Tracking Dynamic Interactions Between Structuraland Functional Connectivity:

A TMS/EEG-dMRI Study

Enrico Amico,1,2 Olivier Bodart,1 Mario Rosanova,3 Olivia Gosseries,1,4 Lizette Heine,1 Pieter Van Mierlo,5

Charlotte Martial,1 Marcello Massimini,3 Daniele Marinazzo,2,* and Steven Laureys1,*

Abstract

Transcranial magnetic stimulation (TMS) in combination with neuroimaging techniques allows to measure theeffects of a direct perturbation of the brain. When coupled with high-density electroencephalography (TMS/hd-EEG), TMS pulses revealed electrophysiological signatures of different cortical modules in health and disease.However, the neural underpinnings of these signatures remain unclear. Here, by applying multimodal analyses ofcortical response to TMS recordings and diffusion magnetic resonance imaging (dMRI) tractography, we inves-tigated the relationship between functional and structural features of different cortical modules in a cohort ofawake healthy volunteers. For each subject, we computed directed functional connectivity interactions betweencortical areas from the source-reconstructed TMS/hd-EEG recordings and correlated them with the correspon-dent structural connectivity matrix extracted from dMRI tractography, in three different frequency bands (a,b, c) and two sites of stimulation (left precuneus and left premotor). Each stimulated area appeared to mainlyrespond to TMS by being functionally elicited in specific frequency bands, that is, b for precuneus and c for pre-motor. We also observed a temporary decrease in the whole-brain correlation between directed functional con-nectivity and structural connectivity after TMS in all frequency bands. Notably, when focusing on the stimulatedareas only, we found that the structure–function correlation significantly increases over time in the premotor areacontrolateral to TMS. Our study points out the importance of taking into account the major role played by dif-ferent cortical oscillations when investigating the mechanisms for integration and segregation of information inthe human brain.

Keywords: brain-directed functional connectivity; directed functional connectivity; DTI; structural connectivity;structure–function; TMS/EEG

Introduction

Transcranial magnetic stimulation (TMS) has beenused for more than 20 years to investigate connectivity

and plasticity in the human cortex. By combining TMSwith high-density electroencephalography (hd-EEG), onecan stimulate any cortical area and measure the effects pro-duced by this perturbation in the rest of the cerebral cortex(Ilmoniemi et al., 1997). It has been shown that cortical po-

tentials elicited by TMS (TMS-evoked potentials [TEPs])mainly generate a significant EEG response during thefirst 300 msec in wakefulness (Massimini et al., 2005;Rosanova et al., 2009). The effects of TMS might alsolast for as long as 600 msec, during their spread from thearea of stimulation to remote interconnected brain areas(Bonato et al., 2006; Lioumis et al., 2009). To date, TMS/EEG recordings have provided new and reliable insightson the whole brain cortical excitability with reasonable

1Coma Science Group, Cyclotron Research Center & GIGA Research Center, University and University Hospital of Liege, Liege, Belgium.2Department of Data-Analysis, University of Ghent, Ghent, Belgium.3Department of Biomedical and Clinical Sciences ‘‘Luigi Sacco,’’ University of Milan, Milan, Italy.4Department of Psychiatry, University of Wisconsin, Madison, Wisconsin.5Medical Image and Signal Processing Group, Department of Electronics and Information Systems, Ghent University-IBBT, Ghent, Belgium.*These authors contributed equally to this work.

BRAIN CONNECTIVITYVolume 7, Number 2, 2017ª Mary Ann Liebert, Inc.DOI: 10.1089/brain.2016.0462

84

spatial and excellent temporal resolution (Gosseries et al.,2015; Rogasch and Fitzgerald, 2013).

The amount of information contained in the hd-EEG re-sponse to TMS appears to contain inner signatures of thefunctional organization in a brain network. Two recent stud-ies (Ferrarelli et al., 2012; Rosanova et al., 2009) in healthyawake subjects showed that TMS can also induce EEG oscil-lations at different frequencies. The TMS pulse gives rise todifferent connected cortical regions in the brain, generating acomplex EEG pattern composed of strong fluctuations at the‘‘natural’’ frequency of the stimulated area. These oscilla-tions are believed to reflect neurophysiological activity thatis transiently elicited by the TMS pulse and possibly engagedthrough brain connections (Cona et al., 2011; Ferrarelli et al.,2012; Rosanova et al., 2009).

Recently, researchers have started to investigate how theTMS/hd-EEG perturbation might be constrained and shapedby brain structure, either by exploring the correlation be-tween TMS-induced interhemispheric signal propagationand neuroanatomy (Groppa et al., 2013; Voineskos et al.,2010) or by improving the modeling of the TMS-inducedelectric field by using realistic neural geometry (Bortolettoet al., 2015; De Geeter et al., 2015). Besides, it has latelybeen shown that cortical networks derived from sourceEEG connectivity partially reflect both direct and indirectunderlying white matter (WM) connectivity in a broadrange of frequencies (Chu et al., 2015).

In this respect, the development of diffusion magnetic res-onance imaging (dMRI) might add information on the struc-tural architecture of the brain (Catani et al., 2002). Theapplication of deterministic and probabilistic tractographymethods allows for the spatial topography of the WM,which represents bundles of coherently organized and mye-linated axons (Song et al., 2002). The output of tractographyalgorithms permits anatomically plausible visualization ofWM pathways and has led to reliable quantification of struc-tural connections between brain regions (i.e., the human con-nectome) (Bullmore and Sporns, 2009; Sporns et al., 2005).

The purpose of this proof-of-concept article is to investi-gate EEG changes of directed functional connectivity inthe brain induced by TMS from both a functional anda structural perspective, using multimodal modeling ofsource-reconstructed TMS/hd-EEG recordings and dMRItractography. The study of functional connectivity changesafter the perturbation can possibly help in understanding thestructure–function modulation caused by TMS (i.e., the ex-tent to which TMS-induced EEG dynamics is constrainedby WM pathways) and the specific frequency bands of the in-volved brain regions. Taking the aforementioned recent find-ings as a starting point, we here aim at assessing: (1) the roleof the ‘‘natural frequencies’’ in the TMS-induced functionalconnectivity changes (Ferrarelli et al., 2012; Rosanova et al.,2009) and (2) the extent to which functional connectivity, asa consequence of the induced perturbation, is shaped by brainstructure.

We will first present the processing pipelines for TMS-EEG and dMRI data. Second, the mathematical methodol-ogy for the evaluation of the directed functional connectiv-ity between brain regions and its correlation with thestructural connectome will be presented. Finally, resultsobtained in a cohort of healthy volunteers will be presentedand discussed.

Materials and Methods

TMS/hd-EEG recordings: acquisition and preprocessing

TMS/hd-EEG data were acquired in 14 healthy awakeadults (6 men, age range 23–37 years) as published else-where (Casali et al., 2010, 2013; Rosanova et al., 2012). Inbrief, subjects were sitting on a reclined chair with eyesopen looking at a fixation point on a screen. All participantsgave written informed consent and underwent clinical exam-inations to rule out any potential adverse effect of TMS.

