, language and so on will be gradually reduced or damaged, and the final patient will be in a coma, usually with complications such as infection. As the aging population of the country is becoming more and more serious, the incidence of AD is getting higher and higher, and the expenses and losses caused by it are also inestimable. For AD patients and their families, in addition to the huge economic burden, the more difficult to bear is the great pressure and torment of the spirit and the emotion. The clinical and neuropathological study has greatly advanced people's understanding of the pathophysiology of AD and the development of the disease, but there is no diagnostic method for the diagnosis of AD in living individuals, only after the individual has died Therefore, it is important to find a non-traumatic AD clinical diagnosis method. With the development of the medical image technology, the clinical evaluation method based on the medical image has gradually become an important part in the clinical diagnosis of AD, in which the structure magnetic resonance imaging (sMRI) is used. The most widely used. sMRI images can objectively record the changes in the biological markers of the brain structure of AD patients from the latent period of the disease to the whole process of the attack period, which can fundamentally change people's understanding of the disease, and can influence and guide the follow-up of the disease. the traditional sMRI image processing method has the advantages of large workload, long time consumption, complex process, high requirements on the prior knowledge of the user, and a fully-automatic sMRI data analysis method is needed to identify and detect the subpotential in the brain structure of the subject. the progress of signal processing methods such as blind source separation, subspace learning, and machine learning provides a possible technical means to extract key information from the subject's sMRI data and determine whether it An independent component analysis (ICA) is used as a blind source separation method. In recent years, a number of scholars have been used to The sMRI data analysis of AD. In these studies, the researchers assumed that the sMRI images of each subject were independent of each other and then extracted the quantized sMRI image with ICA and the extracted features were used for Follow-up to the subject's classification diagnosis. However, this model requires a set of subject's sMRI images when using ICA for feature extraction, and it is not possible to immediately characterize the sMRI data when there is a new single subject The disadvantage is that this diagnostic method does not meet the need for clinical diagnosis, i.e., clinically, there is a need for every additional subject This paper presents a new ICA feature extraction model based on the feature extraction model. Add a single subject to make a diagnosis. The model assumes that each of the voxels of each sMRI image is independent of each other, and then uses a set of sMRI training data to train the demixing matrix so that when there is new sMRI data, it can be characterized by a well-trained demixing matrix Take, and follow up The simulation results show that the new ICA-based diagnostic method can achieve the same diagnostic accuracy as the original ICA diagnostic method, and More realistic diagnosis needs. Many linear feature extraction methods, including ICA, show good performance in the sMRI data analysis of AD, but these linear feature extraction methods require the original three-dimensional sMRI image After quantization, the data can be analyzed. The consequence of this process is that the spatial information in the original three-dimensional image data is destroyed, resulting in a loss of a large amount of effective information; at the same time, the amount of data of the sMRI image is very large, so that the number of vector dimensions obtained after quantization is not Often high, and the number of subjects is limited, which may result in a small sample problem (under In view of these problems, a non-correlated multi-linear principal component analysis (UMPCA) and a Laplacian value (laplacian sco) are proposed in this paper. the new classification and diagnosis method of re, ls) is a multi-linear subspace learning method, which can be used for extracting the sMRI data, and the three-dimensional image data can be represented by a direct tensor model and processed without the need to vectorize the original three-dimensional sMRI data, the space structure information of the original data is reserved, the problem caused by the backward quantization mentioned above is avoided; in addition, the extracted information can still have certain redundancy, and the process of selecting the LS feature selection after extracting the sMRI data characteristic information can further the redundant information is reduced, the computational complexity is reduced, the characteristics of high difference are selected, The accuracy of the post-diagnosis process is improved. The simulation results show that the UMPCA presented in this paper is compared with the existing diagnostic method.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41;R445.2
本文编号：2351216
