基于脑网络社团结构和深度学习的自闭症诊断研究
发布时间：2018-11-15 15:32
【摘要】：虽然科学家们已经探明了自闭症患者的大脑与健康人的大脑有着明显的不同,但是两者之间的具体差异却一直存在争议。因为这个原因,医生无法从一个客观的视角来对自闭症进行诊断。事实上,自闭症的诊断在医学界是一个十分主观的过程,由于诊断是以行为为导向,以诊断表为准绳,因此目前并不能把医学评估数据作为诊断的依据。随着科学的进步,一些新方法为诊断自闭症指出了新的方向。近年来,科学家们广泛应用静息态功能核磁共振技术(rs-fMRI,Resting-state Functional Magnetic Resonance Imaging)探索脑部疾病。这是因为fMRI能够在无创的条件下,通过检测血氧水平获得更高分辨率的图像。正因如此,使用fMRI数据构建大脑功能网络,通过进一步分析而得到被试大脑的特征无疑成为了一种十分有效的方法。脑功能网络是一个复杂网络,因此具备了复杂网络的属性。在这些属性中,社团结构占据着极其重要的地位,当然也存在于脑网络中。由于相似的节点处在同样的社团中,而相异节点处在不同的社团中,因此社团划分也可以看作是在复杂网络中寻找相似结构的一种方法。然而,社团划分方法繁多,评价标准各有不同,加之社团划分问题本身是一个NP-hard问题,因此找到一种适用于脑功能网络的社团划分方法是很困难的。深度学习是机器学习方法的一种。它通过模仿人类的思考方式对数据进行重新整理,以此得到事物的更高维更抽象的特征。当前使用深度学习最引人注意和最成功的例子是语音识别和图像识别。深度学习还尤其擅长分类问题,这也使得众多科学家把深度学习方法用于疾病诊断问题上。然而,获取有效的特征数据和选择深度分类器是两个重要的问题。针对以上这些问题,本文通过设计和使用新的社团划分算法GAcut(Genetic Algorithm Cut)提取和分析了自闭症和对照组的脑网络社团结构特征。然后以此作为依据,使用深度降噪自动编码器对自闭症和对照组进行区分,最终得到了较高的诊断准确率。本文的主要工作如下:(1)运用目前流行的方法对rs-fMRI数据进行了预处理,并在此基础上分别设计个体相关性矩阵和组相关性矩阵对一个被试和一组被试构建了脑功能网络。(2)为了准确的划分脑网络的社团结构,本文在遗传算法和模块度Q的基础上设计和实现了算法GAcut。在真实网络和fMRI数据上的实验表明,GAcut算法有效。(3)使用GAcut算法分别对正常被试和自闭症被试的脑功能网络进行社团划分并证明了脑网络具备社团属性,然后结合自闭症的病理对患者的脑网络社团结构进行了医学上的相关性分析,详细描述了自闭症患者和正常人脑网络社团结构的不同之处和可能造成这些差异的病理原因。最后发现自闭症和对照组之间的社团结构的具体差异可以通过标准化互信息(Normalized Mutual Information,NMI)定量的展现出来。(4)通过构建NMI统计矩阵将所有被试的脑网络社团结构特征浓缩到一个低维度的矩阵中,然后将其作为深度降噪自动编码器的输入,从而对自闭症和对照组进行区分。大量对比实验表明,使用NMI统计矩阵作为深度降噪自动编码器的输入,不但能够得到更加准确的诊断结果,而且时间成本也更低。
[Abstract]:Although scientists have found that the brain of a patient with autism is distinct from the brain of a healthy person, the specific difference between the two has been controversial. For this reason, doctors can't diagnose autism from an objective perspective. In fact, the diagnosis of autism is a very subjective process in the medical field, because the diagnosis is guided by behavior, and the diagnosis table is the quasi-rope, so the medical evaluation data cannot be used as the basis for diagnosis. With the progress of science, some new approaches have identified new directions for the diagnosis of autism. In recent years, scientists have been widely used in the exploration of brain diseases by resting-state magnetic resonance (rs-fMRI). This is because fMRI is able to obtain higher resolution images by detecting blood oxygen levels without invasive conditions. For this reason, the use of fMRI data to construct the brain function network is a very effective way to get the characteristics of the brain to be tested by further analysis. The brain function network is a complex network, so it has the properties of complex networks. In these attributes, the community structure plays a very important role, of course in the brain network. Because similar nodes are in the same community, and different nodes are in different communities, the division of associations can also be seen as a method of finding similar structures in a complex network. However, it is difficult to find a kind of community division method which is suitable for the brain function network, because the division method of the community is different, the evaluation standard is different, and the problem of the division of the community itself is an NP-hard problem. Depth learning is one of the methods of machine learning. It makes the data refresh by simulating the human thinking mode, so as to obtain the higher-dimensional and more abstract features of the object. The most interesting and successful examples of current use of depth learning are speech recognition and image recognition. Deep learning is also particularly good at the classification problem, which also makes it possible for many scientists to use depth-learning methods in the diagnosis of disease. However, the acquisition of valid feature data and the selection of a depth classifier is two important issues. In view of the above problems, this paper extracts and analyzes the structural features of the brain network community in the autistic and control group by the design and use of the new community-based algorithm GAcut (Genetic Algorithm Cut). and then using the depth noise reduction automatic coder to distinguish the autism and the control group as the basis, and finally the high diagnosis accuracy rate is obtained. The main work of this paper is as follows: (1) Using the current method to pre-process the rs-fMRI data, and on this basis, the individual correlation matrix and the group correlation matrix are designed to construct the brain function network. (2) In order to accurately classify the community structure of the brain network, this paper designs and implements the algorithm GAut based on the genetic algorithm and the module degree Q. The experiments on real network and fMRI data show that the GAut algorithm is effective. (3) The brain function network of the normal and autistic people is divided by the GAcut algorithm, and the association property of the brain network is proved, and then the association analysis of the brain network community structure of the patient is carried out in combination with the pathology of the autism. The differences in the structure of the autistic and normal human brain networks and the possible causes of these differences are described in detail. Finally, it is found that the specific difference of the community structure between the autism and the control group can be quantified by the standardized mutual information (NMI). and (4) concentrating all the tested brain network community structure features into a low-dimension matrix by constructing the NMI statistical matrix, and then using the NMI statistical matrix as the input of the depth noise reduction automatic encoder, thereby distinguishing the autism and the control group. A large number of comparison experiments show that the NMI statistical matrix is used as the input of the depth noise reduction automatic encoder, so that the accurate diagnosis result can be obtained, and the time cost is lower.