The TMS/hd-EEG experimental procedure, approved by theLocal Ethical Committee of the University of Liege, was per-formed by using a figure-of-eight coil driven by a mobile unit(eXimia TMS Stimulator; Nexstim Ltd.), targeting two corticalareas (left superior parietal and left premotor) for at least 200trials. These areas were selected for the following reasons:(1) They are easily accessible and far from major head or facialmuscles whose activation may affect EEG recordings; (2) pre-vious TMS/EEG studies have been successfully performed inthese areas during wakefulness (Casarotto et al., 2010; Massi-mini et al., 2005; Rosanova et al., 2009); and (3) because theseareas are part of the fronto-parietal network and, thus, are struc-turally and functionally highly connected. The left superior pa-rietal and left premotor targets were identified on the subject’s3D T1 brain scan and reached through the neuronavigation sys-tem (NBS; Nexstim Ltd.) by using a stereoscopic infraredtracking camera and reflective sensors on the subject’s headand the stimulating coil.

Electrical brain activity was recorded by using a 60-channel TMS-compatible EEG amplifier (Nexstim eXimia;Nexstim Plc) with a sample-and-hold circuit, which preventsthe amplifier from saturation (see Supplementary Data fordetails; Supplementary Data are available online at www.liebertpub.com/brain).

Channels dominated by artifacts, such as 50 Hz noise, elec-trode movements, or strong TMS-induced artifact, were re-moved (maximum of 10 per session) and later interpolated.The session was discarded if the removed channels were clus-tered in such a way that interpolation would have been impos-sible or limited in accuracy. A minimum of 200 average trialswas kept for further analyses. The percentage of excluded tri-als varied from subject to subject but was less than 10%. TMStrials containing noise, muscle activity, or eye movementswere detected and rejected (Rosanova et al., 2012), and thesession was performed until 200 accepted trials were recorded,or a maximum of 400 per site (never reached in this healthysubjects population). Independent component analysis (ICA)was also used as the last step in the signal preprocessing insome subjects, to remove artifacts from ocular movementsor remaining 50 Hz noise, that could not be dealt with in theprevious steps.

Out of the initial 14 subjects, we excluded 5 of them forthe superior parietal target and 2 for the premotor target, be-cause of a low signal-to-noise ratio of TMS/EEG-evokedresponses [i.e., when the ratio between post-stimulus andpre-stimulus amplitude is below 1.4, as in Casali et al.(2010, 2013)]. EEG data were referenced by means of the ref-erence electrode standardization technique (Liu et al., 2015;Yao, 2001), downsampled at half of the original samplingrate (from 725 to 362 Hz), and bandpass filtered (2 to 80 Hz).

Source reconstruction was performed as described byCasali and coworkers (2010, 2013). Conductive head volume

TMS/EEG-DMRI INTERACTIONS IN THE HUMAN BRAIN 85

was modeled according to the 3-spheres BERG method(Berg and Scherg, 1994) and constrained to the cerebral cor-tex that was modeled as a three-dimensional grid of 3004fixed dipoles that were oriented normally to the cortical sur-face. This model was adapted to the anatomy of each subjectby using the Statistical Parametric Mapping software pack-age (SPM8, freely available at www.fil.ion.bpmf.ac.uk/spm) as follows: Binary masks of skull and scalp obtainedfrom individual MRIs were warped to the correspondingcanonical meshes of the Montreal Neurological Institute(MNI) atlas. Then, the inverse transformation was applied tothe MNI canonical mesh of the cortex for approximating toreal anatomy. Finally, the single-trial distribution of elec-trical sources in the brain was estimated by applying theempirical Bayesian approach as described by Mattout andassociates (2006) and Phillips and colleagues (2005).

To summarize significant functional measures over ana-tomically and/or functionally identifiable brain regions, thetime courses of the 3004 reconstructed sources were then av-eraged into the specific 90 cortical and subcortical areas ofthe automated anatomical labeling (AAL) (Tzourio-Mazoyeret al., 2002) parcellation (Fig. 1), according to their positionon the cortical mesh (Casali et al., 2010). Specifically, theAAL template was first registered onto the cortical mesh,allowing for a mapping between the 3004 fixed dipoles andthe regions of interest (ROIs) of the template. Then, thetime series of the dipoles falling in the same AAL brain re-gion were averaged together.

dMRI data: acquisition and preprocessing

A series of diffusion-weighted magnetic resonance images(dwi) of brain anatomy were acquired in each participant byusing a Siemens Trio Magnetom 3 Tesla system (SiemensTrio, University Hospital of Liege, Belgium). dwi imageswere acquired at a b-value of 1000 sec/mm2 by using 64 encod-ing gradients that were uniformly distributed in space by anelectrostatic repulsion approach (Jones et al., 1999). Voxelshad dimensions of 1.8 · 1.8 · 3.3 mm3, and volumes were ac-quired in 45 transverse slices by using a 128 · 128 voxel ma-trix. A single T1-weighted 3D magnetization-prepared rapidgradient echo sequence (MPRAGE) image, with an isotropicresolution of 1 mm3, was also acquired for each subject.

Diffusion volumes were analyzed by using typical prepro-cessing steps in dMRI (Caeyenberghs et al., 2012; Zaleskyet al., 2014). Eddy current correction for each participantwas achieved by using FDT, v2.0, the diffusion toolkit withinFSL 5.0 (FMRIB Software Library; www.fmrib.ox.ac.uk/fsl). Rotations applied to the diffusion-weighted volumeswere also applied to the corresponding gradient directions(Leemans and Jones, 2009). A fractional anisotropy (FA)image was estimated by using weighted linear least squaresfitted to the log-transformed data for each subject.

dMRI data: registration of the anatomical imageand atlas parcellation

We segmented each subject’s T1-weighted image intowhole-brain WM, gray matter (GM), and cerebrospinal fluidmasks by using FAST, a part of FSL (FMRIB Software Libraryv 5.0). The corresponding WM mask image was registeredwithout resampling to the relevant dwi series (Smith et al.,2004). Next, the AAL atlas was first registered to the T1

space by using linear (FSL flirt) and non-linear warping (FSLFNIRT) to achieve the best registration into each subject’sspace. Then, the single-subject AAL template was finally reg-istered without resampling to the dwi space by using the affinetransform resulting from the WM registration. This transforma-tion matrix was also applied to the T1-derived GM mask, whichwas used as a termination mask for the tractography analysis.

dMRI data: tractography and connectome construction

The fiber response model was estimated for each subjectfrom the high b-value (b = 1000 sec/mm2) dwi. A mask ofsingle-fiber voxels was extracted from the thresholded anderoded FA images (Tournier et al., 2004, 2008). Using non-negativity constrained spherical deconvolution, fiber orienta-tion distribution (FOD) functions were obtained at eachvoxel by using the MRTRIX3 package (J-D Tournier; BrainResearch Institute, https://github.com/jdtournier/mrtrix3)(Tournier et al., 2012). For both the response estimation andspherical deconvolution steps, we chose a maximum harmonicorder lmax of 6, based on a successful application of con-strained spherical deconvolution with a similar b-value andnumber of directions (Roine et al., 2015).

Probabilistic tractography was performed by using ran-domly placed seeds within subject-specific WM masks, reg-istered as mentioned in the latter. Streamline trackingsettings were as follows: number of tracks = 10 million,FOD magnitude cutoff for terminating tracks = 0.1, mini-mum track length = 5 mm, maximum track length = 200 mm,minimum radius of curvature = 1 mm, tracking algorithmstep size = 0.5 mm. Streamlines were terminated when theyextended out of the WM-GM mask interface, or could notprogress along a direction with an FOD magnitude or a cur-vature radius higher than the minimum cutoffs.

