ess

un

cisz

eb

Naple

Germany

rland

erlan

05

D 2005 Elsevier Inc. All rights reserved.

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,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 radio-

frequency 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

Keywords: Functional magnetic resonance imaging (fMRI); Triggered

event-related functional magnetic resonance imaging; Multivariate filter;

Eigenfilter; Principal and independent component analysis (PCA, ICA);

Auditory cortex

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) acquisitionA multivariate approach for proc

effects in triggered event-related f

resonance imaging time series

Fabrizio Esposito,a,* Francesco Di Salle,b Fran

Marcus Herdener,d Klaus Scheffler,e Rainer Go

aDepartment of Neurological Sciences, University of Naples Federico II,

II Policlinico (Nuovo Policlinico) Padiglione 17, Via S. Pansini 5, 80131bDepartment of Neuroscience, University of Pisa, ItalycMR Methods Development Department, BRUKER BioSpin MRI, Ettlingen,dUniversity Hospital of Clinical Psychiatry, University of Bern, SwitzerlandeMR-Physics, Department of Medical Radiology, University of Basel, SwitzefDepartment of Cognitive Neuroscience, University of Maastricht, The NethgDepartment of Psychiatry, University of Basel, Switzerland

Received 12 April 2005; revised 25 August 2005; accepted 5 September 20

Available online 19 October 2005

Triggered event-related functional magnetic resonance imaging

requires sparse intervals of temporally resolved functional data

acquisitions, whose initiation corresponds to the occurrence of an1053-8119/$ - see front matter D 2005 Elsevier Inc. All rights reserved.

doi:10.1016/j.ne

* Corresponding author. Fax: +39 081 546 3663.

E-mail address: faesposi@unina.it (F. Esposito).

Available online on ScienceDirect (www.sciencedirect.com).ds

remaining technical limitations and the additional neurophysiological

issues of the triggered acquisition.ing magnetization

ctional magnetic

ek Hennel,c Ornella Santopaolo,a

el,f and Erich Seifritzd,g

s, Italy

www.elsevier.com/locate/ynimg

NeuroImage 30 (2006) 136 143with 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 isuroimage.2005.09.012

roImarelated 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 sequen-

ces 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 (47 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

F. Esposito et al. / Neufact, with the use of sporadic sampling, not only magnetization but

also temperature changes in the gradient and shim systems may notreach 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,b). 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 eigenvalueeigen-

vector (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 dimen-

sionality 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 variancecovariance 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

ge 30 (2006) 136143 137improve sometimes the ratio of variance contribution between

brain activity components and other non-informative but strong

oImaMultivariate 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

datamatrix X is represented through a linear combination of N

signal components, with N being equal or less then T, in the P-

dimensional space domain of observations. Organizing these

components as rows of an N P matrix C, we can write:

X A IC 1

In this formulation, the rows of C, Ci, are spatial processes or

maps, the columns of A, Ai, are time courses, and each pair of

corresponding vectors represents a mode of the input data.

In component-based analyses, the matrices C and A are

typically calculated from a secondary matrix W with the formulas:

C W IX; IA WT IW 1 IWT 2The matrix W is estimated with suitable formulas or algorithms

that gather the resulting matrix C with the required statistical

properties for the output components Ci.

Apart from the algebra formulation, PCA and ICA have

different mathematical and statistical properties which give them

very distinct functionalities, especially in fMRI applicationssignal 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,b).

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).

Materials and methods

Theory: multivariate analysis of fMRI time series with PCA and

ICA

F. Esposito et al. / Neur138(McKeown et al., 1998). Referring to the spatial variant that is

used in this work, while (spatial) PCA imposes that the Ci areuncorrelated or orthogonal (spatial) ICA constrains higher than

second order statistics of the Ci, attempting the ideal statistical

independence (Papoulis, 1991). In both PCA and ICA approaches,

the data in X are preliminary centered about the spatial mean.

In PCA, the standard solution for W in pls. link: (2) is found in

terms of the eigenvalues and eigenvectors (or eigenimages) of the

spatial covariance of the data:

Rx 1=P 4X IXt 3If E is the matrix whose columns are the unit-norm eigenvector

of Rx and D is the diagonal matrix of eigenvalues, the matrix W

and the components for PCA can be estimated in closed form by:

W PCA D1=2 IEt; IC PCA W PCA IX 4In PCA, the components Ci are naturally ordered in terms of the

amount of explained variancecovariance of the data. Plotting

the eigenvalues, typically in a logarithmic scale, gives the so-called

eigenspectrum.

