Dale-Dynamic Statistical Parametric NeurotechniqueMapping-Combining fMRI and MEG for High-Resolution Imaging of Cortical Activity

 Documents

 53 views
of 13

Please download to get full document.

View again

All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Description
Neuron, Vol. 26, 55–67, April, 2000, Copyright ©2000 by Cell Press Dynamic Statistical Parametric Neurotechnique Mapping: Combining fMRI and MEG for High-Resolution Imaging of Cortical Activity Anders M. Dale,* Arthur K. Liu,* Bruce R. Fischl,* Randy L. Buckner,§ John W. Belliveau,* Jeffrey D. Lewine,† and Eric Halgren*†‡ * Massachusetts General Hospital Nuclear Magnetic Resonance Center Charlestown, Massachusetts 02129 † Department of Radiology University of Utah Salt Lake City, Utah 84108 ‡ I
Share
Tags
Transcript
  Neuron, Vol. 26, 55–67, April, 2000, Copyright © 2000 by Cell Press Dynamic StatisticalParametric NeurotechniqueMapping:CombiningfMRIandMEGforHigh-ResolutionImagingofCorticalActivity exist that measure changes in these hemodynamic andelectromagnetic signals. Considered individually, thesetechniques offer trade-offs between spatial and tempo-ral resolution (Churchland and Sejnowski, 1988). Hemo-dynamic assessment of brain activity is temporally lim- AndersM.Dale,*  ArthurK.Liu,*BruceR.Fischl,*RandyL.Buckner, §  JohnW.Belliveau,* JeffreyD.Lewine, † andEric Halgren* †‡ *Massachusetts General Hospital NuclearMagnetic Resonance CenterCharlestown, Massachusetts 02129 ited by thelatencyof thehemodynamic response(about1 s) but can provide millimeter spatial sampling (Belli- † Department of RadiologyUniversity of Utah veau et al., 1991; Kwong et al., 1992). Conversely, meth-ods based on direct measurement of the electric andSalt Lake City, Utah 84108 ‡ Institut National de la Sante´et magnetic fields produced by neuronal activity can pro-vide temporal resolution of less than 1 ms, adequate forde la Recherche Me´dicaleE9926, Marseilles detecting theorchestrationofcomplex cognitiveactivity(Regan,1989).However,the spatial  configurationofneu-France § Department of Psychology Anatomy, and ronal activity cannot be derived uniquely based on elec-troencephalography (EEG)and/or magnetoencephalog-Neurobiology, and RadiologyWashington University raphy(MEG)recordingsalone(Nunez,1981;Ha¨ma¨la¨inenet al., 1993). In order to make this so-called inverse  St. Louis, Missouri 63130 problem  wellposed,it is necessaryto imposeadditionalconstraints on the solution.One common approach is to assume that the EEG/  Summary MEG signals are generated by a relatively small numberof focal sources (Sherg and VonCramon, 1985; Schmidt Functional magnetic resonance imaging (fMRI) can et al., 1999). An additional constraint can be derived providemapsofbrainactivationwithmillimeterspatial from the assumption that the sources are temporally resolutionbutis limitedinits temporal resolutionto uncorrelated (Mosher et al., 1992). These assumptions theorderofseconds.Here,wedescribeatechnique are particularly appropriate when analyzing early sen- thatcombinesstructuralandfunctionalMRIwithmag- sory responses, where the activity might reasonably be netoencephalography(MEG)toobtainspatiotemporal expected to be relatively focal and constrained to a mapsofhumanbrainactivitywithmillisecondtempo- few primary sensory areas. On the other hand, such ralresolution.Thisnewtechniquewasusedtoobtain assumptions are less justified in higher-level cognitive dynamicstatisticalparametricmapsofcorticalactiv- experiments, which have been found by intracranial re- ityduringsemantic processing ofvisuallypresented cordings in humans to involve extensive networks of words.Aninitialwaveofactivitywasfoundtospread more or less synchronously activated brain areas (Hal- rapidlyfromoccipitalvisualcortextotemporal,pari- gren et al., 1994a, 1994b, 1995a, 1995b; Baudena et al., etal,andfrontalareaswithin185ms,withahighde- 1995). Similarly, the interictal spikes characteristic of greeoftemporaloverlapbetweendifferentareas.Rep- partial epilepsy typically spread very rapidly to involve etition effects were observed in many of the same a network extended across multiple cortical and limbic areasfollowingthisinitialwaveofactivation,providing regions (Chauvel et al., 1987). evidencefortheinvolvementoffeedbackmechanisms An alternative approach to analyzing EEG/MEG sig- inrepetitionpriming. nals is to impose constraints based