Available online at www.sciencedirect.com

ng 26 (2008) 905–913

Magnetic Resonance Imagi

Independent component model of the default-mode brain function:combining individual-level and population-level analyses

in resting-state fMRIFabrizio Espositoa,b,⁎, Adriana Aragric, Ilaria Pesaresid, Sossio Cirilloc,e,

Gioacchino Tedeschic,e, Elio Marcianoa, Rainer Goebelb, Francesco Di Salleb,daDepartment of Neuroscience, University of Naples “Federico II,” 80131 Naples, Italy

bDepartment of Cognitive Neuroscience, Maastricht University, 6200 MD Maastricht, The NetherlandscDepartment of Neurological Sciences, Second University of Naples, 80138 Naples, Italy

dDepartment of Neurosciences, University of Pisa, 56127 Pisa, ItalyeNeurological Institute for Diagnosis and Care “Hermitage Capodimonte,” 80131 Naples, Italy

Received 31 October 2007; revised 20 December 2007; accepted 14 January 2008

Abstract

Resting-state functional magnetic resonance imaging (RS-fMRI) is a technique used to investigate the spontaneous correlations of blood-oxygen-level-dependent signals across different regions of the brain. Using functional connectivity tools, it is possible to investigate aspecific RS-fMRI network, referred to as “default-mode” (DM) network, that involves cortical regions deactivated in fMRI experiments withcognitive tasks. Previous works have reported a significant effect of aging on DM regions activity.Independent component analysis (ICA) is often used for generating spatially distributed DM functional connectivity patterns from RS-fMRIdata without the need for a reference region. This aspect and the relatively easy setup of an RS-fMRI experiment even in clinical trials haveboosted the combined use of RS-fMRI and ICA-based DM analysis for noninvasive research of brain disorders.In this work, we considered different strategies for combining ICA results from individual-level and population-level analyses and used themto evaluate and predict the effect of aging on the DM component.Using RS-fMRI data from 20 normal subjects and a previously developed group-level ICA methodology, we generated group DM maps andshowed that the overall ICA–DM connectivity is negatively correlated with age. A negative correlation of the ICA voxel weights with ageexisted in all DM regions at a variable degree.As an alternative approach, we generated a distributed DM spatial template and evaluated the correlation of each individual DM componentfit to this template with age. Using a “leave-one-out” procedure, we discuss the importance of removing the bias from the DM template-generation process.© 2008 Elsevier Inc. All rights reserved.

Keywords: Functional magnetic resonance imaging; fMRI; Default-mode network; Resting state; Independent component analysis; Group-level analysis

1. Introduction

Resting-state functional magnetic resonance imaging(RS-fMRI [1]) is a noninvasive technique used to investigate

⁎ Corresponding author. Department of Neuroscience, University ofNaples Federico II, Unita' di Audiologia e Vestibologia, II Policlinico(Nuovo Policlinico) Padiglione 13, 80131 Naples, Italy. Tel.: +39 81 7462883; fax: +39 81 746 3581.

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

0730-725X/$ – see front matter © 2008 Elsevier Inc. All rights reserved.doi:10.1016/j.mri.2008.01.045

and characterize in vivo the spontaneous correlations ofblood-oxygen-level-dependent (BOLD) signals within andbetween different regions of the human brain. The ultimateinterpretation of these effects pertains to the amount ofsynchronized neural activity existing between distinct brainlocations in a given period of time in the absence of input(sensory) or output (motor) tasks.