[Abstract]:Alzheimer's Disease (AD) is one of the most common forms of dementia. AD is a neurodegenerative disorder, and its pathogenesis is still not well-defined. The development of AD is a gradual and gradual process, and the cognitive ability of the patient's memory, attention, language and so on will be gradually reduced or damaged, and the final patient will be in a coma, usually with complications such as infection. As the aging population of the country is becoming more and more serious, the incidence of AD is getting higher and higher, and the expenses and losses caused by it are also inestimable. For AD patients and their families, in addition to the huge economic burden, the more difficult to bear is the great pressure and torment of the spirit and the emotion. The clinical and neuropathological study has greatly advanced people's understanding of the pathophysiology of AD and the development of the disease, but there is no diagnostic method for the diagnosis of AD in living individuals, only after the individual has died Therefore, it is important to find a non-traumatic AD clinical diagnosis method. With the development of the medical image technology, the clinical evaluation method based on the medical image has gradually become an important part in the clinical diagnosis of AD, in which the structure magnetic resonance imaging (sMRI) is used. The most widely used. sMRI images can objectively record the changes in the biological markers of the brain structure of AD patients from the latent period of the disease to the whole process of the attack period, which can fundamentally change people's understanding of the disease, and can influence and guide the follow-up of the disease. the traditional sMRI image processing method has the advantages of large workload, long time consumption, complex process, high requirements on the prior knowledge of the user, and a fully-automatic sMRI data analysis method is needed to identify and detect the subpotential in the brain structure of the subject. the progress of signal processing methods such as blind source separation, subspace learning, and machine learning provides a possible technical means to extract key information from the subject's sMRI data and determine whether it An independent component analysis (ICA) is used as a blind source separation method. In recent years, a number of scholars have been used to The sMRI data analysis of AD. In these studies, the researchers assumed that the sMRI images of each subject were independent of each other and then extracted the quantized sMRI image with ICA and the extracted features were used for Follow-up to the subject's classification diagnosis. However, this model requires a set of subject's sMRI images when using ICA for feature extraction, and it is not possible to immediately characterize the sMRI data when there is a new single subject The disadvantage is that this diagnostic method does not meet the need for clinical diagnosis, i.e., clinically, there is a need for every additional subject This paper presents a new ICA feature extraction model based on the feature extraction model. Add a single subject to make a diagnosis. The model assumes that each of the voxels of each sMRI image is independent of each other, and then uses a set of sMRI training data to train the demixing matrix so that when there is new sMRI data, it can be characterized by a well-trained demixing matrix Take, and follow up The simulation results show that the new ICA-based diagnostic method can achieve the same diagnostic accuracy as the original ICA diagnostic method, and More realistic diagnosis needs. Many linear feature extraction methods, including ICA, show good performance in the sMRI data analysis of AD, but these linear feature extraction methods require the original three-dimensional sMRI image After quantization, the data can be analyzed. The consequence of this process is that the spatial information in the original three-dimensional image data is destroyed, resulting in a loss of a large amount of effective information; at the same time, the amount of data of the sMRI image is very large, so that the number of vector dimensions obtained after quantization is not Often high, and the number of subjects is limited, which may result in a small sample problem (under In view of these problems, a non-correlated multi-linear principal component analysis (UMPCA) and a Laplacian value (laplacian sco) are proposed in this paper. the new classification and diagnosis method of re, ls) is a multi-linear subspace learning method, which can be used for extracting the sMRI data, and the three-dimensional image data can be represented by a direct tensor model and processed without the need to vectorize the original three-dimensional sMRI data, the space structure information of the original data is reserved, the problem caused by the backward quantization mentioned above is avoided; in addition, the extracted information can still have certain redundancy, and the process of selecting the LS feature selection after extracting the sMRI data characteristic information can further the redundant information is reduced, the computational complexity is reduced, the characteristics of high difference are selected, The accuracy of the post-diagnosis process is improved. The simulation results show that the UMPCA presented in this paper is compared with the existing diagnostic method.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41;R445.2
【参考文献】
相关期刊论文 前2条
1 赵小杰;龙志颖;郭小娟;姚力;;阿尔茨海默氏症研究中的磁共振成像数据分析[J];软件学报;2009年05期
2 吕彬;何晖光;赵明昌;吕科;张志强;卢光明;;基于磁共振图像的脑皮层厚度测量方法[J];中国医学影像技术;2008年06期
，本文编号：2351216
本文链接：https://www.wllwen.com/yixuelunwen/fangshe/2351216.html
