FMRI Experimental Design
FMRI Experimental Design7/10/2014YingyingMRI studies brain anatomy.
Functional MRI (fMRI) studies brain function.MRI vs. fMRI22
Slice Thicknesse.g., 6 mm
Number of Slicese.g., 10SAGITTAL SLICE
IN-PLANE SLICEField of View (FOV)e.g., 19.2 cmVOXEL(Volumetric Pixel)3 mm3 mm6 mmSlice TerminologyMatrix Sizee.g., 64 x 64In-plane resolutione.g., 192 mm / 64= 3 mm334
Coordinates - Anatomy3 Common Views of Brain:Coronal (head on)Axial (birds eye), aka Transverse. Sagittal (profile)sagittalcoronalaxialOverview of SPM
5OverviewCategorical designsSubtraction - Pure insertion, evoked/differential responsesConjunction - Testing multiple hypothesesParametric designsLinear - Adaptation, cognitive dimensionsNonlinear - Polynomial expansions, neurometric functionsFactorial designsCategorical - Interactions and pure insertionParametric - Linear and nonlinear interactions - Psychophysiological interactions
6Cognitive subtractionAimNeural structures underlying a single process Y (e.g. face recognition)?Procedure:Contrast: [Task with Y] [control task without Y] = Y the critical assumption of pure insertionExample:[Task with Y] [task without Y] = Y7Cognitive subtractionReality:
8Cognitive subtraction: Baseline problemsWhich neuronal structures support face recognition?
Distant stimuli
Several components differ!Related stimuli
Y implicit in control condition?Same stimuli, different task
Interaction of task and stimuli (i.e. do task differences depend on stimuli chosen)?
Name Person! Name Gender!President? Aunt Yingying
9Evoked responses
10Differential responses
11A categorical analysis
12Categorical design
13OverviewCategorical designsSubtraction - Pure insertion, evoked/differential responsesConjunction - Testing multiple hypothesesParametric designsLinear - Adaptation, cognitive dimensionsNonlinear - Polynomial expansions, neurometric functionsFactorial designsCategorical - Interactions and pure insertionParametric - Linear and nonlinear interactions - Psychophysiological interactions
14ConjunctionsOne way to minimize the baseline/pure insertion problem is to isolate the same process by two or more separate comparisons, and inspect the resulting simple effects for commonalities
A test for such activation common to several independent contrasts is called conjunction
Conjunctions can be conducted across a whole variety of different contexts: tasks stimuli senses (vision, audition) etc.
Note: the contrasts entering a conjunction must be orthogonal !15non-overlapping, uncorrelated, or independent15Conjunctions16
17Test of global null hypothesis: Significant set of consistent effectsWhich voxels show effects of similar direction (but not necessarily individual significance) across contrasts?Null hypothesis: No contrast is significant: k = 0does not correspond to a logical AND !Test of conjunction null hypothesis: Set of consistently significant effectsWhich voxels show, for each specified contrast, significant effects?Null hypothesis: Not all contrasts are significant: k < ncorresponds to a logical AND A1-A2 B1-B2p(A1-A2) < +p(B1-B2) < +Friston et al. (2005). Neuroimage, 25:661-667.Nichols et al. (2005). Neuroimage, 25:653-660.Two types of conjunctions18SPM offers both conjunctions
Friston et al. 2005, Neuroimage, 25:661-667.specificitysensitivityGlobal null:k = 0Conjunction null:k < n19F-test vs. conjunction based on global null
Friston et al. 2005, Neuroimage, 25:661-667.20Using the conjunction null is easy to interpret, but can be very conservative
Friston et al. 2005, Neuroimage, 25:661-667.21OverviewCategorical designsSubtraction - Pure insertion, evoked/differential responsesConjunction - Testing multiple hypothesesParametric designsLinear - Adaptation, cognitive dimensionsNonlinear - Polynomial expansions, neurometric functionsFactorial designsCategorical - Interactions and pure insertionParametric - Linear and nonlinear interactions - Psychophysiological interactions
22Parametric designsParametric designs approach the baseline problem by:
Varying the stimulus-parameter of interest on a continuum, in multiple (n>2) steps...