on anatomical andphysiological information derived from other imaging Introduction modalities. The anatomical constraint is based on theobservation that the main cortical generators of EEGThorough understanding of the functional organizationand MEG signals are localized to the gray matter andof the brain requires knowledge of several aspects oforiented perpendicularly to the cortical sheet (Nunez,functional neuroanatomy, including the specific loca-1981).Thus,oncetheexact shapeof thecorticalsurfacetions of processing areas, the type of processing per-is known (for a specific subject), this information can beformed, the time course of processing, and the natureused to greatly reduce the EEG/MEG solution spaceof the interactions between these areas. Local alter-(Dale and Sereno, 1993).ations in neuronal activity induce local changes in theThe solution space can be further reduced by makingelectric and magnetic fields (Ha¨ma¨la¨inen et al., 1993;use of information derived from metabolic or hemody-Mitzdorf, 1985), cerebral metabolism, and cerebral per-namic measures of brain activity during the same taskfusion (blood flow, blood volume, and blood oxygen-(Nenov et al., 1991; Dale and Sereno, 1993; Heinze etation)(Mazziottaet al.,1983; Fox and Raichle,1986;Foxal., 1994; Snyder et al., 1995; Liu et al., 1998; Mangunet al., 1988; Belliveau et al., 1991; Prichard et al., 1991;et al., 1998; Ahlfors et al., 1999). This is based on theKwong et al., 1992). Several noninvasive techniqueshypothesis that the synaptic currents generating theEEG/MEG signals also impose metabolic demands,which in turn lead to a hemodynamic response measur-  To whom correspondence should be addressed (e-mail: dale@nmr.mgh.harvard.edu). ableusingpositronemissiontomography(PET)(Raichle,  Neuron56 1987) or functional magnetic resonance imaging (fMRI) ResultsandDiscussion (Belliveau et al., 1991; Kwong et al., 1992). Although the EstimationofSpatiotemporalActivityPatterns precisenatureof thecoupling betweenneuronal activityThe goal of the method described here is to obtainandhemodynamicsignalsisunknown,thereisconsider-estimates of brain electrical activity with the best possi-able evidence for a strong correlation between the spa-ble spatial and temporal accuracy. Given the inherenttial patterns of hemodynamic changes and neuronaldifferencesinthespatialand temporalresolutionofnon-electrical activity over time in both animals (Grinvald etinvasive imaging modalities, as discussed above, weal., 1986) and humans (Benson et al., 1996; Puce et al.,wish to combine the different measures in a way that1997).takes maximal advantage of the strengths of each tech-Here we present a general framework for integratingnique. In other words, we wish to obtain spatiotemporalinformation from different imaging modalities with a pri-activity estimates that are maximally consistent with allori anatomical and physiological information to produceavailable observables (i.e., fMRI, EEG, and/or MEG) asspatiotemporal estimates of brain activity. By normaliz-wellas apriori anatomicaland physiologicalinformationing these estimates in terms of noise sensitivity at each(for a more formal discussion, see Experimental Proce-spatial location, we obtain statistical parametric mapsdures below, and Liu, 2000). This, of course, requires(SPMs)that provide information about the statistical re-that we have some idea of the coupling between electri-liability of the estimated signal at each location in thecalactivityinthebrainand ournoninvasiveobservables.map with millisecond accuracy. These SPMs can thenInthecaseof EEGand MEG,this isrelativelystraightfor-be visualized as movies of brain activity (dynamicward, as the electric and magnetic fields generated bySPMs). In order to quantitate the spatial resolution ofneuronal activity follows from well understood funda-these maps, we compute the pointspread function formental laws of physics (Nunez, 1981; Ha¨ma¨la¨inen et al.,different locations on the cortical surface. This reflects1993).the spatial “blurring”of the true activity patterns in ourThe primary generators of EEGand MEGare synapticspatiotemporal maps or movies.currents, where the current flows crossing neuronalThis new method is applied to MEG and fMRI mea-membranes act as tiny current sources or sinks for cur-surements in a task involving semantic judgments ofrent outflow and inflow, respectively. Note