Spatial independent component analysis of functional MRI time-series: To what extent do results depend on the algorithm used?
Spatial independent component analysis of functional MRI time-series: To what extent do results depend on the algorithm used?
复制标题
DOI：
10.1002/hbm.10034
复制
发表时间：
2002-07-01
影响因子：
4.8
通讯作者：
Di Salle, F
中科院分区：
文献类型：
--
作者：
Esposito, F;Formisano, E;Di Salle, F
独立分量分析(ICA)已被成功地用于将功能磁共振(FMRI)时间序列分解成一组激活图及其相关的时间进程。在神经网络的文献中已经提出了几种独立分量分析算法。将这些算法应用到功能磁共振成像中，可能会导致不同的大脑激活的空间或时间读数。比较了目前用于fMRI时间序列空间ICA(SICA)的两种ICA算法：Infomax(Bell和Sejnowski[1995]：NeuralComput7：1004-1034)算法和定点ICA(Hyvarinen[1999]：ADV NeuroinProc Syst 10：273-279)算法。我们使用一系列测量方法评估了基于Infomax和固定点的模拟运动和真实运动和视觉激活fMRI时间序列的SICA分解。对数似然率(McKeown等人[1998]：Hum Brain Mapp 6：160-188)被用来衡量估计的独立源与数据统计结构的显著匹配程度；接收器操作特征(ROC)和线性相关分析被用来评估算法在估计模拟和真实激活的空间布局和时间动力学方面的准确性；集群大小计算和组件内残留高斯噪声项的估计被用来检查ICA组件的解剖结构和评估降噪能力。虽然这两种算法都产生了非常准确的结果，但只要以推断统计作为基准，定点算法在空间和时间精度方面都优于Infomax。相反，Infomax SICA在ICA模型的全局估计和降噪能力方面更具优势。由于其适应性，Infomax方法似乎更适合于研究无法通过推理技术预测或充分建模的激活现象。(C)2002年Wiley-Liss，Inc.
Independent component analysis (ICA) has been successfully employed to decompose functional MRI (fMRI) time-series into sets of activation maps and associated time-courses. Several ICA algorithms have been proposed in the neural network literature. Applied to fMRI, these algorithms might lead to different spatial or temporal readouts of brain activation. We compared the two ICA algorithms that have been used so far for spatial ICA (sICA) of fMRI time-series: the Infomax (Bell and Sejnowski [1995]: Neural Comput 7:1004-1034) and the Fixed-Point (Hyvarinen [1999]: Adv Neural Inf Proc Syst 10:273-279) algorithms. We evaluated the Infomax- and Fixed Point-based sICA decompositions of simulated motor, and real motor and visual activation fMRI time-series using an ensemble of measures. Log-likelihood (McKeown et al. [1998]: Hum Brain Mapp 6:160-188) was used as a measure of how significantly the estimated independent sources fit the statistical structure of the data; receiver operating characteristics (ROC) and linear correlation analyses were used to evaluate the algorithms' accuracy of estimating the spatial layout and the temporal dynamics of simulated and real activations; cluster sizing calculations and an estimation of a residual gaussian noise term within the components were used to examine the anatomic structure of ICA components and for the assessment of noise reduction capabilities. Whereas both algorithms produced highly accurate results, the Fixed-Point outperformed the Infomax in terms of spatial and temporal accuracy as long as inferential statistics were employed as benchmarks. Conversely, the Infomax sICA was superior in terms of global estimation of the ICA model and noise reduction capabilities. Because of its adaptive nature, the Infomax approach appears to be better suited to investigate activation phenomena that are not predictable or adequately modelled by inferential techniques. (C) 2002 Wiley-Liss, Inc.