... and relating measured BOLD signal to this parameter
Possible tests for such relations are manifold:LinearNonlinear: Quadratic/cubic/etc. (polynomial expansion)Model-based (e.g. predictions from learning models)
23Parametric modulation of regressors inverted U response toincreasing word presentationrate in the DLPFC
SPM{F}Polynomial expansion:f(x) ~ b1 x + b2 x2 + b3 x3 ...
LinearQuadratic
F-contrast [0 1 0] on quadratic parameterSPM5 offers polynomial expansion as option for parametric modulation of regressorsBchel et al. 1996, NeuroImage 4:60-6624Parametric modulation of regressors by time
Bchel et al. 1998, NeuroImage 8:140-14825User-specified parametric modulation of regressors
Polynomial expansion&orthogonalisationBchel et al. 1998, NeuroImage 8:140-14826Investigating neurometric functions (= relation between a stimulus property and the neuronal response)
StimulusawarenessStimulusintensityPainintensity
Pain threshold: 410 mJP1P2P3P4P0-P4: Variation of intensity of a laser stimulus applied to the right hand (0, 300, 400, 500, and 600 mJ)Bchel et al. 2002, J. Neurosci. 22:970-97627Neurometric functions
Stimulus presence
Pain intensity Stimulus intensity
Bchel et al. 2002, J. Neurosci. 22:970-97628Model-based regressorsgeneral idea:generate predictions from a computational model, e.g. of learning or decision-makingCommonly used models:Rescorla-Wagner learning modeltemporal difference (TD) learning modelBayesian modelsuse these predictions to define regressorsinclude these regressors in a GLM and test for significant correlations with voxel-wise BOLD responses 29Model-based fMRI analysisGlscher & ODoherty 2010, WIREs Cogn. Sci.30
AB
toutcome30
Glscher & ODoherty 2010, WIREs Cogn. Sci.310 131
Fixation crossAuditoryAuditoryVisualFixation crossTime (ms)0200400600800100012002000 500
or
Visual
DistractorTargetDistractorTarget1400Incidental learning of audio-visual associationsHypothesis: Incidental learning of this relation is reflected by prediction-error dependent changes in connectivity between auditory and visual areas.
80%
80%
den Ouden et al. 2009, Cereb. Cortex32Rescorla-Wagner model of associative learningp < 0.05, correctedrandom effects, n=16V1 & PUT increasingly activate the more surprising the visual outcome is V1 & PUT increasingly deactivate the more expected the visual outcome isRescorla-Wagner learningcurve (=0.075):
During learning, predictive tones
V1
PUT
den Ouden et al. 2009, Cereb. Cortex33OverviewCategorical designsSubtraction - Pure insertion, evoked/differential responsesConjunction - Testing multiple hypothesesParametric designsLinear - Adaptation, cognitive dimensionsNonlinear - Polynomial expansions, neurometric functionsFactorial designsCategorical - Interactions and pure insertionParametric - Linear and nonlinear interactions - Psychophysiological interactions
34Main effects and interactionsA1A2B2B1Task (1/2)Viewing NamingStimuli (A/B)Objects Colours Colours Objects Colours Objectsinteraction effect (Stimuli x Task)ViewingNamingMain effect of task:(A1 + B1) (A2 + B2)
Main effect of stimuli: (A1 + A2) (B1 + B2)
Interaction of task and stimuli: Can show a failure of pure insertion
(A1 B1) (A2 B2)35
Factorial design36A1A2B2B1Task (1/2)Viewing Naming Stimuli (A/B)Objects ColoursA1 B1 A2 B2Main effect of task: (A1 + B1) (A2 + B2)36
37A1A2B2B1Task (1/2)Viewing Naming Stimuli (A/B)Objects ColoursA1 B1 A2 B2Main effect of stimuli: (A1 + A2) (B1 + B2)Factorial design37
38A1A2B2B1Task (1/2)Viewing Naming Stimuli (A/B)Objects ColoursA1 B1 A2 B2 Interaction of task and stimuli: (A1 B1) (A2 B2)Factorial design38Example: evidence for inequality-aversion
Tricomi et al. 2010, Nature39
39Parametric interactionsp < 0.05, correctedrandom effects, n=16V1 & PUT increasingly activate the more surprising the visual outcome is V1 & PUT increasingly deactivate the more expected the visual outcome isSignificant four-way interaction in V1 and putamen:
