Original contribution
Performance of blind source separation algorithms for fMRI analysis using a group ICA method

https://doi.org/10.1016/j.mri.2006.10.017Get rights and content

Abstract

Independent component analysis (ICA) is a popular blind source separation technique that has proven to be promising for the analysis of functional magnetic resonance imaging (fMRI) data. A number of ICA approaches have been used for fMRI data analysis, and even more ICA algorithms exist; however, the impact of using different algorithms on the results is largely unexplored. In this paper, we study the performance of four major classes of algorithms for spatial ICA, namely, information maximization, maximization of non-Gaussianity, joint diagonalization of cross-cumulant matrices and second-order correlation-based methods, when they are applied to fMRI data from subjects performing a visuo-motor task. We use a group ICA method to study variability among different ICA algorithms, and we propose several analysis techniques to evaluate their performance. We compare how different ICA algorithms estimate activations in expected neuronal areas. The results demonstrate that the ICA algorithms using higher-order statistical information prove to be quite consistent for fMRI data analysis. Infomax, FastICA and joint approximate diagonalization of eigenmatrices (JADE) all yield reliable results, with each having its strengths in specific areas. Eigenvalue decomposition (EVD), an algorithm using second-order statistics, does not perform reliably for fMRI data. Additionally, for iterative ICA algorithms, it is important to investigate the variability of estimates from different runs. We test the consistency of the iterative algorithms Infomax and FastICA by running the algorithm a number of times with different initializations, and we note that they yield consistent results over these multiple runs. Our results greatly improve our confidence in the consistency of ICA for fMRI data analysis.

Introduction

Independent component analysis (ICA) is a popular blind source separation (BSS) technique that has been successfully applied in the analysis of functional magnetic resonance imaging (fMRI) data [1]. The majority of ICA applications for fMRI use the Infomax [2] algorithm or the FastICA [3] algorithm. A few comparative studies [4], [5], [6] have compared Infomax to FastICA and/or a BSS approach using second-order statistics, the Molgedey–Schuster algorithm [7]. The Statistical Parametric Mapping (SPM2) software package [8] has been used extensively to study fMRI data; however, it requires that model time courses be parameterized a priori; hence, it can be used primarily in the study of task-related components. ICA, on the other hand, extracts both spatially independent brain maps and their activation time courses from fMRI data without a priori specification of time courses and is shown to be promising for the study of components whose time courses are not known or cannot be easily modeled (see, e.g., Calhoun et al. [9]).

Our purpose in this paper is to investigate the performance of different ICA algorithms that are commonly used to analyze fMRI data along with SPM. We use fMRI data from 12 subjects performing a visuo-motor task and, instead of entering each subjects' data into a separate ICA analysis, we use a group ICA technique [10] to estimate one set of components and then back-reconstruct from an aggregate mixing matrix to obtain individual subject maps. This method has the advantage of ordering components in different subjects in the same way, which is a tedious task when individual ICA analyses are performed because, unlike the general linear model (GLM) used by SPM and other univariate methods that can easily be generalized to group analysis, subjects as processed by ICA do not share common regressors, thus making it difficult to match components across subjects. In Calhoun et al. [10], it has been shown that back-reconstructed ICA results using our group approach are similar to those performed for each individual subject, which shows that the group ICA technique allows for cross-subject variability.

To date, several approaches have been presented to perform group ICA [10], [11], [12]. Schmithorst and Holland [13], using the FastICA algorithm on simulated group data, compared three group ICA methods and noted that the group ICA technique proposed in Calhoun et al. [10] provided the best overall performance in terms of an accurate estimation of sources and associated time courses. In our study, we used this group ICA approach, which has been implemented in the user-friendly environment of the group ICA of fMRI toolbox (GIFT) [14]. We use this toolbox to study the performances of four BSS/ICA algorithms when applied to fMRI data. Thus far, nine algorithms have been implemented in GIFT; the algorithms we have used for the comparative study in this paper are: Infomax (maximum likelihood) [2], FastICA [3], joint approximate diagonalization of eigenmatrices (JADE) [15] and eigenvalue decomposition (EVD) [16]. This selection includes an algorithm from each of the major approaches to performing ICA, namely, information maximization, maximization of non-Gaussianity, joint diagonalization of cross-cumulant matrices and second-order correlation-based methods.

