The future of fMRI in cognitive neuroscience

Download The future of fMRI in cognitive neuroscience

Post on 30-Nov-2016

214 views

Category:

Documents

2 download

TRANSCRIPT

  • ce

    Texa

    Available online 11 August 2011

    hanit

    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1218. . .and co. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1219. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1219. . . . . . . . . . . . . . . . . . . . . . . . . . .

    was unaware of the ongoing race to develop functional magnetic continues to be used in cognitive neuroscience, I think that the last few

    NeuroImage 62 (2012) 12161220

    Contents lists available at SciVerse ScienceDirect

    NeuroIm

    e lresonance imaging (fMRI). It may come as no surprise to those whoknow my writings that when I rst encountered imaging studies as agraduate student, I was highly skeptical that it could inform us aboutmental function. I was, in this sense, in good company; in those days, itwas not uncommon to hear conversations between cognitivepsychologists likening fMRI to trying to understand how an engineworks by measuring the temperature of the exhaust manifold. When Imoved to Stanford in 1995 to start my postdoc, I didn't actually go there

    years havewitnessed a number of encouraging newdevelopments, and Iremain very hopeful that fMRI will continue to provide us with usefulinsights into the relation between mind and brain. In the foregoing, Ioutline what I see as the the most promising new directions for fMRI incognitive neuroscience, with an obvious bias towards the some ofdirections that my own research is currently taking.

    Methodological rigor

    to do fMRI, but circumstances led me to getwork that was ongoing in John Gabrieli's lneuroimager was born.

    E-mail address: poldrack@mail.utexas.edu.

    1053-8119/$ see front matter 2011 Elsevier Inc. Aldoi:10.1016/j.neuroimage.2011.08.007ve psychology in 1989, Iwould be better characterized as hopes rather than predictions. Despitewhat I see as serious fundamental problems in how fMRI has been andWhen I began graduate school for cognitiDescribing mental processes: ontologiesConclusions . . . . . . . . . . . .Acknowledgments . . . . . . . . .References . . . . . . . . . . . . .

    Introduction Mythoughtsbelowabout the futureof fMRI in cognitiveneuroscienceinvolved in the imagingab, and my career as a Foremost, I h

    neuroscience wiaging methods.correlations (Khighlighted theConsideration o

    l rights reserved.. . . . . . . . . . . . . . . . . . . . . . . . 1219Open science and data aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1218mputational models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1219From blobology to pattern analysis . . . . . . . . . . . . . . . . . .Selective inference . . . . . . . . . . . . . . . . . . . . . . . . . .Methodological rigor . . . . . . . . . . . . . . . . . . . . . . . . .Keywords:fMRICognitive neuroscienceMeta-analysisOntologyConnectivity

    increased methodological rigor, an increasing focus on connectivity and pattern analysis as opposed toblobology, a greater focus on selective inference powered by open databases, and increased use ofontologies and computational models to describe underlying processes.

    2011 Elsevier Inc. All rights reserved.

    Contents

    Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1216. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1216. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1217Review

    The future of fMRI in cognitive neuroscien

    Russell A. PoldrackDepartments of Psychology and Neurobiology and Imaging Research Center, University of

    a b s t r a c ta r t i c l e i n f o

    Article history:Accepted 5 August 2011

    Over the last 20 years, fMRInext 20 years of fMRI in cog

    j ourna l homepage: www.s, Austin, TX 78759, USA

    s revolutionized cognitive neuroscience. Here I outline a vision for what theive neuroscience might look like. Some developments that I hope for include

    age

    sev ie r.com/ locate /yn imgope that in the next 20 years the eld of cognitivell increase the rigor with which it applies neuroim-The recent debates about circularity and voodooriegeskorte et al., in press; Vul et al., 2009) haveneed for increased care regarding analytic methods.f similar debates in genetics and clinical trials led

  • (Ioannidis, 2005) to outline a number of factors thatmay contribute toincreased levels of spurious results in any scientic eld, and thedegree to which many of these apply to fMRI research is rathersobering:

    small sample sizes small effect sizes large number of tested effects exibilty in designs, denitions, outcomes, and analysis methods being a hot scientic eld