V1
PUT
A+A-V+V-V+V-V+V-V+V-A+A-Primary visual cortex
Rescorla-Wagner learningcurves (=0.075):den Ouden et al. 2009, Cereb. Cortex40Psycho-physiological interactions (PPI)We can replace one main effect in the GLM by the time series of an area that shows this main effect.E.g. let's replace the main effect of stimulus type by the time series of area V1:Task factorTask ATask BStim 1Stim 2Stimulus factorTA/S1TB/S1TA/S2TB/S2
GLM of a 2x2 factorial design:main effectof taskmain effectof stim. typeinteractionmain effectof taskV1 time series main effectof stim. typepsycho-physiologicalinteraction41PPI example: attentional modulation of V1V5
attentionno attentionV1 activityV5 activitySPM{Z}timeV5 activityFriston et al. 1997, NeuroImage 6:218-229Bchel & Friston 1997, Cereb. Cortex 7:768-778 V1V1 x Att.=V5V5Attention42PPI: interpretationTwo possible interpretations of the PPI term:V1Modulation of V1V5 by attentionModulation of the impact of attention on V5 by V1.V1V5V1V5attentionV1attention
43FMRI Design Types44Blocked Designs
Event-Related DesignsPeriodic Single Trial Jittered Single Trial
Mixed Designs- Combination blocked/event-relatedBlock Design45Block Design SequencesConsider the simplest case, a block design with two conditionse.g., alternate tapping of two fingers vs. restlets assume 2 sec/volumeHow long should a run be?
Short enough that the subject can remain comfortable without moving or swallowing. Long enough that youre not wasting a lot of time restarting the scanner. Ideal is ~5 2 minutesimagestime courseof activationhaemodynamicresponse functionfinger tappingbaseline restSource: Jody Culhams web slides46Adapted from Gusnard & Raichle (2001)(E - Bad Control Design)
4747fmri-fig-11-12-0.jpg
Block Design SequencesHow fast should the conditions cycle?Every 4 sec (2 images) signal amplitude is weakened by HRF not too far from range of breathing frequency (every 4-10 sec) could lead to respiratory artifacts if design is a task manipulation, subject is constantly changing tasks, gets confused
pre-HRF
post-HRFEvery 96 sec (48 images) more noise at low frequencies linear trend confound subject will get bored very few repetitions hard to do eyeball test of significance
post-HRFSource: Jody Culhams web slides48Block Design SequencesEvery 16 sec (8 images) allows enough time for signal to oscillate fully not near artifact frequencies enough repetitions to see cycles by eye a reasonable time for subjects to keep doing the same thingpost-HRF
Other factors:
symmetric design some use longer rest vs. activation periods add a few extra images at the end to allow the hemodynamic response to catch up add extra time at the beginning to allow for the magnet to warm up and the subject to warm up (let the startle response die down)
Source: Jody Culhams web slides49But I have 4 conditions to compare!Adding conditions makes things way more complicated. Theres no right answer, but like everything else in fMRI, there are various tradeoffs. Lets consider the case of four conditions plus a baseline.
1. Main condition epochs all separated by baseline epochsPro: simplest case if you want to use event-related averaging to view time courses and dont want to have to worry about baseline issuesCon: spends a lot of your n on baseline measures
A. Orderly progressionPro: SimpleCon: May be some confounds (e.g., linear trend if you predict green&blue > pink&yellow)
B. Symmetric order
Pro: No linear trends to confoundCon: Each condition occurs at a different frequency, including some (e.g., yellow in top sequence) that occur at low frequencies which are noisierC. Random order in each runPro: order effects should average outCon: pain to make various protocols, no possibility to average all data into one time course, many frequencies involvedSource: Jody Culhams web slides502. Clustered conditions with infrequent baselines.