The most popular ICA algorithms (see Esposito et al., 2002)

estimate W using iterative rather than closed form methods based

on information theory principals and measures (Hyvarinen et al.,

2001). Since independent components are anyway orthogonal, the

PCA expansion is sometimes used as a first decomposition stage

before ICA (also called whitening). In the context of whitening, the

ICA solution can be seen as a further rotation to the principal

components that pursues the more stringent constraint of mutual

statistical independence in change of the simpler variance

covariance constraint:

C ICA W ICA IC PCA W ICA W PCA IX 5

As a consequence, unlike PCA, ICA components will not be

intrinsically ordered, and the amounts of variancecovariance in

the data explained by each component will not constitute a useful

spectrum.

In this paper, we focus on the aspect that, besides doing the first

part of ICA statistical job, PCA is also a natural step in which a

dimension reduction of the data is performed if not all the rows of

W(PCA) enter the ICA separation in pls. link: (5). This has the effect

of (i) defining the final number of components in ICA and (ii)

providing a multivariate orthogonal filtering (eigenfilter) of the

data when back-reconstructing the dataset from pls. link: (1) like in

Thomas et al. (2002).

Experiments: subjects and image acquisition

The experiment was part of a larger experiment with details

reported elsewhere (Seifritz et al., 2002). Eight healthy subjects (N =

4 females/4 males; mean age and standard deviation, 36.4 T 8.9years) were examined and instructed to lay quietly supine in the

scanner and not to perform any output task. Images were collected

on a 3-T Medspec Avance whole body system (Bruker, BioSpin

MRI, Germany) equipped with a BGA38 head gradient system

(gradient strength, 28 mT/m; slew rate, 280 T/m/s on all three axes)

and a circularly polarized head coil. After anatomical imaging,

functional volumes consisting of ten gradient recalled echo-planar

images (slice thickness, 4 mm; matrix, 64 64 pixels; field of view,240 240 mm2; flip angle, 60-; echo time, 30 ms; slice acquisitiontime, 100 ms, repetition time, 1000 ms) positioned along the lateral

ge 30 (2006) 136143sulcus to cover the superior temporal gyrus including primary and

secondary auditory cortices were acquired. After collecting a first set

of 60 conventional echo-planar volumes, 60 silent dummy

repetitions consisting of radiofrequency and slice selection gradient

pulses with long sinusoidal ramps were carried out to obtain

magnetization steady state for the second set of 60 echo-planar

volumes. This way, the functional session lasted 3 min and consisted

of three distinct 1-min-lasting intervals: 60 scans of echo-planar

acquisition (repetition time TR, 1 s), 60 scans of only radiofrequency

(RF) excitation (TR, 1 s) without readout gradients activated (thus

silent), 60 scans of echo-planar acquisition (TR, 1 s). The approach

is illustrated in Fig. 1 and allowed to acquire two distinct datasets per

subject with the same sequence parameters but completely different

contribution of the magnetization effect. This effect is expected to be

maximally present in the first series (non-RF prepared) and totally

absent in the second series (RF prepared). Since the RF preparation

of the second echo-planar imaging sequence was completely silent,

we expected to achieve exactly the same auditory stimulation by the

acoustic scanner noise starting on the first volume after a baseline

period of silence in both series. Each trial was repeated five times in

each subject during one experimental session and resulting non-RF-

prepared and RF-prepared time series were averaged before

submission to data analysis.

Data analysis

After coregistration with anatomical scans, the functional

images were warped into standard Talairach space, resampled into

3-mm isotropic voxels, rescaled in intensity to account for receiver

sensitivity run-to-run differences, corrected for slice acquisition

time, head motion and intra-session between run differences using

BrainVoyager 2000 (Brain Innovation B.V., www.brainvoyager.

com). The resulting single-trial voxel time series (60 scans) were

separately averaged for the non-RF-prepared case and the RF-

prepared case and filtered in space using a 4-mm full-width at half-

maximum Gaussian kernel. A common anatomical mask generated

Fig. 1. Simple scheme of the echo-planar imaging sequence used for the acquisition of non-RF-prepared and RF-prepared fMRI time series.