A specific RS-fMRI network, referred to as “default-mode” (DM) network [2–4], has recently attracted consider-able interest in the clinical neuroscience community for its

906 F. Esposito et al. / Magnetic Resonance Imaging 26 (2008) 905–913

possible interpretation as the baseline cognitive state of asubject and its link to memory and executive function innormal and pathological conditions. In fact, the DM networkinvolves the anterior cingulate cortex (ACC) and theposterior cingulate cortex (PCC), which are known to beinvolved in attention-related processes [5], and a number ofother regions that are often detectable as transiently orconsistently deactivated during many different types ofcognitive tasks [6]. For this reason, Raichle et al. [3], whofirst targeted this type of brain activity with positronemission tomography imaging, introduced the concept of“DM” activity and attempted to differentiate a “cognitive”baseline state from a “general” resting state in the humanbrain. After the study by Greicius et al. [4], the DM networkhad often been conceptualized as a “stand-alone” function orsystem to be analyzed with data models specifically orientedto functional connectivity [7].

Independent component analysis (ICA [8]) is a multi-variate statistical technique that has been successfullyapplied to fMRI voxel time series for generating spatiallydistributed patterns of functional connectivity in a fully data-driven fashion (spatial ICA [9]) from BOLD signals. Byvirtue of this property, ICA performs similarly in fMRIstudies based on a variety of experimental paradigms andsettings, as well as in RS-fMRI. When applied to strictlytimed fMRI paradigms, ICA has been shown to characterizenot only consistently task-related BOLD activities but alsotransiently task-related activities [9,10]. This has favored theapplication of ICA to more loosely timed paradigms [11,12]and to complex naturalistic stimuli [12,13].

To date, ICA appears to be the method of choice foranalyzing RS-fMRI time series because no formal approachexists for predicting the expected timing and amplitude ofsignal evolution [14,15]. Compared to other functionalconnectivity analysis tools, ICA improves and generalizesprincipal component analysis (PCA [7,9]) in achieving afully distributed representation of the fMRI voxelwisecorrelations from whole-brain data sets without the need toprespecify voxels or regions of interest (ROI) as seeds orbuilding blocks for the network analysis. In this sense, theapplication of spatial ICA to RS-fMRI data decompositionfor the analysis of spontaneous correlations of BOLD signalsmight be paralleled to the use of Fourier decomposition forelectromagnetic data in the spectral analysis of spontaneousbrain rhythms. A comparative analysis of the outcomes fromthe two methodologies has been recently presented in asimultaneous electroencephalogram–fMRI study, where theneurophysiological connection between fMRI–ICA compo-nents and electroencephalogram spectral components wasexplored in a resting-state experiment [16].

In almost all fMRI studies where ICA was adopted, theresults have highlighted its capability to extract a reliableestimate of the DM network in the form of one spatialpattern, which was stable and consistent at the individual-subject level and generalized well at the group level [14,15].When applied to fMRI data from working-memory experi-

ments with subjects alternating between cognitive tasks ofdifferent difficulties (n-back [17]), the ICA–DM estimationhas proven to be, at the same time, stable across subjects andlevels in its network layout, but also modulated at the level ofisolated regions of activity, thus showing the potential ofreflecting and discriminating between different states of thecognitive engagement of the subjects [17].

Altogether, the relatively easy setup of an RS-fMRIexperiment, the seed-independent general modeling of thespontaneous functional connectivity provided by ICA andthe bridge from resting states to cognitively active brainstates mediated by the DM function have boosted theneuroimaging research of brain disorders according to thesemethodology and neurophysiology. The use of simpleresting-state designs was promising and attractive particu-larly for studying cognitively impaired patient populations,given the inherent problems of matching the scoresbetween patients and normal controls if a cognitive taskis chosen [18]. Greicius et al. [18] have first evaluated theICA-based DM network as a potential neurological markerin the evaluation of mild cognitive impairment and earlyAlzheimer's disease that produces significant evidence forthis potential.

Similarly to Greicius et al. [18], all DM componentsextracted from individual RS-fMRI data sets can besubmitted to any type of random-effects analysis, with oneor more categorical factors (grouping) or covariates used forpopulation-level hypothesis testing. Besides that, a veryattractive framework for the DM analysis in clinical trials isrepresented by the use of a single individual score, such as agoodness-of-fit measure with respect to a given template[18], characterizing globally the ICA-based DM connectivityof a single subject. Ideally, these scores should be suitable foraiding or automating the classification of the individual DMcomponent in a diagnostic framework and possibly allowextension of the study to large samples of subjects.