The streamlines obtained were mapped to the relevantnodes defined by the AAL parcellation registered in the sub-ject’s dwi space, using MRTRIX3 (Tournier et al., 2012).Each streamline termination was assigned to the nearestGM parcel within a 2 mm search radius. The resulting con-nectome was finally examined by determining the connec-tion density (number of streamlines per unit surface)between any two regions of the AAL template, as describedby Caeyenberghs and coworkers (2012) (Fig. 1). This correc-tion was needed to account for the variable size of the corti-cal ROIs of the AAL template (Hagmann et al., 2008).

TMS/hd-EEG-directed functional connectivity estimation:spectrum-weighted adaptive-directed transfer function

We evaluated directed functional connectivity betweenthe 90 EEG reconstructed brain signals (see TMS/hd-EEGRecordings: Acquisition and Preprocessing section) byusing an extension of a data-driven technique based onGranger causality (Granger, 1969). According to Grangercausality, if a signal x1 ‘‘Granger causes’’ (or ‘‘G-causes’’)a signal x2, then past values of x1 should contain informationthat helps to predict x2 above and beyond the informationcontained in past values of x2 alone (Granger, 1969). Thisstatistical concept can be mathematically modeled as a mul-tivariate autoregressive model (MVAR):

xn = +p

m = 1

Amxn�mþ en (1),

86 AMICO ET AL.

where, for a number K of signals, xn is the matrix containingthe K signals at time n, en is a matrix containing the uncorre-lated white noise at time n, p is the model order (i.e., howmany past values are taken into account), and Am is theK · K coefficient matrix for delay m.

The directed transfer function (DTF) is a measure basedon Granger causality to analyze the propagation of activitybetween multiple signals in the frequency domain, bymeans of a multivariate model of spectral coefficients(Kaminski and Blinowska, 1991):

DTFij( f ) = Hij( f )��

��2

(2)

The information on the interactions between signals (i.e.,brain regions, in our case) in the frequency domain is nowcontained in the K · K matrix H (the equivalent of A in

Eq. 1 in the frequency domain), also known as the transfermatrix of the model, which contains information on the di-rected interactions between signals xi and xj at frequency f.

These models assume stationary signals as input. To copewith the non-stationary nature of most natural signals, an ex-tension of the DTF, the adaptive DTF (ADTF) (Arnold et al.,1998; Astolfi et al., 2008; Wilke et al., 2008), was proposed.To adapt the autoregressive model to non-stationary signals,the coefficients of the autoregressive model (i.e., matrix H inEq. 2) were allowed to vary in time, by means of a Kalmanfiltering procedure (Arnold et al., 1998) [the interested readermay refer to Van Mierlo et al. (2011, 2013) for details on themethodology].

This time-variant multivariate autoregressive (TVAR)model with time-varying parameters has been successfullyused for connectivity modeling of epileptic intracranial

FIG. 1. Flowchart of TMS/EEG-dMRI modeling. Up: The time courses of the 3004 reconstructed dipoles were averagedinto the parcels of the AAL atlas (Tzourio-Mazoyer et al., 2002), consisting of 90 unique brain regions (cerebellar regionswere excluded from the analysis). The 90 time courses obtained were modeled by using swADTF (Van Mierlo et al., 2011,2013). swADTF returns the causal interactions between the cortical regions (90 · 90 time-varying directed functional con-nectivity matrices) at a specific frequency interval (f1, f2). Bottom: For each dMRI dataset, whole-brain probabilistic trac-tography was performed by using a combination of FSL and MRTRIX (see Materials and Methods section). The AAL atlaswas then used to segment the streamlines fiber bundles between each pair of ROIs. Next, we determined the percentage oftracts between each pair of regions of the AAL template, resulting in a 90 · 90 structural connectivity matrix. AAL, auto-mated anatomical labeling; dMRI, diffusion magnetic resonance imaging; ROIs, regions of interest; swADTF, spectrum-weighted adaptive directed transfer function; TMS, transcranial magnetic stimulation.

TMS/EEG-DMRI INTERACTIONS IN THE HUMAN BRAIN 87

EEG data (Van Mierlo et al., 2011, 2013). To take into ac-count the non-stationary nature of the TMS pulse in ourEEG source-reconstructed signals, here, we adopted thespectrum-weighted version of the ADTF (swADTF) (VanMierlo et al., 2011, 2013):

swADTFij(t) =+f2

f = f1

jHij( f , t)j2 +K

k = 1

jHjk( f , t)j2

+K

k = 1

+f2

f = f1

jHik( f , t)j2 +K

k = 1

jHjk( f , t)j2(3),

where Hij( f , t) in Eq. 3 is the time-variant transfer matrix ofthe system describing the directed functional connectivityfrom signal j to i at frequency f at time t, for each of the Ksignals.

Each term Hij( f , t) is weighted by the autospectrum of thesending (in this case j) signal.

The swADTF allows us to investigate directional interac-tions between all the signals at a predefined frequency bandover time. The measure weights all outgoing directed func-tional connectivity present in the terms Hij( f , t) by thepower spectrum of the sending signal j, and it, therefore,does not depend on the power amplitude. Each swADTFvalue corresponds to the directed time-variant strength of thedirected functional connectivity between two nodes. This dy-namic interaction between nodes can also be represented as aseries of time-varying directed matrices (Fig. 1). The swADTFis normalized so that the sum of incoming directed functionalconnectivity into a channel at each time point is equal to 1:

+K

k = 1

swADTFik(t) = 1 (4)

TMS/EEG-dMRI multimodal integration:outdegree computation and statistical assessment

We computed directed functional connectivity (swADTF)on the brain network defined by the anatomical atlas (AAL)reconstructed sources for each subject. A detailed discus-sion on the implementation and the setup of the parameterscan be found as described by Van Mierlo and coworkers(2011).

The swADTF was calculated in three frequency bands: a(8–12 Hz), b (13–20 Hz), and b2=c (21–50 Hz), as describedby Rosanova and associates (2009). We chose these threebands for reasons of comparison with the study by Rosanovaet al., which provided evidence that TMS on healthy awakesubjects consistently evoked dominant EEG oscillations indifferent cortical areas.

To track modulations of directed functional connectiv-ity due to TMS, we considered 3 different non-overlappingwindows of 300 msec: a ‘‘baseline,’’ pre-stimulus, extendedfrom �300 to 0 msec before the TMS pulse; a ‘‘post-stimulus,’’ after the TMS pulse, which captures the dynamicsfrom 20 to 320 msec after the pulse (the first 20 msec werediscarded to minimize the effect of possible artifacts occur-ring at the time of stimulation) (Rogasch et al., 2013; Rosa-nova et al., 2009); and a ‘‘late post-stimulus,’’ from 320 to620 msec after the stimulation.

We obtained the mean global outgoing flow from a regionj before and after the stimulation by averaging the swADTF

time courses in each of the three time windows and by sum-ming the average amount of directed connectivity transferredfrom j to each node of the network. In network terms, thisquantity is called Outdegree. In our case, for each frequencyband and window (i.e., baseline or post-stimulus):

Outdegreej = +K

k = 1

Cjk, 8k, j = 1 . . . K, (5),

where K = 90 in our case (i.e., the number of AAL regions),and C is the connectivity matrix constructed by averaging theswADTF time courses within each window. All self-edgeswere set to 0. By using this procedure, we aimed at obtainingan illustrative snapshot of the total directed functional con-nectivity from a region j at a specific stage of the TMS pro-cess (i.e., baseline or post-stimulus).