    Some simple methodological improvements could make a bigdifference. First, the eld needs to agree that inference based onuncorrected statistical results is not acceptable (cf. Bennett et al., 2009).Many researchers have digested this important fact, but it is stillcommon to see results presented at thresholds such as uncorrectedpb .005. Because such uncorrected thresholds do not adapt to the data(e.g., the number of voxels tests or their spatial smoothness), they arecertain to be invalid in almost every situation (potentially being eitheroverly liberal or overly conservative). As an example, I took the fMRIdata from Tom et al. (2007), and created a random individualdifference variable. Thus, there should be no correlations observedother than Type I errors. However, thresholding at uncorrected pb .001andaminimumcluster size of 25 voxels (a commonheuristic threshold)

    legitimate, but one often gets the feeling that the masks used for smallvolume correction were chosen after seeing the initial results(perhaps after a whole-brain corrected analysis was not signicant).In such a case, any inferences based on these corrections are circular andthe statistics are useless. Researchers who plan to use small volumecorrections in their analysis should formulate a specic analysis planprior to any analyses, and only use small volume corrections that wereexplicitly planned a priori. This sounds like a remedial lesson in basicstatistics, but unfortunately it seems to be regularly forgotten byresearchers in the eld.

    Third, the eld needs to move toward the use of more robustmethods for statistical inference (e.g., Huber, 2004). In particular,analyses of correlations between activation and behavior acrosssubjects are highly susceptible to the inuence of outlier subjects,especially with small sample sizes. Robust statistical methods canensure that the results are not overly inuenced by these outliers,either by reducing the effect of outlier datapoints (e.g., robustregression using iteratively reweighted least squares) or by separatelymodeling data points that fall too far outside of the rest of the sample(e.g., mixture modeling). Robust tools for fMRI group analysis areincreasingly available, both as part of standard software packages(such as the outlier detection technique implemented in FSL:Woolrich, 2008) and as add-on toolboxes (Wager et al., 2005). Given

    2.5

    2

    1.5

    1

    0.5

    0

    0.5

    -1

    1.5

    -2

    on. Ttedcan

    1217R.A. Poldrack / NeuroImage 62 (2012) 12161220showed a signicant region near the amygdala; Fig. 1 shows this regionalongwith a plot of the beautiful (but artifactual) correlation betweenactivation and the random behavioral variable. This activation was notpresent when using a corrected statistic. A similar point was made in amore humorous way by Bennett et al. (2010), who scanned a deadsalmonbeingpresentedwith a social cognition task and foundactivationwhen using an uncorrected threshold. There are now a number of well-established methods for multiple comparisons correction (Poldrack etal., 2011), such that there is absolutely no excuse to present results atuncorrected thresholds. The most common reason for failing to userigorous corrections formultiple tests is that with smaller samples thesemethods are highly conservative, and thus result in a high rate of falsenegatives. This is certainly a problem, but I don't think that the answer isto present uncorrected results; rather, the answer is to ensure that one'ssample is large enough to provide sufcient statistical power to nd theeffects of interest.

    Second, I have become increasingly concerned about the use ofsmall volume corrections to address the multiple testing problem.The use of a priori masks to constrain statistical testing is perfectly

    rando

    m c

    ovar

    iate

    -

    -

    Fig. 1. An example of artifactual activation that survives a heuristic statistical correcti25 voxel extent threshold) for the correlation between activation and a randomly generathe map and the randomly generated variable, showing that completely spurious data

    circular data analysis.the frequency with which outliers are observed in group fMRI data,these methods should become standard in the eld. However, it's alsoimportant to remember that they are not a panacea, and that itremains important to apply sufcient quality control to statisticalresults, in order to understand the degree to which one's resultsreect generalizeable patterns versus statistical gments.