【学位授予单位】：江苏大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R749;O157.5
本文编号：2333694
[Abstract]:Although scientists have found that the brain of a patient with autism is distinct from the brain of a healthy person, the specific difference between the two has been controversial. For this reason, doctors can't diagnose autism from an objective perspective. In fact, the diagnosis of autism is a very subjective process in the medical field, because the diagnosis is guided by behavior, and the diagnosis table is the quasi-rope, so the medical evaluation data cannot be used as the basis for diagnosis. With the progress of science, some new approaches have identified new directions for the diagnosis of autism. In recent years, scientists have been widely used in the exploration of brain diseases by resting-state magnetic resonance (rs-fMRI). This is because fMRI is able to obtain higher resolution images by detecting blood oxygen levels without invasive conditions. For this reason, the use of fMRI data to construct the brain function network is a very effective way to get the characteristics of the brain to be tested by further analysis. The brain function network is a complex network, so it has the properties of complex networks. In these attributes, the community structure plays a very important role, of course in the brain network. Because similar nodes are in the same community, and different nodes are in different communities, the division of associations can also be seen as a method of finding similar structures in a complex network. However, it is difficult to find a kind of community division method which is suitable for the brain function network, because the division method of the community is different, the evaluation standard is different, and the problem of the division of the community itself is an NP-hard problem. Depth learning is one of the methods of machine learning. It makes the data refresh by simulating the human thinking mode, so as to obtain the higher-dimensional and more abstract features of the object. The most interesting and successful examples of current use of depth learning are speech recognition and image recognition. Deep learning is also particularly good at the classification problem, which also makes it possible for many scientists to use depth-learning methods in the diagnosis of disease. However, the acquisition of valid feature data and the selection of a depth classifier is two important issues. In view of the above problems, this paper extracts and analyzes the structural features of the brain network community in the autistic and control group by the design and use of the new community-based algorithm GAcut (Genetic Algorithm Cut). and then using the depth noise reduction automatic coder to distinguish the autism and the control group as the basis, and finally the high diagnosis accuracy rate is obtained. The main work of this paper is as follows: (1) Using the current method to pre-process the rs-fMRI data, and on this basis, the individual correlation matrix and the group correlation matrix are designed to construct the brain function network. (2) In order to accurately classify the community structure of the brain network, this paper designs and implements the algorithm GAut based on the genetic algorithm and the module degree Q. The experiments on real network and fMRI data show that the GAut algorithm is effective. (3) The brain function network of the normal and autistic people is divided by the GAcut algorithm, and the association property of the brain network is proved, and then the association analysis of the brain network community structure of the patient is carried out in combination with the pathology of the autism. The differences in the structure of the autistic and normal human brain networks and the possible causes of these differences are described in detail. Finally, it is found that the specific difference of the community structure between the autism and the control group can be quantified by the standardized mutual information (NMI). and (4) concentrating all the tested brain network community structure features into a low-dimension matrix by constructing the NMI statistical matrix, and then using the NMI statistical matrix as the input of the depth noise reduction automatic encoder, thereby distinguishing the autism and the control group. A large number of comparison experiments show that the NMI statistical matrix is used as the input of the depth noise reduction automatic encoder, so that the accurate diagnosis result can be obtained, and the time cost is lower.
【学位授予单位】：江苏大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R749;O157.5
【参考文献】
相关期刊论文 前3条
1 邹小兵;;儿童孤独症诊断与治疗新动向[J];中国儿童保健杂志;2012年04期
2 邓赐平,刘明;自闭症诊断理论的进展与困顿[J];中国特殊教育;2005年03期
3 邓赐平,刘明;自闭症的认知神经发展研究:回顾与展望[J];华东师范大学学报(教育科学版);2004年03期
，本文编号：2333694
本文链接：https://www.wllwen.com/kejilunwen/yysx/2333694.html