F. Esposito et al. / NeuroImage 30 (2006) 136143 139Fig. 2. Auditory ICA component maps thresholded at z = 3 (left) and region of act

prepared data (lower panel). Note how the spatial maps are highly overlapping whi

prepared echo-planar time series.ivity time courses (right) from RF-prepared data (upper panel) and non-RF-

le the magnetization-related signal obscures the BOLD in signal in non-RF-

by a simple intensity threshold on the anatomical scans was applied

to the non-RF-prepared and the RF-prepared time series from each

subject, and the voxels outside the brain were excluded. The

remaining voxels were used to fill the data matrices X in pls. link:

(1) for non-RF-prepared and RF-prepared datasets. PCA and ICA

as well as linear correlation analyses were performed on these

datasets within the same component analysis framework.

In PCA, the eigenspectrum of each dataset was estimated and

evaluated using Matlab routines from the fastica package (http://

www.cis.hut.fi) producing plots in a logarithmic scale of the

eigenspectra of the data. Multiple eigenfilters were defined in

which 30 eigenmodes were retained but sequentially 0, 1, 2,. . .12,. . . 15 of the first (dominant) eigenmodes were filtered out fromthe PCA-transformed (whitened) datasets before ICA transforma-

tion or before data back-reconstruction. For ICA, the matlab

program runica (http://www.cnl.salk.edu) which implements the

infomax algorithm (Bell and Sejnowski, 1995) was used, and the

resulting component values across all the voxels (rows of the

matrix C) were scaled to z scores (i.e., the number of standard

deviations from the map mean) as in (McKeown et al., 1998)

before being overlaid as colour coded activation map on the

anatomical scans with a threshold of z = 3. Instead, the

reconstructed (eigenfiltered) data were applied a simple univariate

linear correlation analysis with a double gamma reference function

modeling the transient auditory response to the onset of the scanner

induced auditory stimulus.

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 secon-

dary auditory cortices were adequately covered by the component

activation map, thresholded at z = 3 and overlaid on the individual

anatomy (see Fig. 2 for one representative subject). However, the

average region-of-activity time courses extracted from the two

r the a

ed tim

empo

sting

F. Esposito et al. / NeuroImage 30 (2006) 136143140Fig. 3. (a) ICA component maps and region of activity signals estimated afte

various cut-off (1, 2, 3, 10, 12 in the yellow box) compared to the RF-prepar

off was on the 1st, 2nd, 3rd, 10th and 12th eigenmode. (b) Spatial (left) and t

filtered non-RF-prepared data and the spatial and temporal templates consiactivity of the auditory components from non-RF-prepared data compared to that

an average time course of all the eight case subjects (right) after the applicationpplication of the high-pass eigenfilter to the non-RF-prepared time series at

e series (red box). For the non-RF-prepared data, from left to right, the cut-

ral (right) linear correlation analysis between the auditory components of the

in the auditory component of RF-prepared time series. (c) Time course offrom the RF-prepared data for one single representative subject (left) and as

of the eigenfilter with cut-offs on the 1, 2, 3 and 4 dominant eigenmodes.

pass cut-off for the eigenfilter, ranging from 1 to 12, before spatial

ICA decomposition (Fig. 3a). Despite the strong reduction in the

total variance of the data, we found that the spatial layout of the

auditory components did not vary significantly from the default case

of a full eigenspectrum preserved to the case of few eigenmodes

removed (at least up to four). This was evident at a visual inspection

of the maps and confirmed numerically by the analysis of the linear

correlation coefficient between the auditory components of the

filtered non-RF-prepared data and the corresponding spatial

template. Remarkably, the time course of activity of these

components was significantly improved by the eigenfilter: filtering

out dominant eigenmodes produced a progressive reduction of the

magnetization effect in the resulting time courses and a clear

enhancement of the functional contrast of the BOLD response. The

linear correlation analysis with the templates (Fig. 3b) and the visual

inspection of the time courses (Fig. 3c) revealed in single- and

Fig. 4. Normalized (and equalized on the first dominant eigenmode)

logarithmic eigenspectra of RF-prepared and non-RF-prepared datasets in

one subject. Eigenmodes (2) and (3) become significantly higher for non-

RF-prepared data compared to RF-prepared datasets suggesting the

F. Esposito et al. / NeuroImage 30 (2006) 136143 141datasets using the same components as functional mask exhibited a

clearly different behavior and a strongly different scale of the

signal change within the first 30 s of measurements (Fig. 2).

