Elsevier

NeuroImage

Volume 30, Issue 1, March 2006, Pages 136-143
NeuroImage

A multivariate approach for processing magnetization effects in triggered event-related functional magnetic resonance imaging time series

https://doi.org/10.1016/j.neuroimage.2005.09.012Get rights and content

Abstract

Triggered event-related functional magnetic resonance imaging requires sparse intervals of temporally resolved functional data acquisitions, whose initiation corresponds to the occurrence of an event, typically an epileptic spike in the electroencephalographic trace. However, conventional fMRI time series are greatly affected by non-steady-state magnetization effects, which obscure initial blood oxygen level-dependent (BOLD) signals. Here, conventional echo-planar imaging and a post-processing solution based on principal component analysis were employed to remove the dominant eigenimages of the time series, to filter out the global signal changes induced by magnetization decay and to recover BOLD signals starting with the first functional volume. This approach was compared with a physical solution using radiofrequency preparation, which nullifies magnetization effects. As an application of the method, the detectability of the initial transient BOLD response in the auditory cortex, which is elicited by the onset of acoustic scanner noise, was used to demonstrate that post-processing-based removal of magnetization effects allows to detect brain activity patterns identical with those obtained using the radiofrequency preparation. Using the auditory responses as an ideal experimental model of triggered brain activity, our results suggest that reducing the initial magnetization effects by removing a few principal components from fMRI data may be potentially useful in the analysis of triggered event-related echo-planar time series. The implications of this study are discussed with special caution to remaining technical limitations and the additional neurophysiological issues of the triggered acquisition.

Introduction

Event-related functional magnetic resonance imaging (er-fMRI) allows to detect and localize the blood oxygen level-dependent (BOLD) sources of rapid and transient signal changes in T2*-weighted images that are evoked by conceptually instantaneous and perceptually separated, mental or behavioral discrete events (Bandettini et al., 1992, Buckner et al., 1996). Provided a faithful temporal model of the hemodynamic impulse response (Friston et al., 1994, Boynton et al., 1996), the accuracy of the detection and the statistical power of the resulting picture of neural correlates are strongly affected by the exact knowledge of the event timing and the precision of the time-locked dynamic (echo-planar) acquisition of the evoked signals. Triggered event-related fMRI (ter-fMRI) is a variant of er-fMRI that does not assume knowledge of the events' timing but requires more advanced experimental design and setup, with sparse intervals of measurements initiated manually or automatically by the occurrence of the event of interest (Zimine et al., 2003). Probably the most natural application of ter-fMRI is related to combined electroencephalographic (EEG) and fMRI studies (Allen et al., 1998, Krakow et al., 1999). To date, EEG-fMRI is considered an approved technique, and EEG systems are available that allow for continuous EEG-fMRI acquisition and scanning. Nonetheless, using the same EEG spikes as triggering events for dynamic echo-planar acquisition and fMRI spatio-temporal pattern extraction allows to avoid the event jitter that occurs when scanning continuously, reduces bias due to systematic timing offset (especially when scanning over the whole brain) and reduces uncertainty in the peak response delay. As a consequence, this acquisition design type may crucially improve the spatial localization of the EEG event generator and may become a convenient solution for the non-invasive localization of epileptic foci in patients with epilepsy (Krakow et al., 1999).

In more general terms, ter-fMRI represents a convenient solution to all fMRI applications in which a typical event-related design is desired (e.g., a time-locked rapid and transient signal change is to be detected and characterized at a good temporal resolution), but the sequence of events cannot precisely be predicted and defined before starting the functional scan session (e.g., epileptic seizures, hallucinations), or in which discontinuous fMRI acquisition is a favorable choice (e.g., sleep studies, drug action studies). In such experimental conditions, the use of conventional (continuous echo-planar acquisition) designs can be applied but requires to oversize the imaging protocol with respect to the worst case prediction of events' occurrences before discarding all the stored data except those following the trigger event. Although there are currently no specific absorption rate (SAR) limitations for prolonged examination (the exposure limits to radiofrequency pulses apply to a temporal window of three minutes), the tissue temperature changes in the scanned subjects may become an issue (Collins et al., 2004) if high-power sequences are used much longer than typical SAR averaging periods and with mounted EEG electrodes. Moreover, although the temporal resolution of continuous EPI can be sufficiently high to ensure a dense sampling of spontaneous events, it remains the impossibility to predefine the sequence of image time points with respect to the triggering event.

