de Pasquale F, Del Gratta C, Romani G L
ITAB, Institute for Advanced Biomedical Technologies, University G. D'Annunzio, Chieti, Italy.
Neuroimage. 2008 Aug 1;42(1):99-111. doi: 10.1016/j.neuroimage.2008.04.235. Epub 2008 Apr 29.
In this work an Empirical Markov Chain Monte Carlo Bayesian approach to analyse fMRI data is proposed. The Bayesian framework is appealing since complex models can be adopted in the analysis both for the image and noise model. Here, the noise autocorrelation is taken into account by adopting an AutoRegressive model of order one and a versatile non-linear model is assumed for the task-related activation. Model parameters include the noise variance and autocorrelation, activation amplitudes and the hemodynamic response function parameters. These are estimated at each voxel from samples of the Posterior Distribution. Prior information is included by means of a 4D spatio-temporal model for the interaction between neighbouring voxels in space and time. The results show that this model can provide smooth estimates from low SNR data while important spatial structures in the data can be preserved. A simulation study is presented in which the accuracy and bias of the estimates are addressed. Furthermore, some results on convergence diagnostic of the adopted algorithm are presented. To validate the proposed approach a comparison of the results with those from a standard GLM analysis, spatial filtering techniques and a Variational Bayes approach is provided. This comparison shows that our approach outperforms the classical analysis and is consistent with other Bayesian techniques. This is investigated further by means of the Bayes Factors and the analysis of the residuals. The proposed approach applied to Blocked Design and Event Related datasets produced reliable maps of activation.
在这项工作中，我们提出了一种经验马尔可夫链蒙特卡罗贝叶斯方法来分析功能磁共振成像（fMRI）数据。贝叶斯框架很有吸引力，因为在图像和噪声模型的分析中都可以采用复杂模型。在这里，通过采用一阶自回归模型来考虑噪声自相关，并假设一个通用的非线性模型用于与任务相关的激活。模型参数包括噪声方差和自相关、激活幅度以及血流动力学响应函数参数。这些参数在每个体素处根据后验分布的样本进行估计。通过一个4D时空模型纳入先验信息，该模型用于描述空间和时间上相邻体素之间的相互作用。结果表明，该模型可以从低信噪比数据中提供平滑的估计，同时能够保留数据中的重要空间结构。我们进行了一项模拟研究，探讨了估计的准确性和偏差。此外，还给出了关于所采用算法收敛诊断的一些结果。为了验证所提出的方法，我们将结果与标准广义线性模型（GLM）分析、空间滤波技术和变分贝叶斯方法的结果进行了比较。这种比较表明，我们的方法优于经典分析，并且与其他贝叶斯技术一致。通过贝叶斯因子和残差分析进一步对此进行了研究。所提出的方法应用于组块设计和事件相关数据集时，生成了可靠的激活图。