There remains the problem, however, of how to preparean optimal template of the DM component for objectivecomponent selection and scoring of the individual subjectICA decomposition. A possible alternative to the preparationof a manual template [18] is to rely on a group-level ICAanalysis methodology for generating a distributed andthreshold-independent template.

Different approaches have been so far proposed forgroup-level ICA [19–22]. It is possible to combine the datasets beforehand and perform a joint dimensionality reductiontowards a computationally treatable and unique ICAdecomposition [19–21], or, using individual ICA decom-positions, submit components to a clustering algorithm forextracting maximally similar components across individualsand summary results [22].

In this study, we consider different strategies of combiningICA results from individual-level and population-levelanalyses in an RS-fMRI study on adult subjects across awide range of age. Previous works have reported a significanteffect of aging onDMregions activity in the normal population

907F. Esposito et al. / Magnetic Resonance Imaging 26 (2008) 905–913

[18,23,24]. Here we apply single-subject ICA and thepreviously developed self-organizing group-level ICA (sog-ICA [22]) to evaluate statistically and to predict the effect ofnormal aging on the DM component. Specifically, we addressthe point of how to prepare and evaluate the bias of a DMtemplate, starting from the individual components clustered bythe used group-level ICA approach.

2. Materials and methods

2.1. Subjects and fMRI experiments

Twenty healthy subjects (with ages ranging from 23 to86 years) participated in an RS-fMRI study. Functional andanatomical MRI scans were performed on a 1.5-T SiemensMRI system (Siemens, Erlangen, Germany).

Functional MR data consisted of 120 volumes of a repeatedgradient-echo echo planar imaging T2

*-weighted sequence(TR=2100 ms, axial slice=20, matrix=64×64, field ofview=220mm, thickness=5mm, interslice gap=1mm).Duringthe functional scan, subjects were asked to simply relax and tokeep their eyes closed. Three-dimensional T1-weightedisometric images (Magnetization-Prepared RApid Gradient-Echo sequence, TR=2160 ms, TI=1100 ms, TE=3.9 ms, flipangle=10, voxel size=1×1×1 mm3) were acquired in the samesession in order to have high-resolution spatial references forregistration and normalization of the functional images.

2.2. Data preprocessing

Functional image time series were first corrected fordifferences in slice acquisition times using a “sinc”interpolation technique, then realigned with T1 volumes,warped into the standard anatomical space of Talairach andTournoux [25] and finally resampled to 3×3×3-mm3

isotropic voxels. The resulting voxel time series were filteredin time and space: very-low-frequency (drift) fluctuationswere reduced using a high-pass temporal filter (three cycles);a modest spatial smoothing with an isotropic Gaussian kernelof 5 mm full width at half-maximum was performed beforegroup-level analysis. For display purposes on the volumetricanatomy, individual maps were projected on the averagenormalized volumetric image.

All image data preparations (preprocessing and post-processing), as well as the map volumetric projection, wereperformed in BrainVoyager QX (Brain Innovation, Maas-tricht, The Netherlands), version 1.9.

2.3. Single-subject and group-level ICA

Individual-level ICA and sog-ICA [22] were applied tothe preprocessed functional time series using two “plug-in”extensions of BrainVoyager QX [26].

The single-subject ICA plug-in corresponded to a C++implementation of the fastICA algorithm [8]. Prior to ICAdecomposition, the initial dimensions of the functional dataset were reduced to 20 using PCA. Then, an equal number ofspatially independent components were estimated using the

deflation variant of fastICA and the “skew” function asnonlinearity. The final number of dimensions was chosenafter a reliability analysis had been performed on eachsubject data set. Using the methods described in Himberget al. [27], we estimated the number of reliable independentcomponents estimated by ICA for multiple dimensionalitiesand chose the value above which adding additionaldimensions did not provide additional reliable componentestimates in any of the subjects.