Specifically, the presence of the magnetization steady-state effect-

related signal clearly obscured in the non-RF-prepared time series

the transient BOLD response evoked by the acoustic stimulus and

clearly distinguishable in the region of activity signal from the RF-

prepared data. Moreover, while the selection of the auditory

component was possible in both datasets using a rough spatial

template (region of interest-based anatomical selection, Van de Ven

et al., 2004), a straightforward selection based on a temporal

template (double gamma response synchronized on the onset of the

acquisition) was possible only for RF-prepared data. In fact, the

time course of the auditory component of RF-prepared data

exhibited a high maximal linear correlation (r = 0.83) with the

reference function, but this was not the case of the component

derived from non-RF-prepared data (r = 0.29).

The spatial ICA decomposition of non-RF-prepared data was,

then, repeated after a different dimension reduction: the first

principal components of the analyzed time series were, now,

removed, and the maximum number of independent components

presence of signals strongly affecting the eigenstructure of the non-RF-

prepared datasets.was estimated. Using the spatial and temporal templates provided by

the spatial ICA decompositions of RF-prepared data, we compared

the auditory components of non-RF-prepared data at various high-

Fig. 5. (a) Comparison of the activation maps obtained from non-RF-prepared tim

gamma reference function before (left) and after (right) the eigenfilter on the fir

auditory spatio-temporal template provided by ICA.averaged multisubject data an optimal cut-off at 3 and 4 dominant

eigenmodes. In our sample of subjects, a cut-off on the 4th

eigenmode always resulted in a favorable trade-off between

preserved variance in the data and functional BOLD contrast of

the auditory response.

The effect of the eigenfilter on the non-RF-prepared time series is

also graphically illustrated by the comparison of the eigenspectra of

the two datasets (Fig. 4). Assuming a dominant eigenmode

equalized between the two datasets, the logarithmic plot clearly

shows how the relative contribution to the total variance of at least

the 2nd and 3rd dominant eigenmodes are significantly higher for

non-RF-prepared data compared to RF-prepared datasets. This

outcome turns out to be associated to the presence of signals strongly

affecting the eigenstructure of the non-RF-prepared datasets.

The removal of the dominant eigenmodes had a beneficial

effect on BOLD contrast sensitivity even in the classical frame-

work of univariate inferential analysis. Fig. 5a compares the

activation maps obtained from non-RF-prepared time series after

simple linear regression analysis of the data on a double gamma

reference function before and after the eigenfilter of the first three

eigenimages. The spatial sensitivity of the analysis is greatly

improved by the multivariate filter with obvious advantages in

terms of signal detection capabilities of the method for the event-

related BOLD response. Of course, the read-out produced by linear

regression analysis is affected by the choice of the reference

function that partly explains the differences between the correlation

and the ICA maps.

e series after simple linear regression analysis of the time series on a doublest three eigenimages. (b) PCA auditory component map selected with the

may cause a number of additional spots of activity and signal

fluctuations that will be superimposed to the effects of interest.

without a baseline measurement using post-processing only and

suggested the opportunity of the multivariate component-based

oImaDiscussion

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 510

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

intrinsically triggered brain activity.

The use of a special eigenfilter removing the first principal

components of the non-RF-prepared datasets allowed a very

convenient multivariate filtering, with a significant recovery of

the expected shape of the auditory fMRI responses and a

minimal loss of accuracy in the spatial layout of the component

maps. The filtering of the dominant eigenmodes enabled the

selection of the auditory ICA component with a simple

correlation of the ICA time courses and a temporal template

and restored a reasonable level of accuracy of conventional

univariate linear regression analysis of the data in detecting the

triggered event-related hemodynamic response from most of

activated voxels.

This finding demonstrates how filtering the eigenimages of

fMRI time series before ICA can be useful not only for the non-

dominant part of the eigenspectrum (low-pass eigenfilter) as is

often done for model order selection and noise reduction but

also for the dominant part of the eigenspectrum (high-pass

eigenfilter) when strong signals introduce clutter in the data in

way that the multivariate statistical structure (eigenstructure) is

affected. In our study, the clutter is represented by the

magnetization effect, the prominent confounding and limiting

factor of triggered event-related fMRI designs at irregular

intervals of acquisition.

The use of component-based analysis as a noise reduction stage

to improve BOLD contrast sensitivity was pointed out by Thomas

et al. (2002) as an alternative approach to pattern extraction and

signal detection. The difference between the two perspectives is

even more emphasized in the presented application where PCA is

essentially used for a highly tailored global effect reduction filter

before and independently of the signal detection analysis stage. In

our case, PCA per se was not adequate for accurate signal

detection, as could be confirmed by the visual inspection of the

auditory principal component map (i.e., the PCA map exhibiting

the highest spatial correlation with the auditory template, reported

in Fig. 5b). As for the clutter magnetization-related signals, the

variancecovariance of BOLD signals is somehow spread in a

finite interval of dominant eigenmodes and is not summarized by

one principal component. The positive observation here was that

the BOLD signal contribution turned out to be much more spread

in the eigenspectrum than the clutter signal, and we could improve

BOLD contrast sensitivity using a high-pass cut-off for the

eigenfilter.