that for eachvisually presented words. Such tasks are known to in-neuron, the net current inflow and outflow through itsvolve a large number of cortical areas (Buckner andmembrane has to be zero (for conservation of charge).Koutstaal,1998;Fiezand Petersen,1998),but thetimingThus, in order for these fields not to cancel out at theoftheirinvolvementhasnotpreviouslybeendetermined.noninvasive sensors, there has to be some net spatialThe putative roles assigned to these structures suggestseparationbetweenthecurrent sourcesand sinkswithina possible sequence of activation during this task, withtheneurons.Ofcourse,thefieldsproduced byindividualvisual word form processing preceding semantic asso-neurons are far too weak to be observed noninvasivelyciative activation, which in turn would precede workingby EEGor MEG. Thus, to generate externally detectablememory and response mapping. It is unknown if, in fact,signals, the neurons within a volume of tissue must beprocessing occurs according to this sequence. Morealigned and their synaptic current flows correlated intime.Thescalp-recorded EEGand MEGreflect thelineargenerally, the degree to which processing is sequentialsuperposition of the fields generated by all such synap-and modular in specialized regions versus parallel andtic currents across all neurons.distributed is unknown.Of all the neurons in the human brain, the corticalIn addition, we investigated the effect of item repeti-pyramidal cells are particularly well suited to generatetion on the spatiotemporal activity patterns evoked byexternally observable electric and magnetic fields due tothe task. Repetition is also known to alter the activationtheir elongated apical dendrites, systematically aligned inevoked in several cortical regions (Buckner and Kouts-a columnar fashion perpendicular to the cortical sheet.taal, 1998; Gabrieli et al., 1998), but again, the timingInhibitory and excitatory synaptic inputs from differentof these effects is not known. One hypothesis wouldcell populations have characteristic laminar distribu-suggest that processing of repeated items is facilitatedtions, resulting in characteristic spatial and temporalbeginning with relatively early perceptual stages. An al-patternsof net synaptic current flowsat different depthsternative hypothesis is that repetition effects are medi-through the cortical sheet (Nicholson and Freeman,ated by a top-down mechanism, srcinating in higher-1975; Mitzdorf, 1985; Barth and Di, 1991; Schroederlevel brain areas. One way to distinguish between theseet al., 1995). These current flows are typically stronglyhypotheses would be to know the order in which differ-correlated laterally along the cortical sheet (Sukov andent anatomical areas show the effects of repetition. TheBarth, 1998). Since the thickness of the cortical sheetarea with the earliest changes might then be reasonablyis much smaller than the distance to the EEG and MEGsupposed to help produce the later changes noted insensors, the current source/sink distribution within aother areas. Notethat an observed decreasein metabo-small slab of cortex can be represented by a currentlism in cortical areas that are “early” in the anatomicaldipole oriented perpendicularly to the local cortical sur-sense (e.g., sensory-perceptual areas) might nonethe-face (Dale and Sereno, 1993), whose strength (moment)less be due to a change that occurs late in time (i.e.,varies with time. Thus, we can represent the net corticallate in the processing stream) due to top-down effectssynaptic current flows by a scalar function s  ( r , t  ), re-from higher areas. Thus, the actual timing, as well asflecting dipole strength as a function of location r andthe location of activation, is crucial for understandingtime t  . The coupling between the dipole strengths andthe observed electric and magnetic fields can then bethe functional dynamics of cognitive processing.  