Sog-ICA [22] was used to summarize relevant indepen-dent components at the group level. This frameworkrequires, first, analyzing individual data sets and then, on agroup level, “clustering” the independent components in thesubject space. The clustering algorithm is based oncomponents' mutual similarity measures, here implementedas linear spatial correlations in a common whole-brain maskdefined in the 3-mm resampled Talairach space. The clustersize, which is the number of individual components that areallowed to be grouped into one group-level component, wasset equal to the number of subject data sets included in thegroup-level analysis.

Cluster “group” components were calculated as random-effects maps using the Random-Effects Analysis of Covar-iance module of BrainVoyager QX. According to thisimplementation, a random-effects statistic for each voxelcan be calculated from the individual ICA z-value of thatvoxel across all the individual maps within the standardframework of repeated-measures analysis of variance.

After performing a sog-ICA run of the entire group of 20subjects, we selected the DM component cluster andgenerated two random-effects group t-maps of the DMcomponent for two groups including, respectively, the 10youngest subjects and the 10 oldest subjects (young andelderly subject groups). In addition, a linear correlationanalysis of the individual ICA z-scores with age wasevaluated in each region activated in the random-effectsDM map of the young subject group after local spatialaveraging (ROI-based approach).

2.4. Individual-level and population-level analyses of theICA–DM network

The entire spatial distribution of a single individual DMcomponent was also evaluated in one threshold-independentscore (DM score) according to a template-based proceduresimilar to that in Greicius et al. [18]. This procedure attemptsto generalize the spatial layout and the value distribution ofthe investigated sample of subjects for possible use in largerpopulation studies and clinical trials. The general scheme ofthis approach is illustrated in Fig. 1.

First, sog-ICA is performed on a “reference” group ofsubjects (template subjects). Then, the group DM cluster isselected, and the average z-map obtained from all subjects orsome of the subjects is kept as DM template. Once thetemplate is chosen, the absolute spatial correlation coeffi-cient between each individual ICA component and the DMtemplate can be calculated for any new subject entering the

Fig. 1. General scheme of the approach used in the study. Group-level ICA is used in a group of subjects to generate a spatial template of the DM network. Spatiacorrelation is performed on all components from all subjects to extract the DM component and to produce a DM score for individual classification andpopulation-level analysis.

908 F. Esposito et al. / Magnetic Resonance Imaging 26 (2008) 905–913

analysis in this second stage (e.g., test subjects or patients).For each test subject, the component with the highest score istaken as the individual DM component, and this score is keptfor individual classification and population-level statistics.

Here we tested the quality of the score in characterizingthe studied sample of subjects and the degree of general-ization of the procedure for two sog-ICA-derived templates.Assuming the factor Age as the unique between-factor ofinterest in one case, we considered explicitly the subjectfactor in the template definition and included the DMcomponents of all young subjects only (template [Y]); inanother case, we ignored the factor and considered the entiregroup of subjects for generating the template (template[Y+E]). The quality of the score was evaluated by means of alinear correlation analysis of the subject-specific scores withthe factor Age.

The degree of generalization of the templates wasevaluated by repeating the linear correlation analysis onthe same subjects using a “leave-one-out” procedure in thetemplate-generation step. According to this procedure, weused a different template for each subject of the studiedsample, in such a way that each subject was evaluated on thebasis of an independent DM template. We identified thesetemplates as template [Y−1] and template [(Y+E)−1],respectively, for the case of only young subjects includedand all subjects included.

3. Results and discussion

Consistently with previous studies [14–17], all healthysubjects showed a component resulting from the coactivation

l

of the PCC and the ACC, with additional spots in the medialprefrontal, lateral and medial temporal, and inferior parietalcortices, as described in details below. We labeled this as theDM component.

