Original contribution
Detecting functional connectivity in the resting brain: a comparison between ICA and CCA

https://doi.org/10.1016/j.mri.2006.09.032Get rights and content

Abstract

Independent component analysis (ICA) and cross-correlation analysis (CCA) are general tools for detecting resting-state functional connectivity. In this study, we jointly evaluated these two approaches based on simulated data and in vivo functional magnetic resonance imaging data acquired from 10 resting healthy subjects. The influence of the number of independent components (maps) on the results of ICA was investigated. The influence of the selection of the seeds on the results of CCA was also examined. Our results reveal that significant differences between these two approaches exist. The performance of ICA is superior as compared with that of CCA; in addition, the performance of ICA is not significantly affected by structured noise over a relatively large range. The results of ICA could be affected by the number of independent components if this number is too small, however. Converting the spatially independent maps of ICA into z maps for thresholding tends to overestimate the false-positive rate. However, the overestimation is not very severe and may be acceptable in most cases. The results of CCA are dependent on seeds location. Seeds selected based on different criteria will significantly affect connectivity maps.

Introduction

Resting-state functional magnetic resonance imaging (fMRI) is instrumental for investigating the functional connectivity of the human brain. It has been successfully applied to study motor [1], [2], [3], visual [4], auditory [5] and language [6] systems for healthy subjects and patients [7], [8], [9]. Most resting-state fMRI processings performed so far used cross-correlation analysis (CCA), in which neuronal activations are detected by evaluating the correlation between the time course of each voxel and a reference function. Although often determined based on experimental designs (i.e., stimulation presentations) and brain hemodynamics for task-induced activation studies, the reference function is much more difficult to be constructed for resting-sate fMRI studies due to the very limited prior knowledge. A common practice is then using the averaged time course of a preselected brain region as the reference function. Thus, the shape, size and location of the seeds will affect the reference function and, consequently, the outcomes of the resulting connectivity map. Currently, the seeds region is commonly determined based on either brain anatomies [10], [11] or additional functional activation studies [2]. The consistency of connectivity maps reported by different laboratories may be deteriorated with the use of different seed selection methods. In the meantime, the seeds approach may also amplify intersubject variability in resting-state connectivity maps because of intersubject variability in anatomy and variability in anatomy–function correspondence.

An alternative to CCA, independent component analysis (ICA) is an exploratory method and requires no reference function or predefined seed [12], [13], [14], [15], [16]. It is a statistical technique that drives a set of measurement data into a number of independent signals. First introduced into neuroimaging data processing by McKeown et al. [15] to determine the spatial and temporal characteristics of task-induced brain activations, ICA gradually became one of the widely used techniques for fMRI data analyses. It assumes that all components of brain activation signals are spatially and/or temporally independent. By maximizing spatial independence or temporal independence, ICA can be classified as spatial ICA (sICA) [15], [17] or as temporal ICA (tICA) [17], [18]. Emphasizing on different aspects (spatial or temporal) of brain activation, sICA and tICA may report different results. The similarities and differences of sICA and tICA have been evaluated and discussed by Calhoun et al. [17]. When both spatial independence assumption and temporal independence assumption are met, sICA and tICA report similar results [17]. In addition to assessing the performances of different ICA techniques, studies have also evaluated the merits of ICA relative to CCA [19] and the temporal clustering technique [20] for task-induced activation.

Requiring no prior knowledge regarding spatial or temporal patterns of brain response, ICA should be best suitable for analyzing resting-state fMRI data, which have very limited prior knowledge of activation. The first application of ICA in resting-state fMRI data analysis was reported by Kiviniemi et al. [7] in 2003. Subsequent applications showed that ICA is able to detect functional connectivity in both primary (motor, visual and somatosensory) and higher-order [21], [22], [23] brain regions. It has also been demonstrated that ICA is capable of formulating the same kind of functional connectivity (e.g., motor systems) into a single map [22].

Although CCA and ICA are capable of resting-state analyses, each has its distinctive advantages and pitfalls. The major shortcoming of CCA is the dependence of the resulting connectivity maps on the selection of the seeds region. The major drawbacks of ICA are that (1) it is often difficult to extract true activation maps from a large number of independent components (ICs) and (2) it is often difficult to determine threshold. How to determine the optimal number of independent maps and how to determine the optimal threshold have become open questions. It therefore should be valuable to assess the effects of the number of ICs, threshold and seeds selection on the resulting connectivity maps and to jointly evaluate the performances of ICA and CCA for resting-state fMRI data analyses. To our best knowledge, such evaluations have not been performed. In this study, we first evaluated the effects of seeds selection on CCA connectivity maps; we then investigated the influences of the number of ICs and threshold on the results of ICA. Finally, we evaluated the relative merits of CCA and ICA based on simulated data and real MRI data.

Section snippets

Simulation

Simulated data were used to investigate the influences of noise level and threshold on the performances of ICA and CCA. The effect of the number of independent maps on the ICA results was also studied. The echo-planar imaging (EPI) scans and the simulated signals used in this section are shown in Fig. 1. A volume of three-slice EPI scans (each slice has 72×72 voxels) was replicated 300 times to simulate 300 time points of noise-free fMRI data. Three simulated signals (Signals A–C) have been

Results

The influences of noise level on the performances of ICA and CCA were first investigated using the simulated data. Fig. 2 shows the effects of noise levels on the detection of functional connectivity networks using ICA and CCA. The threshold was fixed at 3.5. Both ICA and CCA performed well in detecting the signals when the random noise level was low (CNR=1.0; K=0.5). When the random noise was high (CNR=0.5; K=2; Fig. 2A and E), both approaches had difficulty with detecting the interesting

Discussion and conclusion

ICA and CCA are both capable of detecting resting-state functional connectivity networks. However, the performances of ICA and CCA are not necessarily always the same. ICA significantly differs from CCA in terms of the size and location of the detected networks. Importantly, ICA significantly differs from CCA in overall performance and in handling aliased cardiac signals. Based on our results, the overall performance of ICA appears to be significantly better than that of CCA for resting-state

Acknowledgment

This work was financially supported by the NSF (Grant No. BCS 05-09626) and the NIH (Grant No. 5 RO1 NS046082).

References (29)

  • R. Matsumoto et al.

    Functional connectivity in the human language system: a cortico-cortical evoked potential study

    Brain

    (2004)
  • T. Stein et al.

    Functional connectivity in the thalamus and hippocampus studied with functional MR imaging

    AJNR Am J Neuroradiol

    (2001)
  • Y. He et al.

    Detecting functional connectivity of the cerebellum using low frequency fluctuations (LFFs). Medical Image Computing and Computer-Assisted Intervention—MICCAI

    LNCS

    (2004)
  • A.J. Bell et al.

    An information-maximization approach to blind separation and blind deconvolution

    Neural Comput

    (1995)
  • Cited by (86)

    View all citing articles on Scopus
    View full text