To detect significant group changes in the Outdegree be-fore and after the stimulation, a two-sample t-test of thepost-stimulus Outdegree against the correspondent baselineOutdegree was performed in each region. Post-stimulus Out-degree values were considered significant at p < 0:05, andthen false discovery rate (FDR) corrected for multiple com-parisons over the 90 brain regions, the three frequency bands,and the two sites of stimulation.

Furthermore, to investigate how the integration and segre-gation of the functional brain network change after the TMSpulse, we evaluated some key topological properties of thedirected functional connectivity networks, before and afterthe TMS pulse. The following network measures were com-puted: clustering coefficient, betweenness centrality, richclub coefficient, and global efficiency (Bullmore and Sporns,2009; Sporns, 2011), by using the Brain Connectivity Tool-box (brain-connectivity-toolbox.net) (Rubinov and Sporns,2010). The differences pre-post TMS pulse of these networkmeasures were then tested based on Wilcoxon signed-ranktest, and then FDR corrected for multiple comparisons overthe 90 brain regions, the three frequency bands, and thetwo sites of stimulation.

Finally, for each subject, the structural degree of a node j(SCdegree) was simply calculated from the structural con-nectivity matrix S by summing over its columns.

SCdegreej = +K

k = 1

Sjk, 8 j = 1 . . . K, (6)

Structure–function correlations and statistical assessment

The dynamic interaction between regions modeled byswADTF can be represented as a series of time-varying di-rected connectivity matrices (Fig. 1). In each frequencyband, dynamic spatial correlation was defined as the meanrow-by-row Pearson’s correlation at each time point betweeneach subject’s directed functional connectivity matrix andthe correspondent structural connectivity matrix. The 95%confidence intervals for the Pearson’s correlation distributionat the baseline were calculated by using a non-parametricbootstrap procedure (Efron and Tibshirani, 1986). Specifi-cally, the correlation coefficient was recomputed n = 100times on the resampled data obtained by n random permuta-tions of the values in the directed functional connectivity ma-trices at each time point of the baseline, while leaving thestructural connectivity matrix unchanged. The empirical

88 AMICO ET AL.

distribution of the resampled dynamic spatial correlation val-ues at the baseline was used to approximate the sampling dis-tribution of the statistic. A 95% confidence interval for thebaseline was then defined as the interval spanning from the2.5th to the 97.5th percentile of the obtained distribution.Values in dynamic spatial correlation that were falling out-side this interval were considered significantly differentfrom the baseline correlation.

Results

The significant differences (Table 1) in directed functionalconnectivity, between baseline and post-stimulus window,across cortical regions after TMS perturbation are illustratedby projecting the Outdegree onto the anatomical template(Fig. 2). In the first 300 msec after the pulse, the two sitesof stimulation have significant Outdegree peaks at differentfrequency bands. However, we did not find any significantdifferences between Outdegree values of baseline and latepost-stimulus windows.

In particular, the superior parietal cortex has a maximumof directed functional connectivity in the b band in proximityof the stimulation site, whereas the premotor has a maxima in

the b2=c band, which is more spread toward the controlateralhemisphere. Statistical comparisons (i.e., double-sided t-tests, p < 0:05) of directed functional connectivity peakson the two sites of stimulation using a finer frequency step(i.e., every 2 Hz) provide further evidence that each target re-gion has Outdegree maxima at different frequencies (i.e.,within the b range for the superior parietal cortex, withinthe b2=c for the premotor cortex, Fig. 3).

The peaks in directed functional connectivity at differentfrequencies in the areas depicted in Figure 2 confirm the hy-pothesis that different brain regions might be usually tuned tooscillate at a characteristic rate (i.e., natural frequency) (Fer-rarelli et al., 2012; Rosanova et al., 2009). Furthermore, de-spite the fast and chaotic functional response generated bythe TMS pulse in the brain network, the cortical regions sig-nificantly recruited by TMS seem to maintain peaks of func-tional activation at the specific natural frequency consistentlyover time (Fig. 4 and Supplementary Videos S1 and S2).Notably, the connectivity peaks of the premotor site followthe specific differences in the power spectrum, whereas thepeaks of directed functional connectivity in the superior pa-rietal site appear to be more connectivity specific (Supple-mentary Fig. S1).

Table 1. Percentage Increase Before/After Stimulation for the Automated Anatomical Labeling Areas

Illustrated in Figure 2 Where Directed Functional Connectivity After Transcranial Magnetic

Stimulation Was Significantly Higher Than Baseline (p< 0:05, False Discovery Rate Corrected,

See Materials and Methods Section), for the Two Sites of Stimulation (Left Superior Parietal

and Left Premotor) and the Three Different Frequency Band (a, b, b2=c)

TMS site:left superior parietal TMS site:left premotor

AAL region Increase (%) FB AAL region Increase (%) FB

Precentral_L 120 a Supp_Motor_Area_L 210 aFrontal_Mid_L 133 a Supp_Motor_Area_R 190 aFrontal_Inf_Oper_L 70 a Cingulum_Post_R 60 aRolandic_Oper_L 60 a Cuneus_L 120 aSupp_Motor_Area_L 40 a Occipital_Sup_L 90 aCingulum_Post_R 53 a SupraMarginal_R 50 a

Sup_Parietal_L 80 aParacentral_Lobule_R 110 a

Precentral_L 150 b Frontal_Sup_L 185 bFrontal_Mid_R 70 b Frontal_Mid_L 150 bRolandic_Oper_L 50 b Supp_Motor_Area_L 220 bSupp_Motor_Area_L 90 b Supp_Motor_Area_R 260 bParietal_Inf_L 167 b Insula_L 50 bSup_Parietal_L 280 b Cingulum_Mid_R 52 b

Cingulum_Post_R 35 bCuneus_R 30 bCingulum_Mid_R 45 b

Precentral_L 233 b2/c Precentral_L 118 b2/cFrontal_Sup_L 152 b2/c Frontal_Sup_L 363 b2/cSupp_Motor_Area_L 126 b2/c Frontal_Mid_L 203 b2/cSupp_Motor_Area_R 130 b2/c Frontal_Mid_Orb_L 153 b2/cCingulum_Mid_L 100 b2/c Supp_Motor_Area_L 330 b2/cParietal_Inf_L 110 b2/c Supp_Motor_Area_R 432 b2/cSup_Parietal_L 150 b2/c Frontal_Sup_Medial_R 65 b2/c

Insula_L 33 b2/cCingulum_Mid_R 110 b2/cThalamus_L 33 b2/cTemporal_Mid_L 53 b2/c

Top increment for each site of stimulation is highlighted in bold.AAL, automated anatomical labeling; FB, frequency band; TMS, transcranial magnetic stimulation.

TMS/EEG-DMRI INTERACTIONS IN THE HUMAN BRAIN 89

It is also worth mentioning that none of the network mea-sures computed on the directed functional connectivity matri-ces before and after the TMS pulse survived to the statisticalsignificance threshold (p < 0:05 FDR corrected, see Materialsand Methods section). This might be due to the small samplesize of the cohort under study and to the within-subject vari-ability of the response to TMS (see Limitations section).