Our demonstration considered the detection of the auditory

fMRI responses to the acoustic noise produced by the scanner

gradients as an ideal experimental model for ter-fMRI. Never-

theless, our results do not necessarily demonstrate that ter-fMRI

and multivariate filtering can be used, for instance, in the detection

of epileptic foci: our model only considered the case of a primary

F. Esposito et al. / Neur142sensory stimulation, thus we can only speculate that the same holds

true for the much more subtle and heterogeneous response elicitedfiltering of the image time series in high field and high temporally

resolved ter-fMRI.

RF-preparation produced the maximally accurate character-

ization of the auditory fMRI responses without any special filtering

and, thus, remains the best strategy to avoid magnetization effects,

even if it has not been used yet in a real ter-fMRI setup with

triggering devices.

Conclusions

Our study demonstrates that a combined PCAICA 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 series (high-pass eigenfilter)

may improve the detection capabilities and the functional contrast

sensitivity of the analysis with respect to the presence of global

confounding signal fluctuations like the ones produced by the

magnetization vector in non-steady state conditions. The proposed

approach may turn out to be useful in fMRI applications where a

sporadic echo-planar sampling of a brain volume, triggered by

spontaneous episodic events like, for instance, EEG-detected

epileptic spikes, is considered a feasible solution, despite the

known suboptimal performances of the scanner hardware and the

expected additional neurophysiological effects.

AcknowledgmentApart from the primary auditory stimulation investigated in this

work, higher order cognitive responses may be present (see, for

instance, in Fig. 3a, a spot in the frontal brain that becomes visible

when the auditory responses are severely attenuated by the cut-off

of twelve dominant eigenmodes). Triggering the experiment with

some internal signal may even lead to biofeedback phenomena

obscuring the effects of interest in the case of noisy responses. In

addition, the sudden stimulus-related onset of scanner noise may

produce stimulus-correlated motion effects and more subtle

neuronal effects by startling and arousal of the subjects. In general,

despite the separation power, the crude application of the illustrated

approach does not guarantee that any Factivation_ found representsa functional BOLD-effect or, instead, some other residual transient

signal.

Given the average values of the T1 relaxation tissue parameter

at 3 T, our experiments ensured a transient and time-locked signal

change in the first 510 volumes of an echo-planar time series

with high temporal resolution (Seifritz et al., 2002), without

engaging the more complex setup of real ter-fMRI studies. We

showed how these signals can be made free of this effect at 3 Tby some Finternal_ events like epileptic spikes. In other experi-ments, whatever activation is to be measured, sporadic sampling

ge 30 (2006) 136143Study was supported by the Swiss National Science Foundation

grant no. PP00B-103012.

References

Allen, P.J., Polizzi, G., Krakow, K., Fish, D.R., Lemieux, L., 1998.

Identification of EEG events in the MR scanner: the problem of pulse

artifact and a method for its subtraction. NeuroImage 8, 229239.

Andersen, A.H., Gash, D.M., Avison, M.J., 1999. Principal component

analysis of the dynamic response measured by fMRI: a generalized

linear systems framework. Magn. Reson. Imaging 17, 795815.

Bandettini, P.A., Wong, E.C., Hinks, R.S., Tikofsky, R.S., Hyde, J.S., 1992.

Time course EPI of human brain function during task activation. Magn.

Reson. Med. 25, 390397.

Bandettini, P.A., Jesmanowicz, A., Van Kylen, J., Birn, R.M., Hyde, J.S.,

1998. Functional MRI of brain activation induced by scanner acoustic

noise. Magn. Reson. Med. 39, 410416.

Bell, A.J., Sejnowski, T.J., 1995. An information-maximisation approach

to blind separation and blind deconvolution. Neural Comput. 7,

10041034.

Boynton, G.M., Engel, S.A., Glover, G.H., Heeger, D.J., 1996. Linear

systems analysis of functional magnetic resonance imaging in human

V1. J. Neurosci. 16, 42074221.

