贝叶斯网络（Bayesian Network，BN）是一种基于概率的图论算法，可以有效的解决人工智能中的不确定问题，近年来被引入到功能磁共振（functional magnetic resonance imaging，fMRI）的研究中，它通过计算各节点间的条件依赖关系来构建脑区间的有效连接模式。由于BN是一种纯数据驱动的全局网络学习方法，而且计算简单，复杂度较低，因此慢慢得到神经影像工作者的广泛关注。但是，如何合适合理的将BN算法应用于构建基于fMRI数据的大脑有效连接网络仍需要进一步的研究。在fMRI数据分析中，基于不同的假设，BN组水平的分析方法分为个体结构、共同结构以及虚拟典型被试（Virtual Typical Structure，VTS）（平均（AVE）、主成分分析（PCA）和连接（CAT））三种，Li等人在2006年基于fMRI数据比较了三种组分析方法在fMRI数据分析中的适用性，却没有任何研究探讨基于VTS的三种组分析方法哪种更适合于fMRI数据。本文通过构建一组含有因果关系的模拟数据，比较了三种基于VTS组分析方法的性能，结果发现，基于CAT组分析方法在噪声水平较低时性能最好，当噪声水平增大到一定程度时，它的性能劣于基于AVE的组分析方法。此外，我们将三种组方法分别应用于构建静息态下正常被试默认网络的有效连接模式，结果显示基于AVE和基于CAT的组分析结果都进一步验证了后扣带回和内侧前额叶在整个默认网络中扮演了重要的角色。结合快速磁共振成像技术和实时的数据处理算法，实时fMRI技术应运而生并受到人们的广泛关注，本文将传统的高斯BN算法引入到实时研究中，首先通过模拟数据，从算法的鲁棒性、实时性和样本点长度验证了该算法实时化的可行性，结果显示，高斯BN算法具有较强的抗噪能力，较快的计算速度以及较短的样本点长度就可以达到较高的准确率，符合实时化的要求。然后，本文基于滑窗和累积窗两种方法完成高斯BN算法的实时化，通过比较两种实时算法，发现基于累积窗的实时算法优于基于滑窗的算法。最后，基于累积窗的实时高斯BN算法被应用于构建正常被试静息态下默认网络的有效连接模式，研究发现，被试在静息状态下大脑的有效连接模式基本保持稳定状态，但也会出现微小的改变。综上所述，本文在组水平上探讨和比较了三种基于VTS的组分析方法，为高斯BN算法在fMRI中的应用提供了参考。此外，将传统的高斯BN算法实时化，为实时fMRI的研究提供了一种新的技术手段。

Bayesian network (BN), which is a probability-based graph theory algorithm, has been widely used in solving the uncertainty problems in the field of artificial intelligence. In recent years, BN was introduced to explore the effective connectivity among brain regions from functional magnetic resonance imaging (fMRI). As no assumption of any prior model and providing a global representation of a system automatically learned from data in a completely exploratory manner, it gets more and more popular in the field of neuroimaging. However, it still needs further researches that how to appropriately apply BN algorithm in fMRI data analysis.In fMRI research, group analysis was divided into three categories, namely individual structure (IS) approach, common structure (CS) approach and virtual typical subjects (VTS) approach. The VTS approach needs to get a virtual typical subject by using some methods, as average group data (AVE), seeking first principal component of group data by using principle component analysis (PCA) and concatenate all subjects (CAT) in the group, to represent the whole group. In 2006, according to comparing the IS, CS and VTS approach, Li et al found that none of three group analysis methods is superior over the others. However, the question that which VTS-based group analysis is most appropriate in fMRI study is still unclear. In our study, based on a set of simulated data, we used Gaussian BN and compared the performances of three VTS-based group analysis. The findings demonstrated that the AVE-based approach was more suitable for the data with high level noise and the CAT-based approach performed better in the situation that the data contained low level noise. Then, based on 12 normal subjects, we respectively used three group analysis in constructing the effective connectivity of the default mode network and the results of AVE-based and CAT-based approaches show that the posterior cingulate cortex (PCC) and medial prefrontal cortex (MPFC) play a vital role in the whole default mode network (DMN).For the real-time fMRI, our study investigated Gaussian BN framework and implemented it into real-time fMRI data analysis. We firstly investigates the feasibility of Gaussian BN including its robustness to the noise, computational efficiency and the findings show that it meets the requirements of real-time algorithm. Then, using either the sliding or the cumulative window method, the real-time Gaussian BN was implemented. Combining indexes of structure learning error and computational speed, we compared sliding window method and cumulative window method and find that the latter one performs better. Finally, the real-time Gaussian BN based on cumulative window was applied to real fMRI data and the results demonstrated that the effective connectivity mode of DMN in the resting state is substantially unchanged.