Kanwisher 4 condition designA. Kanwisher lab design sets of four main condition epochs separated by baseline epochs each main condition appears at each location in sequence of four two counterbalanced orders (1st half of first order same as 2nd half of second order and vice versa) can even rearrange data from 2nd order to allow averaging with 1st order
Pro: spends most of your n on key conditions, provides more repetitionsCon: not great for event-related averaging because orders are not balanced (e.g., in top order, blue is preceded by the baseline 1X, by green 2X, by yellow 1X and by pink 0X). As you can imagine, the more conditions you try to shove in a run, the thornier ordering issues are and the fewer n you have for each condition.
Jodys rule of thumb: Never push it beyond 4 main + 1 baseline.Source: Jody Culhams web slides51Power in Blocked DesignsSummation of responses results in large signals then plateaus (~10 sec)
Response Duration does not plateau and onset does not changeStimulus duration and interval compared with HRF
Stimulus duration and interval compared with HRFISI = 1 secChoosing Length of BlocksLonger block lengths allow for stability of extended responsesHemodynamic response saturates following extended stimulationAfter about 10s, activation reaches plateauMany tasks require extended intervalsBrain processing may differ throughout the task period
Shorter block lengths move your signal to higher temporal frequenciesAway from low-frequency noise: scanner drift, etc.Not possible in O-15 PET rCBF studies
Periodic blocks may result in aliasing of other periodic signals in the dataExample: if the person breathes at a regular rate of 12 per min and the blocks are 10s long (6 blocks/min)Could be problem if the aliased signal falls within the range of desired signals
From Scott Huettel, DukeLimitations of Blocked DesignsSensitive to signal drift or MR instability
Poor choice of conditions/baseline may preclude meaningful conclusions
Many tasks cannot be conducted well repeatedly
Non-Task Brain ProcessingIn experiments activation can be greater in baseline conditions than in task conditions!Requires different processing for interpretation
Suggests the idea of baseline/resting mental processesGathering/evaluation about the world around youAwareness (of self)Online monitoring of sensory informationDaydreamingNeurons that are wired together fire together
This collection of resting state brain processes is often called the Default Mode Network (DMN)
Blocked DesignPros.Cons.Excellent detection power (knowing which voxels are active)
Useful for examining state changesPoor estimation power (knowing the time course of an active voxel)
Relatively insensitive to the shape of the hemodynamic response.Stim ON 10sStim OFF 16s15Hz15Hz15HzStim ON 10sStim ON 10sStim OFF 16sStim OFF 16s10 repetitionsEvent-related Design57Buckner et al., 1998
Event RelatedWhat are Event-Related Designs?Event-related designs associate brain processes with discrete events, which may occur at any point in the scanning session. Can detect transient BOLD responsesSupports adapting task to response such as changing difficulty based on error rate
Why use event-related designs?Some experimental tasks are naturally event-related (future stimuli based on response)Allows studying within-trial effectsImproves relation to behavioral factors (behavior changes within blocks may be masked)Simple analysesSelective averagingGeneral linear models (GLM)Same Event Averaging
Sorting Into Common Groups Behavior Physiological Measure Outlier Rejection Transient vs. Task level ResponsesPeriodic Single Trial DesignsStimulus events presented infrequently with long inter-stimulus intervals (ISIs)
500 ms
500 ms
500 ms
500 ms18 s18 s18 sTrial Spacing Effects: Periodic Designs
ISI = 8sec (~12 trials)
ISI = 4sec (~45 trials)
ISI = 20sec (9 trials)
ISI = 12sec (15 trials)A20A4A8A12Want to maximize amplitude times number of trials per studyBandettini & Cox, 2000 The optimal inter-stimulus interval (ISI) for a stimulus duration (SD), was determined. Empirical Observation: For SD=2sec, ISI=12 to 14 sec. Theory Predicts: For SD2sec, RI = 8+(2*SD).
The statistical power of ER-fMRI relative to blocked-design was determined Empirical: For SD=2 sec, ER-fMRI was ~35% lower than that of blocked-design Simulations that assumed a linear system demonstrated estimate ~65% reduction in power Difference suggest that the ER-fMRI amplitude is greater than that predicted by a linear shift-invariant system models. Jittered Single Trial DesignsVarying the timing of trials within a runVarying the timing of events within a trial
Trial 1Trial 2Trial 3Trial 42 events3 events2 events1 event
Effects of Jittering on Response
StimulusResponseJittering allows us to sample BOLD response in more statesEffects of ISI on Detectability
Birn et al, 2002Jittered ISIConstant ISIDetectabilityEstimatedAccuracy ofHRFMax when stims are task state and stims are control stateDale and Buckner (1997)Detecting Using Selective AveragingLow ResponseFewer SamplesMid ResponseMore SamplesLarge ResponseMost samplesVisual stim duration = 1 s; acquisition 240 secTrials subtracted then correlation analysis with predicted response
68fmri-fig-11-17-0.jpg
Variability of HRF: Evidence
Aguirre, Zarahn & DEsposito, 1998 HRF shows considerable variability between subjects
Within subjects, responses are more consistent, although there is still some variability between sessionsdifferent subjectssame subject, same sessionsame subject, different sessionVariability of HRF: ImplicationsAguirre, Zarahn & DEsposito, 1998 Generic HRF models (gamma functions) account for 70% of variance Subject-specific models account for 92% of the variance (22% more!) Poor modeling reduces statistical power Less of a problem for block designs than event-related (do you know why?) Biggest problem with delay tasks where an inappropriate estimate of the initial and final components contaminates the delay component
Possible solution: model the HRF individually for each subject
Possible caveat: HRF may also vary between areas, not just subjects Buckner et al., 1996: noted a delay of 0.5-1 sec between visual and prefrontal regions vasculature difference? processing latency? Bug or feature? Menon & Kim mental chronometry
Block/epoch design v.s. event-related designs
71Randomised trials orderefMRI: Post-hoc classification of trials
72Post-hoc subjective classification of trialsPost-Hoc Sorting of TrialsFrom Kim and Cabeza, 2007Using information about fMRI activation at memory encoding to predict behavioral performance at memory retrieval.
True Memory Formationvs.False Memory FormationEvent model of block design
74Modeling block designs: Epochs vs Events
75Modeling block designs: Epochs vs Events
76Limitations of Event-Related DesignsLow power (maybe)Collecting lots of data, many runs
The key issues are:Can my subjects perform the task as designed?Are the processes of interest independent from each other (in time, amplitude, etc.)?
Event-Related DesignPros.Cons.Good at estimating shape of hemodynamic response
Provides good estimation power (knowing the time course of an active voxel)
Can have reduced detection power (knowing which voxels are active)
Sensitive to errors in predicted hemodynamic response
Event 1Event 2Event 3Event 4Mixed Design79Mixed: Combination Blocked/EventBoth blocked and event-related design aspects are used (for different purposes)Blocked design: state-dependent effects Event-related design: item-related effects
Analyses can model these as separate phenomena, if cognitive processes are independent.Memory load effects vs. Item retrieval effects
Or, interactions can be modeled.Effects of memory load on item retrieval activation.Thanks for David Glahns slides updated by JLL
Blocked (solid)Event-Related (dashed)Event-related model reaches peak sooner and returns to baseline more slowly.In this study, some language-related regions were better modeled by event-related.From Mechelli, et al., 2003You can model a block with eventsMixed Design
Summary of Experiment DesignMain Issues to ConsiderWhat design constraints are induced by my task?What am I trying to measure?What sorts of non-task-related variability do I want to avoid?
Rules of thumbBlocked Designs: Powerful for detecting activationUseful for examining state changesEvent-Related Designs: Powerful for estimating time course of activityAllows determination of baseline activityBest for post hoc trial sortingMixed DesignsBest combination of detection and estimationMuch more complicated analysesWhat is fMRI Experimental Design?Controlling the timing and quality of cognitive operations to influence brain activation
What can we control?Stimulus properties (what is presented?)Stimulus timing (when is it presented?)Subject instructions (what do subjects do with it?)
What are the goals of experimental design?To test specific hypotheses (i.e., hypothesis-driven)To generate new hypotheses (i.e., data-driven)Appendix (more information)85BOLD impulse response
86BOLD impulse response
87Design efficiency
88Sinusoidal modulation, f=1/33
89Blocked, epoch=20 second
90Randomized, SOAmin=4s, highpass filter = 1/120s
91Checking your design efficiency
92Thank youQuestions?
Thank for the internet resources. I was able to borrow a lot of slides from different presentations.93
