基于t-SNE算法的脑网络状态观测矩阵降维及可视化平台研究

发布时间：2018-03-16 05:09

本文选题：数据降维　切入点：t-SNE　出处：《昆明理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：功能磁共振成像是一种有效研究人脑功能和结构的非介入技术手段,基于功能磁共振成像的血氧水平依赖信号重构的脑功能网络为研究人脑动态属性和特征提供了一种有效的方法。为了深入研究脑网络的动态特征,以血氧水平依赖信号时间序列为对象展开了脑区相关性分析并构建了全脑网络状态观测矩阵,但其维数非常高,这给脑网络动态特征的辨识带来了很大的困难,因此研究脑网络状态观测矩阵的降维问题是深入研究脑网络动态特征的基础。本文主要以静息态磁共振采集得到的血氧水平依赖信号时间序列所构建的全脑脑网络状态观测矩阵为对象,对其在低维空间内的嵌入和可视化为目标展开研究。针对使用主成分分析、等距映射、局部线性嵌入等主流降维算法得到结果较差的问题,对其数据特征展开研究和分析,给出了一种基于t分布随机领域嵌入算法的高维脑网络状态观测矩阵降维方法,实验结果显示,相对于其它降维算法,使用该算法可以对脑网络状态观测矩阵达到更好的降维效果。在此基础上设计实现了基于Python的降维算法及可视化平台,并完成了该平台与现有脑数据处理功能和技术平台的整合,从而为脑网络动态特性的进一步研究提供了理论基础和技术基础。
[Abstract]:Functional magnetic resonance imaging (fMRI) is an effective non-interventional technique for studying the function and structure of human brain. The functional network based on functional magnetic resonance imaging (fMRI) provides an effective method for studying the dynamic properties and characteristics of human brain. Based on the time series of the blood oxygen level dependent signal, the correlation analysis of the brain region and the construction of the state observation matrix of the whole brain network are carried out, but its dimension is very high, which brings great difficulties to the identification of the dynamic characteristics of the brain network. Therefore, to study the dimensionality reduction of the state observation matrix of brain network is the basis for further studying the dynamic characteristics of brain network. In this paper, the whole brain reticulum is constructed from the time series of blood oxygen level dependent signal acquired by resting magnetic resonance. The observation matrix of the state of the network is the object, Aiming at the problem of using principal component analysis (PCA), isometric mapping, local linear embedding and other mainstream dimensionality reduction algorithms to get poor results, the data features are studied and analyzed. A state observation matrix reduction method for high-dimensional brain networks based on t-distribution random domain embedding algorithm is presented. The experimental results show that compared with other dimensionality reduction algorithms, The algorithm can achieve better dimensionality reduction effect on the state observation matrix of brain network. On this basis, the dimensionality reduction algorithm and visualization platform based on Python are designed and implemented. The integration of the platform with the existing brain data processing function and technology platform is completed, which provides a theoretical and technical basis for the further study of the dynamic characteristics of the 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】：TP274;R445.2

【参考文献】