基于DAE的脑网络状态观测矩阵降维方法研究

发布时间：2018-03-31 18:55

本文选题：脑功能网络　切入点：状态观测矩阵　出处：《昆明理工大学》2017年硕士论文

【摘要】：核磁共振成像技术为研究大脑的特性提供了有利的手段,其中基于血氧水平依赖的静息态功能磁共振成像由于具有较高的时间和空间分辨率,为深入研究人脑功能的动态特性提供了一种重要方法,而基于它的脑网络重构技术也成为了研究人脑特性的有力工具之一。鉴于人脑网络的复杂性,在提取人脑网络的状态特征时,由于所构建的人脑网络状态观测矩阵维数过高,所以很难识别它的主要特性,因此对其展开降维和聚类方法的研究是非常有必要的。基于上述现状,本文以深度学习理论为基础,以脑网络状态观测矩阵的降维和聚类方法为重点进行了以下研究:提出了一种基于深度自动编码器的脑网络状态观测矩阵降维方法,构建并实现了一个基于5层受限玻尔兹曼机的深度自动编码器系统,可以将高维的脑网络特征数据映射到低维空间内,从而为高维脑网络状态观测矩阵的降维实现提供了一个新的解决思路。为了进一步验证该方法的可靠性,采用自组织映射方法对降维后的低维空间脑网络状态观测向量进行聚类,最后通过实验和结果分析验证了基于深度自动编码器的降维方法的有效性。该方法为下一步深入研究人脑网络动态特性提供了必要的基础。
[Abstract]:Magnetic resonance imaging (MRI) provides a useful tool for the study of brain properties, in which resting functional magnetic resonance imaging based on the level of blood oxygen has higher temporal and spatial resolution. It provides an important way to study the dynamic characteristics of human brain function, and the brain network reconstruction technology based on it has become one of the powerful tools to study the characteristics of human brain, in view of the complexity of human brain network, When extracting the state characteristics of the human brain network, it is difficult to identify its main characteristics because the dimension of the state observation matrix of the human brain network is too high. Therefore, it is necessary to study the methods of reducing and clustering. Based on the above situation, this paper is based on the theory of depth learning. This paper focuses on the reduction and clustering of the state observation matrix of the brain network. A method of reducing the dimension of the state observation matrix of the brain network based on the depth automatic encoder is proposed. A depth automatic encoder system based on a 5-layer constrained Boltzmann machine is constructed and implemented. The feature data of high-dimensional brain network can be mapped to low-dimensional space. This provides a new solution for the dimensionality reduction of the state observation matrix of high-dimensional brain network, and further verifies the reliability of the method. The state observation vector of low-dimensional spatial brain network after dimensionality reduction is clustered by self-organizing mapping method. Finally, the effectiveness of the dimensionality reduction method based on the depth automatic encoder is verified by experiments and results analysis, which provides a necessary foundation for the further study of the dynamic characteristics of the human brain network.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R445.2;TP391.41

【参考文献】