Elsevier

NeuroImage

Volume 19, Issue 2, June 2003, Pages 253-260
NeuroImage

Regular article
Independent component analysis of nondeterministic fMRI signal sources

https://doi.org/10.1016/S1053-8119(03)00097-1Get rights and content

Abstract

Neuronal activation can be separated from other signal sources of functional magnetic resonance imaging (fMRI) data by using independent component analysis (ICA). Without deliberate neuronal activity of the brain cortex, the fMRI signal is a stochastic sum of various physiological and artifact related signal sources. The ability of spatial-domain ICA to separate spontaneous physiological signal sources was evaluated in 15 anesthetized children known to present prominent vasomotor fluctuations in the functional cortices. ICA separated multiple clustered signal sources in the primary sensory areas in all of the subjects. The spatial distribution and frequency spectra of the signal sources correspond to the known properties of 0.03-Hz very-low-frequency vasomotor waves in fMRI data. In addition, ICA was able to separate major artery and sagittal sinus related signal sources in each subject. The characteristics of the blood vessel related signal sources were different from the parenchyma sources. ICA analysis of fMRI can be used for both assessing the statistical independence of brain signals and segmenting nondeterministic signal sources for further analysis.

Introduction

Controlled neuronal activation of the brain functional cortex produces signal enhancement that can be detected with functional magnetic resonance imaging (fMRI) Ogawa et al 1990, Ogawa et al 1998. Without deliberate neuronal activity the T2*-weighted fMRI time domain signal is influenced by uncontrolled, non-Gaussian signal sources including spontaneous neuronal activity, vasomotor fluctuations, and thermal, instrumentational, and other noise sources Weisskoff et al 1996, Mitra et al 1997, Frank et al 2001. In the past, most of the low-frequency signal sources have been regarded as nuisances that obscure fMRI results and focus has been in attenuating their effects. Recently, however, spontaneous activity of connected neural networks has been related to fMRI signal fluctuations at frequencies lower than 0.1 Hz Biswal et al 1995, Li et al 1999, Cordes et al 2000. In anesthesia, the fMRI signal of primary functional cortices is dominated by 0.03-Hz signal intensity fluctuation that closely resembles spontaneous vasomotor (or Mayer) waves Kiviniemi et al 2000, Kleinfeld et al 1998, Obrig et al 2000.

Analysis methods suited for controlled activity, i.e., methods that can utilize a priori assumptions about the spatial distributions and temporal waveforms of the fMRI signal sources, may not be completely accurate in characterizing the uncontrolled physiological signal sources. More robust analysis of nondeterministic or stochastic processes, like vasomotor waves, is based on the statistical analysis of the measured signal (Oppenheim et al., 1996). Independent component analysis (ICA) has recently been shown to be able to separate activation, physiological, and other signal sources in fMRI studies McKeown et al 1998, McKeown and Sejnowski 1998. ICA separates various sources of the fMRI signal by maximizing both the statistical independence and the non-Gaussianity of the source signals McKeown et al 1998, McKeown and Sejnowski 1998, Hyvärinen 1999, Hyvärinen and Oja 2000. The capability of ICA to separate signal sources based on their non-Gaussian distributions could thus be used in differentiating nondeterministic physiological signal sources from fMRI data Kiviniemi et al 2000, Eke and Hermán 1999, West et al 1999, Obrig et al 2000, McKeown et al 1998, McKeown and Sejnowski 1998.

In order to assess the analysis methods of spontaneous physiological fMRI signal sources, one should have strong nondeterministic signal sources and reduced artifacts in the data. Near infrared spectroscopy has shown that spontaneous background vasomotor fluctuations induce greater blood flow changes than neuronal activation during anesthesia (Kleinfeld et al., 1998). Under anesthesia, the very-low-frequency fluctuation (VLF) of 0.03 Hz actually dominates the fMRI signal in the primary functional regions (Kiviniemi et al., 2000). Anesthetized children usually have practically no motion artifacts. The circulatory system of child subjects is also stable compared to young adults based on the lower spectral power of heart rate variability (Pikkujämsä et al., 1999). Thus, the anesthetized child brain seems like a stable model for analyzing the capability of ICA to separate nondeterministic physiological signal sources from brain fMRI data.

The aim in this study was to see whether ICA could separate the physiological signal sources of the brain. The hypothesis was that the statistically independent source signals are present in primary sensory cortices and that the sources are dominated by very-low-frequency fluctuation in the parenchyma. Also other physiological source signals, including CSF and blood flow pulsation, should be detectable. The spatial, frequency, and connectivity characteristics of statistically independent blood oxygen level dependent (BOLD) signal sources of resting state fMRI data during anesthesia were analyzed.

Section snippets

Materials and methods

Fifteen child subjects (7 females, 8 males, age range 2–9.5 years, mean 5.2) were imaged under thiopental anesthesia with an fMRI BOLD sequence after clinical brain imaging. The subject were recruited randomly in order of admittance. Midatzolam premedication (0.3 mg/kg) was given 2–3 h before the iv anesthesia with thiopental (average dose 6 mg/kg/h). The attending anesthesiologist sedated the subject into a state without voluntary motion while spontaneous breathing was ensured (Kiviniemi et

Spatial distribution

ICA separated large and clustered individual signal sources in the visual cortex in each subject, as shown in Fig. 1. ICA was able to differentiate at least two sources in the visual areas in each case, three sources in 10 cases and four in 3 cases. Multiple signal sources of one example subject are shown in the upper row of Fig. 2. One case showed a source which followed the anatomy of the posterior cerebral artery as the most prominent in the visual cortex. In the rest of the subjects the

Discussion

The ICA was able to separate spatially independent signal sources related to nondeterministic physiological fluctuations in the anesthetized brain. On average, nine clustered signal sources could be identified in each subject at the functional cortices that are known to present dominant vasomotor fluctuation under an anesthetized condition Kiviniemi et al 2000, Kleinfeld et al 1998. The signal sources closely resemble fMRI activation maps of primary and associative sensory and motor areas. ICA

Conclusion

ICA can separate multiple, statistically independent, physiological signal sources in the primary sensory areas and major blood vessels in anesthetized brain fMRI. The spatial and frequency characteristics of the functional signal sources match with the previous knowledge of vasomotor waves in the functional parenchyma during anesthesia. ICA exceeds the capability of previously used frequency and time domain methods in physiological source signal localization. The ICA is a potential tool for

References (29)

  • A. Hyvärinen et al.

    Topographic independent component analysis

    Neural Comput.

    (2001)
  • V. Kiviniemi et al.

    Slow vasomotor fluctuation in the fMRI of the anesthetized child brain

    Magn. Reson. Med.

    (2000)
  • V. Kiviniemi et al.

    The effect of midazolam sedation on baseline fMRI signal

    MAGMA

    (2002)
  • D. Kleinfeld et al.

    Fluctuations and stim+ulus-induced changes in blood flow observed in individual capillaries in layers 2 through 4 of rat neocortex

    Proc. Natl. Acad. Sci. USA

    (1998)
  • Cited by (322)

    • Physiological brain pulsations

      2023, Advances in Resting-State Functional MRI: Methods, Interpretation, and Applications
    View all citing articles on Scopus
    View full text