The dynamic spatial correlation between the directedfunctional connectivity (swADTF) and the connectome, forthe two different sites of stimulation (i.e., left superior pari-etal and left premotor) and for each of the three chosen fre-quency bands (i.e., a, b, b2=c), deviates from baseline after

the TMS pulse (Fig. 5). This global network behavior doesnot depend on the subject or the stimulation site. The stablebaseline configuration is then recovered after 200–300 msec,depending on the frequency band. Specifically, this TMS-induced modulation of EEG rhythms over the brain networkis more pronounced (i.e., higher deviation from the baselinecorrelation) and faster in the b2=c and b bands, whereas thereturn to baseline is slower and less pronounced in the aband. The evidence that different brain areas can be normallytuned by TMS to oscillate at a characteristic rate (i.e., naturalfrequency) might also explain the drop in structure–functioncorrelation depicted in Figure 5.

FIG. 2. Directed functionalconnectivity across corticalregions after TMS. First threerows: Snapshot of differencesbetween baseline and post-TMS stimulus directed func-tional connectivity (i.e., Out-degree) at p < 0:05, falsediscovery rate corrected (seeMaterials and Methods sec-tion) across cortical regions,for the three predefined fre-quency bands (a, b, b2=c)(Rosanova et al., 2009),obtained by averaging theswADTF time courses from20 to 320 msec after thepulse. The red circlescoarsely indicate the targetareas of stimulation. Bottomrow: z-scored map of thestructural connectivity profileof the two stimulated regions,representing amount of con-necting the superior parietalcortex (left) and the premotorcortex (right). The red circlescoarsely indicate the targetareas of stimulation. Notethat the superior parietal cor-tex has a maximum ofdirected functional connec-tivity in the b band in prox-imity of the stimulation site,whereas the premotor cortexhas a maxima in the b2=cband, which is more spreadtoward the hemisphere that iscontrolateral to the stimula-tion site. These brain imageswere obtained by usingBrainNet Viewer (Xia et al.,2013).

90 AMICO ET AL.

In fact, assuming that each of the 90 AAL cortical regionsresponds to TMS by oscillating at its peculiar natural fre-quency, the emergence of this complex between-band interac-tion might generate a consequent deflection in the within-bandstructure–function correlation (Fig. 5). This effect might bedue to the region-specific variability in the intensity and theduration of the cortical response to TMS at the different natu-ral frequencies, but it might also depend on the degree towhich each recruited region is structurally connected to therest of the network.

To further investigate whether the structural connectivityprofiles of the stimulated regions influence subsequentTMS-evoked connectivity, we evaluated the local dynamicspatial correlation between the directed functional connectiv-ity (swADTF) for the targeted cortical regions and the connec-tome, for both sites of stimulation (Fig. 6). Notably, thestructure–function correlation significantly increases overtime in the right premotor cortex after TMS, when its naturalfrequency band (i.e., b2=c) is taken into consideration. Thiseffect was not reproduced in the superior parietal area(Fig. 6). This increase in the structure–function correlationseems to be specific of the TMS-evoked response, and depen-dent on the cortical module elicited, as we did not observesuch a response when the same area was not stimulated, orin regions where the TMS pulse was most likely not causingany significant detectable effect (e.g., deep subcortical struc-tures such as putamen; Supplementary Figs. S2 and S3).

Discussion

In this work, we studied for the first time the interplay be-tween directed functional connectivity computed from TMS-reconstructed EEG sources and the connectome extractedfrom whole-brain dMRI tractography in a cohort of healthyvolunteers. We aimed at assessing: (1) whether naturalfrequencies of the stimulated areas play a role in the TMS-induced functional connectivity changes and (2) to what ex-

tent these functional connectivity changes are shaped bybrain structure. Later follows a detailed discussion of thefindings. First, this work confirms the hypothesis that differ-ent rhythms in the brain emerge after TMS, which was thefirst aim of our study. This dynamic interaction at differentnatural frequencies seems to reflect intrinsic properties ofcortical regions, and the way those are interconnected(Cona et al., 2011; Rosanova et al., 2009). Previous studiesrevealed that distant areas, when activated by TMS,responded with oscillations closer to their own ‘‘natural’’ fre-quency (Ferrarelli et al., 2012; Rosanova et al., 2009). Ouranalysis on peaks of significant changes in directed func-tional connectivity at different frequency bands corroboratedthe hypothesis that TMS evokes dominant oscillations in dif-ferent cortical areas at a characteristic rate. These findingsare in line with previous studies (Ferrarelli et al., 2012; Rosa-nova et al., 2009), where the authors showed that TMS onhealthy awake subjects consistently evokes EEG oscillationswith dominant frequencies that depend on the site of stimu-lation. In particular, when stimulated, the superior parietalcortex was shown to respond to TMS in the b band and thepremotor cortex in b2=c (Rosanova et al., 2009). Here, wehave tested and validated the natural frequency hypothesisby tracking directed functional connectivity interactions be-tween brain regions, and by comparing their response beforeand after TMS.

Each stimulated area appeared to mainly respond to thestimulation by being functionally elicited in specific ‘‘natu-ral’’ frequency bands, that is, b for superior parietal andb2=c for premotor (Figs. 2 and 3 and Table 1). Furthermore,these peaks of functional changes at the natural frequency ofthe stimulation site after TMS seem to be quite stable overtime (Fig. 4 and Supplementary Videos S1 and S2). Interest-ingly, the premotor cortex also showed a less pronounced ac-tivation in the b band (Fig. 2). This could arise from therecruitment after the stimulation of the primary motor cortex,

FIG. 3. Peaks of directed functional connectivity for the two stimulated regions. Barplots show Outdegree values for thetwo sites of stimulation (left, sup parietal; right, premotor) when using finer frequency ranges (2 Hz steps). Errorbars indicatestandard error over subjects. The red crosses on the top of the bars indicate that the Outdegree value is significantly differentfor the two target regions (double-sided t-test, p<0:05). Note how the two sites of stimulation have Outdegree maxima indifferent frequency ranges (i.e., within the b range for the superior parietal cortex, within the b2=c for the premotor cortex).

TMS/EEG-DMRI INTERACTIONS IN THE HUMAN BRAIN 91

FIG

.4.

Flo

wof

dir

ecte

dfu

nct

ional

connec

tivit

yover

tim

e.S

elec

ted

fram

esfr

om

Supple

men

tary

Vid

eos

S1

and

S2,

of

the

funct

ional

connec

tivit

yac

ross

cort

ical

regio

ns,

for

both

site

sof

stim

ula

tion

(top

two

row

s),a

fter

the

TM

Spuls

e.T

he

tem

pora

lch

anges

of

the

Outd

egre

efo

rth

enat

ura

lfr

equen

cyof

the

site

of

stim

ula

tio

n(i

.e.,

bfo

rsu

per

ior

par

ieta

l,b

2=c

for

pre

moto

r)(R

osa

nova

etal

.,2009)

(Fig

.2)

are

illu

stra

ted

atfo

ur

dif

fere

nt

tim

epoin

tsaf

ter

the

TM

Spuls

e.T

he

red

circ

les

coar

sely

indic

ate

the

targ

etar

eas

of

stim

ula

tion.