Infomax and FastICA are iterative algorithms that require proper initialization and setting of parameters. To enable an efficient use of ICA for fMRI data, the issue of consistency of the estimates needs to be addressed. ICASSO [17] is a toolbox that was developed to evaluate similarity among multiple estimates of the FastICA algorithm. We incorporate ICASSO into GIFT to evaluate consistency in estimates of Infomax and FastICA.

The algorithms we examine use different methodologies to achieve separation of the sources causing the estimates from different algorithms to differ. We propose a number of evaluation techniques to compare estimation results among ICA algorithms. One of the techniques we use involves a comparison of estimates with masks that include expected areas of activation. Another comparison method uses Infomax as a reference and calculates the difference images of Infomax and the other algorithms. This method enables us to easily evaluate differences in estimation among algorithms. We also compare the performance of the ICA algorithms based on the correlation of time courses to model paradigms. For time courses, using t tests, we also check for significant departures from Infomax results. Finally, we enter the results from the four algorithms into ICASSO and report how estimates from the different algorithms cluster based on correlation.

Our results show that the ICA and GLM methods generally yield similar estimates for the task-related components. The iterative ICA algorithms yield consistent results over multiple runs for the components examined. The two iterative algorithms that utilize higher-order statistical information, namely, Infomax and FastICA, give the best overall performance. JADE performs comparably well; however, its estimated area of activation is usually smaller than that of Infomax and FastICA. The EVD approach, although very fast, yields results that are quite different from those given by the other algorithms, and it is determined to be a useful analysis tool only when the spectra of the components are different.

Section snippets

Participants and experimental paradigms

Twelve right-handed participants with normal vision (six females, six males; average age, 30 years) participated in the study. Participants provided written informed consent to a protocol approved by the Hartford Hospital Institutional Review Board. The subjects performed a visuo-motor task involving two identical but spatially offset, periodic, visual stimuli, shifted by 20 s from each other. The visual stimulus was presented alternately to the right visual hemifield and the left visual

Comparison of SPM and ICA results

At P<.0005 (t=4.4), for the LTR component, the results of ICA (Infomax) and SPM are quite similar. However, at this threshold, the RTR component gets a cutoff below the threshold. When the threshold is lowered to P<.005 (t=3.11) for SPM, the RTR area barely passes the threshold; when it is further lowered to P<.05 (t=1.6), the SPM result is comparable to the ICA result. Hence, we can conclude that the two components are estimated similarly by both methods with respect to the areas detected,

Discussion

We experimentally studied the estimation results of four spatial ICA/BSS algorithms when applied to fMRI analysis. We compared the estimation results of ICA against those obtained with SPM for task-related components. Using Infomax, we obtained regions that were similar to those obtained using SPM; however, Infomax estimates components with a higher contrast-to-noise ratio, which may be due to ICA algorithms not constraining time courses. SPM, on the other hand, allows for explicit tests on the

Acknowledgment

This research was supported, in part, by National Institutes of Health grant 5R01EB005846-02.

References (37)

  • D. Hu et al.

    Unified SPM-ICA for fMRI analysis

    NeuroImage

    (2005)
  • V.D. Calhoun et al.

    Semi-blind ICA of fMRI: a method for utilizing hypothesis-derived time courses in a spatial ICA analysis

    NeuroImage

    (2005)
  • M.J. Mckeown et al.

    Analysis of fMRI by blind separation into independent spatial components

    Hum Brain Mapp

    (1998)
  • A.J. Bell et al.

    An information maximization approach to blind separation and blind deconvolution

    Neural Comput

    (1995)
  • A. Hyvärinen et al.

    A fast fixed-point algorithm for independent component analysis

    Neural Comput

    (1997)
  • V.D. Calhoun et al.

    Independent component analysis applied to fMRI data: a generative model for validating results

    J VLSI Signal Process Syst

    (2004)
  • F. Esposito et al.

    Spatial independent component analysis of functional MRI time-series: to what extent do results depend on the algorithm used?

    Hum Brain Mapp

    (2002)
  • K.S. Petersen et al.

    On the independent components of functional neuroimages

  • Cited by (158)

    • Advances in data preprocessing for bio-medical data fusion: An overview of the methods, challenges, and prospects

      2021, Information Fusion
      Citation Excerpt :

      The number of eigenvectors is predefined by the user to meet certain requirements. PCA has been extensively used in reducing relevant features in medical data classification tasks such as schizophrenia [377, 378], AD [379], face recognition [380], and psychosis [381]. Notably, there are also regression studies that involved PCA such as age prediction [382] and AD clinical scores prediction [383].

    View all citing articles on Scopus
    View full text