Dynamic Statistical Parametric Mapping57 expressed simply in terms of a sum or integral across Due to the linear nature of the inverse estimation ap-all spatial locations (equations 2 and 3 in Experimentalproach considered here, we can obtain a straightforwardProcedures, below).expression for the uncertainty in our activity estimatesThe coupling between electrical activity and hemody-due to noise in the EEG/MEG recordings (equation 5).namic measures likefMRI is less understood.In particu-By normalizing the activity estimates at each locationlar,thereis littlequantitativedataon how the magnitude  in the brain by the noise sensitivity (the standard errorof the hemodynamic response varies as a function ofof the estimate), we obtain statistical parametric mapsthe amplitude and duration of electrical activity. Thereof activity at every timepoint, with millisecond accuracyis, however, considerable evidence for a strong degree(equations 6 and 7). These dynamic statistical paramet-of spatial  correlation between various measures of localric maps can be displayed as movies of brain activityelectrical activity and local hemodynamic signals.Someovertime,with eachframeinthesemovies indicating if aof the strongest evidence for this comes from a directstatisticallysignificantsignalispresentat aeachcorticalcomparison of maps obtained using voltage-sensitivelocation at that latency.dyes, reflecting depolarization of neuronal membranesOne common and straightforward way to quantitatein superficial cortical layers, and maps derived fromthespatialresolution,ordegreeof “blurring,”at differentintrinsic optical signals, reflecting changes in light ab-locations in the maps is by calculating the spatial point-sorption due to changes in blood volume and oxygen-spread function. This function is simply the image oneation(Shohamet al.,1999).Previousanimalstudieshavewould obtain if all the activity were concentrated at aalso shownstrong correlationbetweenlocalfield poten-given point,assuming no noise in theEEG/MEGsignals.tials, spiking activity, and voltage-sensitive dye signalsDue to the time-invariant and spatially linear nature of(Arieli et al., 1996; Tsodyks et al., 1999). Furthermore,our estimation method, one can derive a simple expres-studies in humans comparing the localization of func-sion for the pointspread at every location in the braintional activity using invasive electrical recordings and(equation 8, below). Moreover, the spatial blurring offMRI also provide evidence of correlation between theany arbitrary spatiotemporal activation pattern can belocal electrophysiological and hemodynamic responsepredicted as the activation-weighted linear superposi-(Benson et al., 1996; Puce et al., 1997). This suggeststion (sum)of the pointspread for each location and timethat the local fMRI response can be used to bias thepoint.electrical activity estimate toward those regions thatFigure 1 shows the pointspread function for three dif-show the greatest fMRI response. This can be accom-ferent locations (indicated by green circles), when theplished by using the fMRI response as an a priori esti-electrical activity is constrained to lie on the corticalmate of the locally integrated dipole activity over timesurface. The left side shows the pointspread functions(Daleand Sereno,1993;Liuet al.,1998).Thisformulationfor the commonly used minimum norm solution (Ha¨ma¨-resultsinastraightforward linearexpressionfortheopti-la¨inen and Ilmoniemi, 1984), calculated assuming 122malestimatebased ontheEEGand/orMEGdataat eachchannels of MEG recordings. This illustrates a commontime point (equation 4, below). Our previous simulationproblem with distributed dipole solutions, namely theirstudies suggest that some care must be taken to avoidtendency to misattribute focal, deep activations to ex-overconstraining the solution, as minor mismatches be-tended,superficial patterns (Ha¨ma¨la¨inen and Ilmoniemi,tween the electrical and hemodynamic signals can then1984; Dale and Sereno, 1993). The first point (top), lo-severely distort the resulting estimates (Liu et al., 1998;cated deep in the insula, has a particularly large pointLiu, 2000). These studies further suggest that by usingspread, covering more than half the lateral extent ofthe fMRI data as a partial constraint (i.e., allowing somethe brain. The second point (middle), also located in aelectrical activity in locations with no detectable fMRIsulcus, has a somewhat smaller pointspread, but thereresponse), one can obtain accurate estimates of electri-is a pronounced bias toward superficial locations, ascal activity from MEG and EEG, even in the presenceevidenced by the shift in center-of-mass away from theof some spatial mismatch between the generators ofactual location. The third point (bottom), located moreEEG/MEG data and the fMRI signals.superficially near the crown of a gyrus, has a muchsmaller