Agra

nd

aver

age

plo

tof

the

EE

G-r

econst

ruct

edso

urc

esis

also

show

nin

the

thir

dro

w.T

he

red

mar

ker

den

ote

sth

eti

me

inst

ant

along

the

ER

Pti

me

seri

esfo

rth

eas

soci

ated

fram

e(o

nto

p).

Az-

score

dm

apof

the

gro

up

aver

aged

SC

deg

ree

(fixed

,does

not

chan

ge

over

tim

e)is

dep

icte

don

the

right.

Thes

ebra

inim

ages

wer

eobta

ined

by

usi

ng

Bra

inN

etV

iew

er(X

iaet

al.,

2013).

ER

P,

even

tre

late

dpote

nti

al;

SC

,st

ruct

ura

lco

nnec

tivit

y.

92

which is also known to display a dominant frequency in thatfrequency range (Van Der Werf and Paus, 2006).

Second, our analysis permitted us to evaluate the dynamicinteractions between directed functional connectivity and an-atomical connectivity, before and after TMS, which was thesecond aim of our study.

We compared structural and directed functional connec-tivity at the whole network level for different EEG bands(a, b, b2=c, Fig. 5). The interplay between directed func-tional connectivity and structural connectivity at baseline isin line with findings reported in recent functional MRI-dMRI studies (Barttfeld et al., 2015), where the rich reper-toire of brain states does not necessarily correlate with thestructural pattern. Here, our directed functional connectivityapproach also allowed the investigation of systematic TMS-induced perturbations of the system, extending the insight onthe relationship between structure and function.

We observed a temporary decrease in the correlation be-tween directed connectivity and structural connectivityafter TMS. Assuming that each region in the brain reactsto the perturbation at a characteristic operating frequency,the decrease in structure–function correlation in each fre-quency (Fig. 5) might be caused by the complex patternof between-frequency interactions rising after TMS in the

whole-brain network. The return to baseline might then de-pend on two things: One is the temporal duration of the func-tional activation of the elicited area; the second is the extentto which it is related to its structural connectivity pattern.

These considerations brought us to explore the link be-tween the ‘‘natural’’ frequency response of the stimulatedcortical areas and their structural architecture. Interestingly,for the premotor cortex that is controlateral to the stimulationsite, the correlation between directed functional connectivityat the natural frequency and structural connections increasesafter the stimulation and reveals a long-lasting effect overtime (Fig. 6). The fact that this effect is not reproduced forthe superior parietal cortex might be due to a number of rea-sons. First, it has been shown that this area has lower corticalexcitability than the premotor cortex, and, thus, it is more dif-ficult to trigger (Ferrarelli et al., 2012; Rosanova et al.,2009). Second, it is possible that the different frequency re-sponses in each cortical area might reflect different anatom-ical backgrounds.

Indeed, recent studies have reported that there is a strongcorrelation between cytoarchitecture and anatomical andfunctional connectivity in cats, macaques, and humans(Beul et al., 2015; Scholtens et al., 2014), with the superiorparietal area showing both a different cytoarchitecture and

FIG. 5. Time-varying spatial correlation between directed functional connectivity and structural connectivity. Each plotshows the average over subjects of the dynamic spatial correlation between the directed functional connectivity (swADTF)matrices and the SC in function of time (blue line, standard error in shaded blue), for the three different frequency bands (a, b,b2=c) (Rosanova et al., 2009). The red line indicates the mean baseline value; the dashed lines represent 95% confidenceinterval of the empirical baseline distribution (see Materials and Methods section). Note the TMS-induced decrease in theobserved structure–function correlation, for both stimulation sites and in each frequency band.

TMS/EEG-DMRI INTERACTIONS IN THE HUMAN BRAIN 93

a different connectivity architecture than supplementarymotor regions (van den Heuvel et al., 2015). This might ex-plain why the functional activation at specific resonant fre-quencies is differently related to the structural coupling(i.e., the amount of tracts connecting them) depending onthe anatomical architecture of the specific brain region.

Limitations

Given the intrinsic limitations of the EEG in terms of spa-tial resolution, it is important to stress that the patterns ofconnectivity detected by TMS/hd-EEG are necessarilycoarse. Even though TEPs are characterized by a goodtest-retest reproducibility (Lioumis et al., 2009), the inter-individual reproducibility of the outgoing flow of informa-tion could be improved by a better computation of the elec-tric field induced by the TMS. More advanced models

(boundary, or finite, element models) could improve the ac-curacy of the source localization (Wagner et al., 2009).

Another limitation of our study concerns the relativelysmall sample size and the inter-subject variability at boththe TMS/hd-EEG response and the tractography level. How-ever, in this article, we investigated for the first time throughTMS-EEG and dwi the relationship between structural anddirected functional connectivity patterns in healthy subjects.The number of subjects included in the study, coupled withappropriate statistics, can definitely be of use for future re-search and hypothesis. Furthermore, the functional and struc-tural connectivity profiles obtained from our cohort appear tobe stable across subjects (Supplementary Fig. S4).

In addition, it has been shown that there are many brain re-gions with a complex structural architecture, which are alsoreferred to as ‘‘crossing fibers’’ ( Jeurissen et al., 2011; Tour-nier et al., 2012). In this context, tractography approaches

FIG. 6. Time-varying spatial correlation for the stimulated cortical regions. Each row shows the average over subjects ofthe dynamic spatial correlation (blue and green line, standard error in shaded blue and green) between the directed functionalconnectivity (swADTF) and SC for the AAL ROIs comprising left and right premotor areas (stimulated and controlateral, i.e.,ST SMA, CL SMA) and the left and right superior parietal areas (stimulated and controlateral, i.e., ST PCC, CL PCC), re-spectively, for the three different frequency bands (i.e., a, b, b2=c, respectively) (Rosanova et al., 2009). The continuous redline indicates the mean baseline value; the dashed lines represent 95% confidence interval of the empirical baseline distri-bution (see Materials and Methods section). Note the constant increase over time in the structure–function correlation forthe controlateral SMA after the TMS pulse, when taking into account its natural frequency (i.e., b2=c, Fig. 2).

94 AMICO ET AL.

based on more advanced diffusion models ( Jeurissen et al.,2011), or on more refined anatomical constraints (Smithet al., 2012) may provide more accurate anatomical connec-tivity patterns of brain networks. Therefore, our approachworks best for studying large-scale interactions than fine-scale, local dynamics.

Finally, a b-value of 1000 s=mm2 is lower than the optimalone for performing CSD, about 2500–3000 s=mm2 (Tournieret al., 2013). However, despite a low b-value, with a suffi-cient amount of directions, crossing fibers can be reliablymodeled with CSD and the result is still significantly betterthan with a simple DTI-based model; for example, seeRoine and coworkers (2015) for a successful application.