When using the ter-fMRI strategy, only short image time series are rapidly acquired in a way to cover the signal effects occurring in a short time window of measurement in a maximally controlled and reliable synchronization with input event (e.g., EEG spike). However, the use of conventional unmodified echo-planar sequences for ter-fMRI bears the problem that the triggered image series are greatly affected by longitudinal magnetization non-steady-state effects. In conventional echo-planar image time series, these effects consist of strong and spatially heterogeneous exponential decays of the MR signal that vanish within intervals in the order of a few times of the tissue T1. While in typical blocked or event-related fMRI designs, the time points where this effect is visible are not analyzed, in ter-fMRI, most of the transient BOLD signal change possibly occurring after an event is obscured by this phenomenon. Moreover, considering that the tissue T1 increases with the strength of the static magnetic field, this effect becomes even more severe at higher field strengths, with the consequence of making the ter-fMRI design practically not feasible in high (3 T) and ultra-high field (4–7 T or higher) fMRI applications at a good temporal resolution. On the other hand, beyond the strong magnetization effect, ter-fMRI is suboptimal in terms of scanner performance. In fact, with the use of sporadic sampling, not only magnetization but also temperature changes in the gradient and shim systems may not reach a steady state, and this may lead to additional global signal fluctuations depending on the frequency of the measurement and the time constants affecting scanner stability.

In previous studies, a simple solution to overcome the magnetization problem based on a univariate signal subtraction has been suggested by Bandettini et al. (1998) and evaluated in the context of ter-fMRI designs at 1.5 T by Zimine et al. (2003). Using this approach, BOLD responses are recovered by subtraction between the triggered “task” series and a “control” or “baseline” series, acquired without any stimulus. However, the subtraction method has both analytical and practical limitations. Analytically, the signal decay related to magnetization saturation and any other physical (e.g., motion artifacts, temperature changes) or cognitive confounding effects are assumed to be identical in task and control image series. In general, the subtraction of the task and control image causes per se a reduction of the overall activation areas by a factor of two because of a signal-to-noise reduction (Parrish et al., 2000).

Here, we illustrate a possible alternative post-processing technique for ter-fMRI based on the use of principal component analysis (PCA) (Friston et al., 1993, Sychra et al., 1994, Andersen et al., 1999) as a multivariate filter for ter-fMRI time series (Thomas et al., 2002).

In general, there are many signal and noise sources that modulate the T2*-weighted images with various temporal profiles and spatial layouts of influence, substantially increasing the complexity of the recorded signals. Since these signal effects introduce both a spatial and a temporal correlation in the image time series, the resulting spatio-temporal datasets will possess a relevant multivariate structure with important spatial and temporal features that can only be addressed by means of multivariate statistical methods (Friston et al., 1995a, Friston et al., 1995b). In the context of ter-fMRI, univariate methods like subtraction do not exploit the relevant aspect that not each single voxel independently but that all voxels experience the strong signal decay due to the magnetization effect at a variable degree.

Here, we investigate how the dynamic effect of magnetization in ter-fMRI image time series alters the multivariate structure of the spatio-temporal datasets and explore how the eigenvalue–eigenvector (eigenmode) decomposition of the covariance matrix (eigenspectrum) is affected by the presence of this special type of noise. We produce accurate spatio-temporal patterns of BOLD activity extracted from ter-fMRI time series corrupted by the magnetization effect by selectively removing some of the dominant eigenmodes of the data (eigenfilter).

In the present study, the kernel of the data analysis is an independent component analysis (ICA, McKeown et al., 1998) with different types of eigenfilters applied to reduce the dimensionality of the training data like in Duann et al. (2002); however, both leading and trailing eigenmodes are considered in this application independently of ICA, for filtering out the unwanted signal.