All individual DM components from all subjects weresuccessfully clustered to the same group component by thesog-ICA procedure. The selection of the group-level DMcomponent after the sog-ICA run was highly facilitated byreranking the cluster in terms of their average intersubjectsimilarity, similarly to a previous study [17]. Besides the DMcomponent, a number of other resting-state componentsexhibited a high level of intersubject similarity. A compre-hensive report and discussion of most typical “resting-state”independent components extracted from resting-state timeseries can be found in van de Ven et al. [14], Damoiseauxet al. [15], Mantini et al. [16] and Esposito et al. [17].

Fig. 2 shows the group-level random-effects DM maps ofthe 10 young subjects (upper slices) and the 10 elderlysubjects (lower slices). The maps are projected onto theaveraged anatomy after reslicing to the original axial cuts.

The group-level DM pattern of the young subjects ischaracterized by the presence of 10 main regions of activitysurviving the random-effects threshold of P=.05 (Bonfer-roni corrected). Based on the average anatomy, theseregions include part of the PCC, extending dorsally into theprecuneus along the midline; two bilateral spots in theinferior parietal cortex (L-IPC and R-IPC) at the occipito-parietal junctions; the perigenual part of the ACC,extending ventrally from the rostral part into the orbito-frontal cortex; two bilateral clusters in the medial prefrontalcortex (L-MPFC and R-MPFC); and two bilateral temporalclusters located laterally in the inferolateral temporal cortex

Fig. 2. Group-level random-effects DM maps of the 10 young subjects (upper slices) and the 10 elderly subjects (lower slices). The maps are projected on theaveraged anatomy after reslicing the statistical image to the original axial cuts. A statistical threshold corresponding to P=.05 (Bonferroni corrected) is applied

Table 1Linear correlation coefficient between the spatially averaged z-ICA valuesfrom the individual ICA maps and the factor Age across all 20 subjects in alDM regions

Region Acronym R P

Whole DM network −.445 .049Anterior cingulate cortex ACC −.658 .0016Posterior cingulate cortex PCC −.424 .062Left inferior parietal cortex L-IPC −.125 .589Right inferior parietal cortex R-IPC −.2809 .23Left inferolateral temporal cortex L-ILTC −.07 .741Right inferolateral temporal cortex R-ILTC −.408 .07Left hippocampus L-H −.191 .418Right hippocampus R-H −.372 .106Left medial prefrontal cortex L-MPFC −.455 .0436Right medial prefrontal cortex R-MPFC −.178 .454

909F. Esposito et al. / Magnetic Resonance Imaging 26 (2008) 905–913

(L-ILTC and R-ILTC) and medially in the hippocampus(L-H and R-H).

The group-level DM pattern of the elderly subjects washighly overlapping with the one from the young subjectsgroup. As expected on the basis of previous studies[23,24], the pattern obtained from the elderly group wasgenerally weaker.

After extracting the regions of activity from the younggroup DM pattern, we performed a linear correlation analysisbetween the spatially averaged z-ICA values from theindividual ICA maps and the factor Age across all 20subjects. The same analysis was performed after spatialaveraging across all the activated voxels in the young DMNnetwork. The results are presented in Table 1 and plotted inFig. 3. According to the performed statistics, the ICA-basedDM connectivity averaged over all significant voxels in therandom-effects DM map was significantly correlated withage (Pb.05). A negative trend for the ICA z-values with agewas present in all DM regions, although at a rather variabledegree. Statistically significant negative correlations werefound in ACC and L-MPFC; PCC and R-ILTC exhibited

.

negative correlation trends very close to significance (seetable and graphs).

Despite the small sample size used here, which certainlylimits the interpretation of the presented results, the outcomeof the ROI analysis reveals the main problems arising when

l

Fig. 3. Scatter plot and linear correlation analysis between the spatially averaged z-ICAvalues from the individual ICA maps and aging across all 20 subjects. Alplots are labeled according to the ROI from which the individual z-ICA values were drawn and averaged. The bigger plot (upper left) is based on the individuaz-ICAvalues averaged across all voxels of all regions activated in the random-effects DMmap. The best linear fit (and the corresponding R2 statistics) is reportedon the same plots.