pointspread. Note that the pointspread for the UncertaintiesandPotentialErrors noise-normalized anatomically constrained estimates, intheActivityEstimates shown on the right in Figure 1, is of much more uniformIt isimportant to notethat estimatesof brainstructureorextent for the different locations. (See Liu, 2000 for afunction based on any noninvasive imaging technology,more extensive analysis of the pointspread propertieswhether based on MRI, CT, PET, SPECT, or EEG/MEG,of different estimation approaches.)are necessarily inexact. This imprecision is caused byThispointisfurtherillustrated inFigure2,whichshowsseveral factors, including (1) noise in the measured sig-pseudo-color maps and histograms of pointspread ex-nals, resulting in errors in the estimates; (2) errors intent for all locations on the cortical surface. Again, theour model of the coupling between the parameters ofpointspread extent (in terms of half-width-half-max) forinterest (e.g., electrical activity, tissue perfusion, bloodtheminimumnormestimatorvariesgreatlyacrossdiffer-flow, or oxygenation)and the measured signals; and (3)ent locations on the cortical surface, reaching as muchfundamental ill-posedness or ambiguity of the inverseas 100 mm or more in certain locations (shown in brightproblem, resulting in limited resolution in the estimates.yellow). The noise-normalized estimator, on the otherThus, in order to properly interpret noninvasive imaginghand, results in a much more spatially uniform point-results, it is essential to have quantitative measures ofspread extent, with values averaging around 20 mm. Atheuncertaintiesand potentialerrorsdueto thesediffer-ent factors. further reduction in pointspread extent can be achieved  Neuron58Figure 1. Spatial Resolution of the Anatomi-cally Constrained Estimates for Three Mod-eled Dipole LocationsThe spatial resolution of the described noise-normalized method is contrasted with that ofthe standard minimum norm approach (rightand left columns, respectively). Shown arepointspread maps for three different dipolelocations (indicated by green dots). Note thatthe pointspread maps for the deep dipole lo-cations (first and second rows) are muchmore extensive with the minimum norm thanwith the noise-normalized estimator. Con-versely, the pointspread map for the superfi-cial dipole location (third row) is somewhatmorefocal with the minimum norm estimator.In all cases, estimates were calculated con-straining activation to the cortical surface. byincreasingthenumberofMEGsensorsand byinclud- activity evoked during a task involving semantic pro-cessing of visually presented words. Subjects were re-ing EEG measurements in the estimation (Liu, 2000).Note that due to the linear nature of the estimation quiredtodecidewhethereachwordreferredtoanobjector animal that is usually more than one foot in sizeapproach, such pointspread maps provide a directquantitation of the local degree of spatial blurring ex- (in any dimension). The words were either novel (i.e.,presented only once in the task)or were repeated multi-pected in the spatiotemporal maps calculated with thismethod for arbitrary activation patterns. Importantly, ple times.these pointspread functions can be used to assess thespatial accuracy of the anatomically constrained esti- AnatomicallyConstrainedEstimatesBased mates obtained in cognitive experiments, as described onMEG(aMEG)inaSingleSubject below.Applying the anatomically constrained noise-normal-ized estimation procedure to the event-related MEGav- SpatiotemporalMappingofBrainActivity erages, we computed spatiotemporal activity estimates inaCognitiveTask  for the novel and repeated word conditions. SnapshotsIn the following, we describe the results of applying themethods described above to spatiotemporally map the of these aMEG “movies”are shown in Figure 3 for four Figure 2. Spatial Resolution of the Anatomi-cally Constrained Estimates for all CorticalLocationsMaps and histograms of pointspread extent,in terms of half-width-half-maximum in milli-meters on the cortical surface, are shown forthe minimum norm and noise-normalized es-timators (left and right side, respectively).Note that the noise-normalized estimator re-sults in much more spatially uniform andoverall lower pointspread extent than theminimum norm estimator, particularly fordeep (sulcal) locations.
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks