动态因果模型和Granger因果映射中模型选择的研究

发布时间：2018-05-03 02:18

本文选题：功能磁共振成像 + 效应链接　；参考：《浙江大学》2014年硕士论文

【摘要】：认知科学中,从功能整合角度研究大脑不同脑区间的交互作用具有重要的意义。在神经影像领域,功能磁共振成像(functional Magnetic Resonance Imaging,简称fMRI)凭借非侵入性,较高的空间分辨率成为从网络角度研究大脑功能的有效工具。在功能整合水平下,目前基于功能磁共振数据探测效应连接(effective connectivity)的动态因果模型(Dynamic Causal Modelling,简称DCM)和Granger因果映射(Granger Causality Mapping,简称GCM)长期竞争共存,都得到了广泛的关注。两种模型虽然基于不同的因果概念,但都是应用fMRI时间序列里时间信息来揭示脑区之间的有向的信息流动。基于这两种方法探测效应连接的过程中,模型选择是一个重要的问题。最优模型的选择,将会直接影响效应连接探测和结果的分析。 DCM中,不能准确设定的生理学参数会对模型选择结果和效应连接强度产生影响。基于F检验的Granger因果映射也在模型选择问题上出现了困难。最小描述长度(Minimum Description Length,简称MDL)是一种将奥卡姆剃刀形式化后的一种形式,可以避免模型过适应问题,因此被普遍应用在统计模型选择领域。本文旨在通过仿真实验和真实数据实验,探索在DCM的反演中,静态血容积比率(%)对于感兴趣区之间效应连接强度和整个效应连接网络的影响。在传统Granger因果映射中,把自回归模型定阶和模型选择纳入到一个模型选择框架中,通过引入最小描述长度准则选择最优模型。利用仿真数据和六只老鼠失神性癫痫发作时的BOLD信号在个体水平和组水平上分别进行有效链接的探测。结果证明在一个框架中,基于MDL的模型选择过程可以有效的避免人为确定置信区间带来的误差,在多模型的选择中可以减少F检测两两比较带来的计算量过大问题,并且对于噪声不敏感,从而克服了传统Granger因果映射的缺陷。
[Abstract]:In cognitive science, it is of great significance to study the interaction between different brain regions from the perspective of functional integration. In the field of neuroimaging, functional Magnetic Resonance imaging (fMRI) has become an effective tool for the study of brain function from a network perspective by virtue of its noninvasive and high spatial resolution. At the level of functional integration, the dynamic Causal Modeling (DCM) and the Granger causality Mapping (Granger Causality Mapping), which are based on the functional Magnetic Resonance data (fMRI) detection effect and effective connectivity, have been paid more and more attention. Although the two models are based on different causal concepts, they both use the time information in the fMRI time series to reveal the flow of information between brain regions. Model selection is an important problem in the process of detecting effect connection based on these two methods. The selection of the optimal model will directly affect the detection of the effect connection and the analysis of the results. In DCM, physiological parameters that cannot be accurately set will affect the model selection results and the effect connection strength. Granger causality mapping based on F test also presents difficulties in model selection. Minimum Description length is a formalized form of Occam razor, which can avoid the problem of model overadaptation, so it is widely used in the field of statistical model selection. The purpose of this paper is to explore the effect of static blood volume ratio on the connection strength of the effect between the regions of interest and the whole effect connection network in the inversion of DCM by means of simulation experiments and real data experiments. In the traditional Granger causality mapping, the autoregressive model order determination and model selection are incorporated into a model selection framework, and the optimal model is selected by introducing the minimum description length criterion. Simulated data and BOLD signals of six mice with apocalyptic seizures were detected at individual level and group level respectively. The results show that the model selection process based on MDL can effectively avoid the error caused by artificial determination of confidence interval in a framework, and reduce the computational complexity caused by the comparison of F detection in the selection of multiple models. And it is insensitive to noise, which overcomes the defect of traditional Granger causality mapping.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：O482.532;R741

【参考文献】