910 F. Esposito et al. / Magnetic Resonance Imaging 26 (2008) 905–913

defining an optimal empirical template for the DM networkand when summarizing the entire DM connectivity in onesingle value to be used (e.g., in a “goodness-of-fit” patternanalysis). In fact, the Age effect appears substantiallyinhomogenous across all DM-connected brain regions.Moreover, defining the ROI on one group (young) mapimplicitly renders the z-scores of any additional subjectlower, on average, because of the regression toward themean, thereby introducing spurious correlations whenever anewly added subject is also older.

A possible strategy for dealing with the spatial hetero-geneity of the effects is to use age as covariate and tocorrelate the subject-specific ICA z-scores at the level ofvoxels. While avoiding the inherent problem of voxelpreselection, this approach requires performing a statisticaltest in each single isolated voxel of the brain without takinginto account the spatial variation of the ICA values.Moreover, the resulting correlation maps are to be correctedfor multiple comparisons over the entire brain, thus reducing

ll

substantially the statistical power of this approach andrequiring bigger sample sizes. When applied to our sample of20 subjects, none of the voxels surpassed the statisticalthreshold after correction for multiple comparisons. At anuncorrected threshold (Pb.01, not shown), only smallclusters around the voxels with the largest effects exhibiteda negative correlation value inside the DM network, whosedegree of generalization for a population analysis appearedrather limited in the present case.

For these reasons, we considered using an entiredistribution of values as template and evaluated the twooptions of estimating it by either taking or not taking intoaccount a subject in the population-level analysis (“leave-one-out” approach). Since the purpose of the spatiallydistributed template is to characterize the average spatialvariation of the component weighting without the addi-tional scaling by the interindividual standard deviationimplied by the t-score necessary for statistical thresholding,we used the averaged z-scores from the individual ICA

Fig. 4. Template map (average z-ICA values from all individual decompositions) for the case of all subjects included ([Y+E]). The maps are projected on theaveraged anatomy after reslicing to the original axial cuts. No statistical threshold is applied. The map values are scaled between 0 and 2 to illustrate descriptivelythe entire distribution of values.

Fig. 5. Scatter plot and linear correlation analysis between the DM scores (i.e., the spatial correlation of the individual best-correlating component with the DMtemplate) and aging across all 20 subjects. As also indicated by the labels, in the upper plots, the templates [Y] and [Y+E] are used. In the lower plots, thetemplates [Y−1] and [(Y+E)−1] are used. The best linear fit (and the corresponding R2 statistics) is reported on the same plots.

911F. Esposito et al. / Magnetic Resonance Imaging 26 (2008) 905–913

912 F. Esposito et al. / Magnetic Resonance Imaging 26 (2008) 905–913

decompositions to generate the map. The entire spatialdistribution of the template (positive and negative values) isdescriptively shown in Fig. 4 for the case of all 20 subjectsincluded ([Y+E]).

Fig. 5 plots the individual DM scores versus age plots forall 20 subjects. In the upper plots, we evaluated the scores forall subjects, whether or not these were included in thetemplate, and correlated them with the factor Age. The linearcorrelation test was significant in both cases (Pb.05). Asexpected, the statistical significance of the test was muchhigher when considering only the 10 young subjects for thetemplate ([Y]: R=−.683, P=.0009) than when considering all20 subjects ([Y+E]: R=−.449, P=.0467).