    From blobology to pattern analysis

    The foregoing methodological issues have taken on particularimportance due to the strong focus of the eld on localization offunction; the goal of nding blobs in a specic region can driveresearchers into analytic gymnastics in order to nd a signicant blobto report. However, for the last few years the most interesting andnovel research has focused on understanding patterns of activationrather than localized blobs. The appreciation of patterns is happeningat multiple scales. At the systems (whole-brain) scale, themodeling ofconnectivity and its relation to behavior continue to grow. Oneimpediment to the analysis of connectivity with fMRI is the lack of a

    -150 -100 -50 0fmri signal

    50 100 150 200

    r=0.781

    he left panel shows the thresholded activation map (pb .001, voxelwise uncorrected,variable. The right plot shows the correlation between activation in the peak voxel fromproduce seemingly compelling results when heuristic corrections are combined with

  • 1218 R.A. Poldrack / NeuroImage 62 (2012) 12161220standardized set of approaches for connectivity modeling; with somany different approaches being used across the eld, it can bedifcult to understand how the results relate to one another. Therecent work by Smith et al. (2011) examining the relative power ofdifferent modeling approaches using simulated data is a good starttoward better understanding the relative strengths andweaknesses ofthese different methods, and I envision that these kinds ofmethodological shoot-outs will eventually help determine whichmethods are most optimal for which questions. I think that the jury isstill out on how well fMRI can ever characterize neuronal connectiv-ity; as we outlined in Ramsey et al. (2010), there are a number offundamental challenges in using fMRI to characterize causal in-teractions between brain regions. Further work combining neuro-physiology and fMRI (e.g., David et al., 2008) will be essential to fullyunderstand the potential and limitations of fMRI for modeling of theunderlying neuronal connectivity.

    At a more local level, the use of analyses that focus on patterns ofactivation over a region rather thanmean activation in that region hasbecome very popular in the last few years (e.g., Kriegeskorte et al.,2006, 2008; Norman et al., 2006), and I expect that this will continue.It is now clear that information can be coded in fMRI data eitherthrough local changes in mean activation or through changes in localpatterns of activity, and that analysis of only the mean activation(which has been the standard over the last 20 years) may bemissing asignicant part of the picture. For example, we recently comparedresults from closely-matched univariate and multivariate (pattern-information) analyses for a decision-making task, and found manyregions in the cortex where the multivariate test was signicantlymore sensitive to task-relevant information than the univariateanalysis, with a smaller set of regions showing the opposite pattern(Jimura and Poldrack, in press). I think that one approach that hasparticularly strong appeal is representational similarity analysis(Kriegeskorte et al., 2008), in which the patterns of activity in fMRIdata are related to behaviorally-relevant variables such as stimulusfeatures or task manipulations. For example, we (Xue et al., 2010)recently used this approach to show that memory is better for itemswhose patterns of activity are more similar across study episodes,which has implications for cognitive theories of memory. I think thatthis approach has great potential to provide much stronger links fromneuroimaging data to cognitive theories (which often make detailedpredictions about patterns across stimuli) and computational models,which will help provide a stronger basis for relating brain activity andmental function.

    Selective inference

    Neuroimaging as it is used by most cognitive neuroscientists facesa fundamental problem, which is evident from even a cursory readingof the literature: Within different subelds in cognitive neuroscience,very different functions can be assigned to the same structure. Theanterior cingulate may be the poster child for this heterogeneity, withproposed functions as various as conict monitoring, interoception,pain, autonomic regulation, effort, and consciousness. This becomesparticularly problematic when researchers are aware of only one ofthose literatures and attempt to make reverse inferences based onactivation in a region (e.g., the anterior cingulate is active, thus thesubject must be experiencing conict) (Poldrack, 2006).

    If the goal of cognitive neuroscience is to map function ontostructure, then this is a serious problem, and the standard approach toneuroimaging does not provide any way to solve it. How can weidentify what the basic function is that a region or network subserves?As I argued recently (Poldrack, 2010), I think that the answer is tomove toward methods that allow us to identify selective associationsbetween functions and structures. This requires that we ask notsimply whether we can nd a region that is engaged by a particular

    mental process, but whether one can nd a region that is engagedselectively, such that activation of the region is actually predictive ofthe mental process. The tools from machine learning that haverecently been brought to bear on fMRI data (Norman et al., 2006;Pereira et al., 2009; Poldrack et al., 2009) can provide importantleverage on this issue, by providing a principled way to assess thepredictive power of activation and thus allowing a quantitativeassessment of the strength of any reverse inference (Poldrack, 2008).

    In addition to these tools, however, we need data sets that arebroad enough to allow the assessment of any particular functionagainst many other functions in order to assess sele...

Recommended

View more >