Conclusions

In conclusion, the findings of this study suggest that theway brain structure and function interact after TMS perturba-tion follows a rather complex and multifaceted dynamic.Notably, when looking at the whole-brain network level,the structure–function relationship is largely reduced afterTMS within each frequency band (Fig. 5). Nonetheless,when looking at the local responses of the target regions,we showed that the way directed functional connectivitychanges due to TMS might depend on both the frequencyat which the cortical module is elicited by TMS and the struc-tural architecture of the specific stimulated cortical region(Fig. 6). Future studies should validate this hypothesis by ex-ploring natural frequency profiles in different sites of stimu-lation and/or by evaluating structure–function correlations atnatural peak frequencies in brain regions other than the stim-ulated ones. An interesting follow-up of this study would beto investigate how the TMS pulse changes the functionalconnectivity profile of different stimulated sites over shortertime periods, that is, by using shorter time windows or a slid-ing window approach. In conclusion, our multimodal whole-brain approach gives new insights on how TMS interfereswith the brain network in healthy controls. More specifically,our study points out the importance of taking into account themajor role played by different cortical oscillations when in-vestigating the mechanisms for integration and segregationof information in the human brain (Casali et al., 2013).Another interesting follow-up of this study would, indeed,be to look at differences in structure–function interactions,either when the cognitive function is pharmacologicallymodulated (i.e., anesthesia) or after pathology, damage, ordisruption in structural connections (i.e., coma and disordersof consciousness).

Acknowledgments

The authors thank Timo Roine, Erik Ziegler, GianlucaFrasso, Andrea Piarulli, and Georgos Antonopoulos fortheir insightful discussion and comments on this article.They also thank Marie-Aurelie Bruno, Athena Demertzi,Audrey Vanhaudenhuyse, and Melanie Boly for help in ac-quiring the data. This research was supported by theWallonia-Brussels Federation of Concerted Research Action(ARC), Human Brain Project, Fonds National de laRecherche Scientifique de Belgique (FNRS), Belgian Sci-ence Policy (CEREBNET, BELSPO), McDonnell Founda-tion, European Space Agency, Mind Science Foundation,University Hospital, and University of Liege, Human Brain

Project (EU-H2020-fetflagship-hbp-sga1-ga720270) and Lumi-nous project (EU-H2020-fetopen-ga686764). O.B. is a researchfellow, O.G. is a post-doctoral fellow, and S.L. is a researchdirector at FNRS.

Author Disclosure Statement

No competing financial interests exist.

References

Arnold M, Milner XHR, Witte H, Bauer R, Braun C. 1998.Adaptive AR modeling of nonstationary time series bymeans of Kalman filtering. Biomed Eng IEEE Transact 45:553–562.

Astolfi L, Cincotti F, Mattia D, de Vico Fallani F, Tocci A, Colo-simo A, et al. 2008. Tracking the time-varying cortical con-nectivity patterns by adaptive multivariate estimators.Biomed Eng IEEE Transact 55:902–913.

Barttfeld P, Uhrig L, Sitt JD, Sigman M, Jarraya B, Dehaene S.2015. Signature of consciousness in the dynamics of resting-state brain activity. Proc Nat Acad Sci U S A 112:887–892.

Berg P, Scherg M. 1994. A fast method for forward computationof multiple-shell spherical head models. ElectroencephalogrClin Neurophysiol 90:58–64.

Beul SF, Grant S, Hilgetag CC. 2015. A predictive model of thecat cortical connectome based on cytoarchitecture and dis-tance. Brain Struct Funct 220:3167–3184.

Bonato C, Miniussi C, Rossini PM. 2006. Transcranial magneticstimulation and cortical evoked potentials: A TMS/EEG co-registration study. Clin Neurophysiol 117:1699–1707.

Bortoletto M, Veniero D, Thut G, Miniussi C. 2015. The contribu-tion of TMS—EEG coregistration in the exploration of the hu-man cortical connectome. Neurosci Biobehav Rev 49:114–124.

Bullmore E, Sporns O. 2009. Complex brain networks: Graphtheoretical analysis of structural and functional systems.Nat Rev Neurosci 10:186–198.

Caeyenberghs K, Leemans A, De Decker C, Heitger M, Drijko-ningen D, Linden CV, et al. 2012. Brain connectivity andpostural control in young traumatic brain injury patients: A dif-fusion MRI based network analysis. NeuroImage 1:106–115.

Casali AG, Casarotto S, Rosanova M, Mariotti M, Massimini M.2010. General indices to characterize the electrical responseof the cerebral cortex to TMS. Neuroimage 49:1459–1468.

Casali AG, Gosseries O, Rosanova M, Boly M, Sarasso S, CasaliKR, et al. 2013. A theoretically based index of consciousnessindependent of sensory processing and behavior. Sci TranslMed 5:198ra105.

Casarotto S, Lauro LJR, Bellina V, Casali AG, Rosanova M,Pigorini A, et al. 2010. EEG responses to TMS are sensitiveto changes in the perturbation parameters and repeatable overtime. PLoS One 5:e10281.

Catani M, Howard RJ, Pajevic S, Jones DK. 2002. Virtualin vivo interactive dissection of white matter fasciculi inthe human brain. Neuroimage 17:77–94.

Chu CJ, Tanaka N, Diaz J, Edlow BL, Wu O, Hamalainen M,et al. 2015. EEG functional connectivity is partially predictedby underlying white matter connectivity. Neuroimage 108:23–33.

Cona F, Zavaglia M, Massimini M, Rosanova M, Ursino M.2011. A neural mass model of interconnected regions simu-lates rhythm propagation observed via TMS-EEG. Neuro-image 57:1045–1058.

De Geeter N, Crevecoeur G, Leemans A, Dupre L. 2015. Effectiveelectric fields along realistic DTI-based neural trajectories for

TMS/EEG-DMRI INTERACTIONS IN THE HUMAN BRAIN 95

modelling the stimulation mechanisms of TMS. Phys Med Biol60:453–471.

Efron B, Tibshirani R. 1986. Bootstrap methods for standard er-rors, confidence intervals, and other measures of statisticalaccuracy. Stat Sci 1:54–75.

Ferrarelli F, Sarasso S, Guller Y, Riedner BA, Peterson MJ, Bel-lesi M, et al. 2012. Reduced natural oscillatory frequency offrontal thalamocortical circuits in schizophrenia. Arch GenPsychiatry 69:766–774.

Gosseries O, Sarasso S, Casarotto S, Boly M, Schnakers C,Napolitani M, et al. 2015. On the cerebral origin of EEG re-sponses to TMS: Insights from severe cortical lesions. BrainStimul 8:142–149.

Granger CW. 1969. Investigating causal relations by economet-ric models and cross-spectral methods. Econometrica 37:424–438.

Groppa S, Muthuraman M, Otto B, Deuschl, G, Siebner HR,Raethjen J. 2013. Subcortical substrates of TMS inducedmodulation of the cortico-cortical connectivity. Brain Stimul6:138–146.

Hagmann P, Cammoun L, Gigandet X, Meuli R, Honey CJ,Wedeen VJ, Sporns O. 2008. Mapping the structural coreof human cerebral cortex. PLoS Biol 6:e159.

Ilmoniemi RJ, Virtanen J, Ruohonen J, Karhu J, Aronen HJ,Katila T, et al. 1997. Neuronal responses to magnetic stimu-lation reveal cortical reactivity and connectivity. Neuroreport8:3537–3540.

Jeurissen B, Leemans A, Jones DK, Tournier J-D, Sijbers J.2011. Probabilistic fiber tracking using the residual bootstrapwith constrained spherical deconvolution. Hum Brain Mapp32:461–479.

Jones DK, Horsfield MA, Simmons A. 1999. Optimal strategiesfor measuring diffusion in anisotropic systems by magneticresonance imaging. Magn Reson Med 42:515–525.