In order to assess the degree of bias of the two templates,we repeated the correlation tests using the “leave-one-out”approach in the template-generation stage. This ensured thatthe score of each subject was based on a template that did notinclude the subject itself for the pattern generation. Hence,we expect these tests to give a better indication of howpredictive the DM score is for one individual with respect toage. The resulting correlations between the new scores andages (lower plots in Fig. 4), although lower compared to thecorresponding previous cases, remain statistically significant(Pb.05). Interestingly, the reduction in the correlation wassubstantial in the case of a young-only template ([Y−1]: R=−.451, P=.0457), but was only modest for the case of atemplate with both young and elderly subjects ([(Y+E)−1]:R=−.444, P=.0494). The reduction observed from [Y] to[Y−1] can be interpreted as the effect of removing the biasfrom the template generation. The fact that both templates[Y−1] and [(Y+E)−1] preserved the significant correlation ofthe DM score with age is an indication of a goodgeneralization of the DM patterns used for the scorecalculations. Particularly, the correlation between DMscore and age based on the [Y+E] template is highlyconsistent with previous studies [18,23,24].

In conclusion, ICA-based functional connectivity analy-sis of RS-fMRI data sets can provide a stable network-leveldescription of the DM system. Crucially, this statement doesnot imply that the single constituent regions of the patternscannot reduce or increase, at different degrees, theirextensions over different experimental runs of the samesubject and across different subjects within the samepopulation. Differently from other functional connectivityapproaches, the ICA globally and distributively modelsfunctional connectivity and poses only general constraints onthe spatial distribution of the components. Thus, this patternremains free from reflecting session-specific and subject-specific factors [17]. When preparing a DM template for thespatial correlations of individual DM components, the choiceof the template might become an issue. Using a distributedtemplate has the advantage of being global and thresholdindependent. Using a two-stage group-level ICA methodsuch as sog-ICA for DM template generation gathers theopportunity of testing the degree of generalization of theresult with respect to known subject factors that might have a

substantial impact on the individual scoring of the DMcomponent from one new subject or patient. However, thismethod is sensitive to the bias of incorporating a givensubject into the template to which the comparison is made,particularly in two-group analyses when the template isformed from one group exclusively. Nonetheless, this biascan be easily removed by either using a separate sample ofsubjects for the template generation or preparing multipletemplates according to a “leave-one-out” procedure.

Here we showed how it is possible to identify a stablecollection of group-level DM networks using the sog-ICAmethod and then prepare an empirical template with goodgeneralization performance. The generation of an unbiasedspatial template appears to be an important step for theintergroup statistical analysis of individual ICA componentsin population-level analyses.

References

[1] Biswal BB, Van Kylen J, Hyde JS. Simultaneous assessment of flowand BOLD signals in resting-state functional connectivity maps. NMRBiomed 1997;10(4–5):165–70.

[2] Gusnard DA, Raichle ME. Searching for a baseline: functional imagingand the resting human brain. Nat Rev Neurosci 2001;2(10):685–94.

[3] Raichle ME, MacLeod AM, Snyder AZ, Powers WJ, Gusnard DA,Shulman GL. A default mode of brain function. Proc Natl Acad SciU S A 2001;98:676–82.

[4] Greicius MD, Krasnow B, Reiss AL, Menon V. Functionalconnectivity in the resting brain: a network analysis of the defaultmode hypothesis. Proc Natl Acad Sci U S A 2003;100:253–8.

[5] Badgaiyan RD, Posner MI. Mapping the cingulate cortex in responseselection and monitoring. Neuroimage 1998;7(3):255–60.

[6] McKiernan KA, Kaufman JN, Kucera-Thompson J, Binder JR.A parametric manipulation of factors affecting task-induced deactiva-tion in functional neuroimaging. J Cogn Neurosci 2003;15:394–408.

[7] Bullmore ET, Rabe-Hesketh S, Morris RG, Williams SC, Gregory L,Gray JA, et al. Functional magnetic resonance image analysis of alarge-scale neurocognitive network. Neuroimage 1996;4(1):16–33.

[8] Hyvarinen A, Karhunen J, Oja E. Independent component analysis.New York: John Wiley and Sons, Inc.; 2001.

[9] McKeown MJ, Makeig S, Brown GG, Jung TP, Kindermann SS,Bell AJ, et al. Analysis of fMRI data by blind separation intospatial independent component analysis. Hum Brain Mapp 1998;6:160–88.