Buckner, R.L., Bandettini, P.A., OCraven, K.M., Savoy, R.L., Petersen,

S.E., Raichle, M.E., Rosen, B.R., 1996. Detection of cortical activation

during averaged single trials of a cognitive task using functional

magnetic resonance imaging. Proc. Natl. Acad. Sci. U. S. A. 93,

1487814883.

Friston, K.J., Jezzard, P., Turner, R., 1994. Analysis of functional MRI time

series. Hum. Brain Mapp. 1, 153171.

Friston, K.J., Frith, C.D., Frackowiak, R.S., Turner, R., 1995a. Character-

izing dynamic brain responses with fMRI: a multivariate approach.

NeuroImage 2, 166172.

Friston, K.J., Holmes, A.P., Worsley, K.J., Poline, J.P., Frith, C.D.,

Frackowiak, R.S.J., 1995b. Statistical parametric maps in functional

imaging: a general linear approach. Hum. Brain Mapp. 2, 189210.

Hyvarinen, A., Karhunen, J., Oja, E., 2001. Independent Component

Analysis. Wiley.

Krakow, K., Woermann, F.G., Symms, M.R., Allen, P.J., Lemieux, L.,

Barker, G.J., Duncan, J.S., Fish, D.R., 1999. EEG-triggered functional

MRI of interictal epileptiform activity in patients with partial seizures.

Brain 122, 16791688.

McKeown, M.J., Makeig, S., Brown, G.G., Jung, T.-P., Kindermann, S.S.,

Bell, A.J., Sejnowski, T.J., 1998. Analysis of fMRI data by blind

separation into spatial independent component analysis. Hum. Brain

Mapp. 6, 160188.

Papoulis, A., 1991. Probability, Random Variables, and Stochastic

Processes. McGraw-Hill, New York.

Parrish, T.B., Gitelman, D.R., LaBar, K.S., Mesulam, M.M., 2000.

Impact of signal-to-noise on functional MRI. Magn. Reson. Med. 44

(6), 925932.

Seifritz, E., Esposito, F., Hennel, F., Mustovic, H., Neuhoff, J.G., Bilecen,

D., Tedeschi, G., Scheffler, K., Di Salle, F., 2002. Spatiotemporal

pattern of neural processing in the human auditory cortex. Science 297,

F. Esposito et al. / NeuroImage 30 (2006) 136143 143Smith, M.B., 2004. Temperature and SAR calculations for a human

head within volume and surface coils at 64 and 300 MHz. J. Magn.

Reson. Imaging 19, 650656.

Duann, J.R., Jung, T.P., Kuo, W.J., Yeh, T.C., Makeig, S., Hsieh, J.C.,

Sejnowski, T.J., 2002. Single-trial variability in event-related BOLD

signals. NeuroImage 15, 823835.

Esposito, F., Formisano, E., Seifritz, E., Goebel, R., Morrone, R., Tedeschi,

G., Di Salle, F., 2002. Spatial independent component analysis of

functional MRI time-series: to what extent do results depend on the

algorithm used? Hum. Brain Mapp. 16, 146157.

Friston, K.J., Frith, C., Liddle, P., Frackowiak, R.S.J., 1993. Func-

tional connectivity: the principal component analysis of large data

sets. J. Cereb. Blood Flow Metab. 13, 514.17061708.

Sychra, J.J., Bandettini, P.A., Bhattacharya, N., Lin, Q., 1994. Synthetic

images by subspace transforms: I. Principal components images and

related filters. Med. Phys. 21, 193201.

Thomas, C.G., Harshman, R.A., Menon, R.S., 2002. Noise reduction in

BOLD-based fMRI using component analysis. NeuroImage 17,

15211537.

Van de Ven, V.G., Formisano, E., Prvulovic, D., Roeder, C.H., Linden,

D.E., 2004. Functional connectivity as revealed by spatial independent

component analysis of fMRI measurements during rest. Hum. Brain

Mapp. 22, 165178.

Zimine, I., Seghier, M.L., Seeck, M., Lazeyras, F., 2003. Brain activation

using triggered event-related fMRI. NeuroImage 18, 410415.Collins, C.M., Liu, W., Wang, J., Gruetter, R., Vaughan, J.T., Ugurbil, K.,

A multivariate approach for processing magnetization effects in triggered event-related functional magnetic resonance imaging time seriesIntroductionMaterials and methodsTheory: multivariate analysis of fMRI time series with PCA and ICAExperiments: subjects and image acquisitionData analysis

ResultsDiscussionConclusionsAcknowledgmentReferences