Kaminski MJ, Blinowska KJ. 1991. A new method of the de-scription of the information flow in the brain structures.Biol Cybernet 65:203–210.

Leemans A, Jones DK. 2009. The B-matrix must be rotatedwhen correcting for subject motion in DTI data. MagnReson Med 61:1336–1349.

Lioumis P, Kicic D, Savolainen P, Makela JP, Kahkonen S.2009. Reproducibility of TMS—Evoked EEG responses.Hum Brain Mapp 30:1387–1396.

Liu Q, Balsters JH, Baechinger M, van der Groen O, WenderothN, Mantini D. 2015. Estimating a neutral reference for elec-troencephalographic recordings: The importance of using ahigh-density montage and a realistic head model. J NeuralEng 12:056012.

Massimini M, Ferrarelli F, Huber R, Esser SK, Singh H, TononiG. 2005. Breakdown of cortical effective connectivity duringsleep. Science 309:2228–2232.

Mattout J, Phillips C, Penny WD, Rugg MD, Friston KJ. 2006.MEG source localization under multiple constraints: An ex-tended Bayesian framework. Neuroimage 30:753–767.

Phillips C, Mattout J, Rugg MD, Maquet P, Friston KJ. 2005. Anempirical Bayesian solution to the source reconstructionproblem in EEG. Neuroimage 24:997–1011.

Rogasch NC, Fitzgerald PB. 2013. Assessing cortical networkproperties using TMS—EEG. Hum Brain Mapp 34:1652–1669.

Rogasch NC, Thomson RH, Daskalakis ZJ, Fitzgerald PB. 2013.Short-latency artifacts associated with concurrent TMS—EEG. Brain Stimul 6:868–876.

Roine U, Salmi J, Roine T, Nieminen-von Wendt T, LeppamakiS, Rintahaka P, et al. 2015. Constrained spherical deconvolution-

based tractography and tract-based spatial statistics show ab-normal microstructural organization in Asperger syndrome.Mol Autism 6:1–12.

Rosanova M, Casali A, Bellina V, Resta F, Mariotti M, Massi-mini M. 2009. Natural frequencies of human corticothalamiccircuits. J Neurosci 29:7679–7685.

Rosanova M, Gosseries O, Casarotto S, Boly M, Casali AG,Bruno MA, et al. 2012. Recovery of cortical effective con-nectivity and recovery of consciousness in vegetative pa-tients. Brain 135 (Pt 4):1308–1320.

Rubinov M, Sporns O. 2010. Complex network measures ofbrain connectivity: Uses and interpretations. Neuroimage52:1059–1069.

Scholtens LH, Schmidt R, de Reus MA, van den Heuvel MP.2014. Linking macroscale graph analytical organization tomicroscale neuroarchitectonics in the macaque connectome.J Neurosci 34:12192–12205.

Smith RE, Tournier J-D, Calamante F, Connelly A. 2012.Anatomically-constrained tractography: Improved diffusionMRI streamlines tractography through effective use of ana-tomical information. Neuroimage 62:1924–1938.

Smith SM, Jenkinson M, Woolrich MW, Beckmann CF, BehrensTEJ, Johansen-Berg H, et al. 2004. Advances in functional andstructural MR image analysis and implementation as FSL.Neuroimage 23:S208–S219.

Song S-K, Sun S-W, Ramsbottom MJ, Chang C, Russell J, CrossAH. 2002. Dysmyelination revealed through MRI as in-creased radial (but unchanged axial) diffusion of water. Neu-roimage 17:1429–1436.

Sporns O. 2011. Networks of the Brain. Cambridge: MIT Press.Sporns O, Tononi G, Kotter R. 2005. The human connectome: A

structural description of the human brain. PLoS Comput Biol1:e42.

Tournier J, Calamante F, Connelly A. 2012. MRtrix: Diffusiontractography in crossing fiber regions. Int J Imaging SystTechnol 22:53–66.

Tournier J, Calamante F, Connelly A. 2013. Determination ofthe appropriate b value and number of gradient directionsfor high-angular-resolution diffusion-weighted imaging.NMR Biomed 26:1775–1786.

Tournier J, Calamante F, Gadian DG, Connelly A. 2004. Directestimation of the fiber orientation density function fromdiffusion-weighted MRI data using spherical deconvolution.Neuroimage 23:1176–1185.

Tournier J, Yeh C-H, Calamante F, Cho K-H, Connelly A, LinC-P, et al. 2008. Resolving crossing fibres using constrainedspherical deconvolution: Validation using diffusion-weighted imaging phantom data. Neuroimage 42:617–625.

Tzourio-Mazoyer N, Landeau B, Papathanassiou D, Crivello F,Etard O, Delcroix N, et al. 2002. Automated anatomical la-beling of activations in SPM using a macroscopic anatomicalparcellation of the MNI MRI single-subject brain. Neuro-image 15:273–289.

van den Heuvel MP, Scholtens LH, Barrett LF, Hilgetag CC, deReus MA. 2015. Bridging cytoarchitectonics and connectom-ics in human cerebral cortex. J Neurosci 35:13943–13948.

Van Der Werf YD, Paus T. 2006. The neural response to trans-cranial magnetic stimulation of the human motor cortex. I.Intracortical and cortico-cortical contributions. Exp BrainRes 175:231–245.

Van Mierlo P, Carrette E, Hallez H, Raedt R, Meurs A, Vanden-berghe S, et al. 2013. Ictal-onset localization through connec-tivity analysis of intracranial EEG signals in patients withrefractory epilepsy. Epilepsia 54:1409–1418.

96 AMICO ET AL.

Van Mierlo P, Carrette E, Hallez H, Vonck K, Van Roost D, BoonP, Staelens S. 2011. Accurate epileptogenic focus localizationthrough time-variant functional connectivity analysis of intra-cranial electroencephalographic signals. Neuroimage 56:1122–1133.

Voineskos AN, Farzan F, Barr MS, Lobaugh NJ, Mulsant BH,Chen R, et al. 2010. The role of the corpus callosum in trans-cranial magnetic stimulation induced interhemispheric signalpropagation. Biol Psychiatry 68:825–831.

Wagner T, Rushmore J, Eden U, Valero-Cabre A. 2009. Bio-physical foundations underlying TMS: Setting the stage foran effective use of neurostimulation in the cognitive neuro-sciences. Cortex 45:1025–1034.

Wilke C, Ding L, He B. 2008. Estimation of time-varying con-nectivity patterns through the use of an adaptive directedtransfer function. Biomed Eng IEEE Transact 55:2557–2564.

Xia M, Wang J, He Y. 2013. BrainNet Viewer: A network visuali-zation tool for human brain connectomics. PLoS One 8:e68910.

Yao D. 2001. A method to standardize a reference of scalp EEGrecordings to a point at infinity. Physiol Meas 22:693–711.

Zalesky A, Akhlaghi H, Corben LA, Bradshaw JL, DelatyckiMB, Storey E, et al. 2014. Cerebello-cerebral connectivitydeficits in Friedreich ataxia. Brain Struct Funct 219:969–981.

Address correspondence to:Enrico Amico

Coma Science GroupCyclotron Research Center & GIGA Research Center

University and University Hospital of LiegeLiege 47907-2050

Belgium

E-mail: eamico@ulg.ac.be

TMS/EEG-DMRI INTERACTIONS IN THE HUMAN BRAIN 97