[10] Seifritz E, Esposito F, Hennel F, Mustovic H, Neuhoff JG, Bilecen D,et al. Spatiotemporal pattern of neural processing in the humanauditory cortex. Science 2002;297(5587):1706–8.

[11] Castelo-Branco M, Formisano E, Backes W, Zanella F, Neuensch-wander S, Singer W, et al. Activity patterns in human motion-sensitiveareas depend on the interpretation of global motion. Proc Natl Acad SciU S A 2002;99(21):13914–9.

[12] Malinen S, Hlushchuk Y, Hari R. Towards natural stimulation in fMRI— issues of data analysis. Neuroimage 2007;35(1):131–9.

[13] Bartels A, Zeki S. Brain dynamics during natural viewing conditions— a new guide for mapping connectivity in vivo. Neuroimage 2005;24(2):339–49.

[14] van de Ven VG, Formisano E, Prvulovic D, Roeder CH, Linden DE.Functional connectivity as revealed by spatial independent componentanalysis of fMRI measurements during rest. Hum Brain Mapp 2004;22(3):165–78.

[15] Damoiseaux JS, Rombouts SA, Barkhof F, Scheltens P, Stam CJ, SmithSM, et al. Consistent resting-state networks across healthy subjects.Proc Natl Acad Sci U S A 2006;103(37):13848–53.

913F. Esposito et al. / Magnetic Resonance Imaging 26 (2008) 905–913

[16] Mantini D, Perrucci MG, Del Gratta C, Romani GL, CorbettaM. Electrophysiological signatures of resting state networks inthe human brain. Proc Natl Acad Sci U S A 2007;104(32):13170–5.

[17] Esposito F, Bertolino A, Scarabino T, Latorre V, Blasi G, Popolizio T,et al. Independent component model of the default-mode brainfunction: assessing the impact of active thinking. Brain Res Bull2006;70(4–6):263–9.

[18] Greicius MD, Srivastava G, Reiss AL, Menon V. Default-modenetwork activity distinguishes Alzheimer's disease from healthy aging:evidence from functional MRI. Proc Natl Acad Sci U S A 2004;101:4637–42.

[19] Calhoun VD, Adali T, Pearlson GD, Pekar JJ. A method for makinggroup inferences from functional MRI data using independentcomponent analysis. Hum Brain Mapp 2001;14:140–51.

[20] Svensen M, Kruggel F, Benali H. ICA of fMRI group study data.Neuroimage 2002;16:551–63.

[21] Beckmann CF, Smith SM. Tensorial extensions of independentcomponent analysis for multisubject FMRI analysis. Neuroimage2005.

[22] Esposito F, Scarabino T, Hyvarinen A, Himberg J, Formisano E,Comani S, et al. Independent component analysis of fMRI groupstudies by self-organizing clustering. Neuroimage 2005;25:193–205.

[23] Persson J, Lustig C, Nelson JK, Reuter-Lorenz PA. Age differences indeactivation: a link to cognitive control? J Cogn Neurosci 2007;19(6):1021–32.

[24] Grady CL, Springer MV, Hongwanishkul D, McIntosh AR, WinocurG. Age-related changes in brain activity across the adult lifespan.J Cogn Neurosci 2006;18(2):227–41.

[25] Talairach J, Tournoux P. Co-planar stereotaxic atlas of the human brain.New York: George Thieme Verlag Stuttgart, Thieme MedicalPublishers, Inc.; 1988.

[26] Goebel R, Esposito F, Formisano E. Analysis of Functional ImageAnalysis Contest (FIAC) data with BrainVoyager QX: from single-subject to cortically aligned group general linear model analysis andself-organizing group independent component analysis. Hum BrainMapp 2006;27:392–401.

[27] Himberg J, Hyvarinen A, Esposito F. Validating the independentcomponents of neuroimaging time series via clustering and visualiza-tion. Neuroimage 2004;22:1214–22.