It has previously been observed (see, for instance, McKeown et al., 1998) that brain activity explains only a tiny fraction of the total variance–covariance of the acquired fMRI data, with the consequence that its contribution to the eigenspectrum may be likely to be located in lower ranks. This boosts the idea that filtering out some high rank principal components, while not becoming a general practice for dimensionality reduction, may improve sometimes the ratio of variance contribution between brain activity components and other non-informative but strong signal sources, such as transient magnetization effects. Using PCA as a separate and general preprocessing step (see also Thomas et al., 2002), we report the effect of the eigenfilter also in a classical model-driven univariate linear regression analysis (Friston et al., 1995a, Friston et al., 1995b).

We evaluate this framework on the extraction and the recovery of a transient signal change from the very first volumes of high temporal resolution dynamic echo-planar sequences acquired at 3 T without a baseline measurement. Specifically, we consider the transient BOLD response in the auditory cortex elicited by the acoustic scanner noise when the sequence starts (Bandettini et al., 1998, Seifritz et al., 2002). By definition, the onset of the auditory responses elicited by the acoustic scanner noise occurs in precise time locking with the input stimulus (the scanner gradient acoustic noise itself), and the sampling of the BOLD signal is intrinsically triggered by the same event (starting of the echo-planar acquisition and read-out gradients), as would be the case of any ter-fMRI responses. Moreover, using this experimental model, there is the chance to validate the resulting spatial and temporal patterns of BOLD responses using the gold standard of an equivalent benchmark time series from the same subject, acquired after radiofrequency preparation of the imaging slab, with identical image acquisition parameters, identical baseline of silence (because the read-out phase is skipped during preparation) but without the confounding magnetization effect (because the magnetization steady-state condition is preserved).

Section snippets

Theory: multivariate analysis of fMRI time series with PCA and ICA

Multivariate techniques in fMRI statistical data analysis differ from univariate techniques in that all the voxel time courses from a given image time series are treated as a unique statistical entity. Given P voxels in the brain and T time points, a TxP “data matrix” X is filled with the available fMRI measurements. Using the matrix algebra formulation of multivariate linear decompositions, the data–matrix X is represented through a linear combination of N signal components, with N being equal

Results

Spatial ICA with standard dimension reduction in the PCA stage (i.e., with the first eigenmodes retained and only a low-pass cut-off for the eigenfilter) was able to extract one unique “auditory” spatial component from both RF-prepared and non-RF-prepared datasets in all the eight subjects. The labeling as auditory components stemmed from that the primary and secondary auditory cortices were adequately covered by the component activation map, thresholded at z = 3 and overlaid on the individual

Discussion

The inspection of the time courses from non-RF-prepared datasets in both activated and non-activated regions clearly confirmed the presence of considerable exponential signal drops in the T2*-weighted echo-planar image series during the first 5–10 s, before magnetization steady state was reached. This signal was about two orders of magnitude greater than the auditory hemodynamic response elicited by the scanner acoustic noise and precluded the easy and accurate detection and mapping of the

Conclusions

Our study demonstrates that a combined PCA–ICA analysis is able to distinguish between transient magnetization changes and a BOLD response for primary auditory stimulation. The results suggest that for special types of signals and noise and designs with irregular intervals of dynamic highly temporally resolved echo-planar acquisition, the application of a special multivariate filter to the spatio-temporal datasets consisting in the removal of the first few principal components of the time

Acknowledgment

Study was supported by the Swiss National Science Foundation grant no. PP00B-103012.

References (24)

  • R.L. Buckner et al.

    Detection of cortical activation during averaged single trials of a cognitive task using functional magnetic resonance imaging

    Proc. Natl. Acad. Sci. U. S. A.

    (1996)
  • C.M. Collins et al.

    Temperature and SAR calculations for a human head within volume and surface coils at 64 and 300 MHz

    J. Magn. Reson. Imaging

    (2004)
  • Cited by (3)

    • Independent vector analysis (IVA) for group fMRI processing of subcortical area

      2008, International Journal of Imaging Systems and Technology
    View full text