fMRI Paradigm Design
Chris Dodds
Property of GlaxoSmithKline
• Aims of an fMRI experiment
• Limitations and trade-offs of fMRI
• Blocked vs event-related designs
• Importance of strong anatomical hypotheses
• Subtraction logic and the importance of the baseline
• Tips on design efficiency
• Good and bad fMRI experiments
• Example fMRI project – effects of BDNF polymorphism on episodic
fMRI and the BOLD signal
• Neurons fire → increase
in blood flow.
• But, the increase in O2
utilisation is much lower.
• Concentration of deoxy
Hb decreases and leads
to a decrease in the
distortion of magnetic field
(e.g., image appears
brighter with decreased
distortion from deoxy Hb).
• T2* relaxation signal is
sensitive to this Blood
Oxygenation Level
(BOLD) change and we
measure this difference
due to psychological
Local Field Potential correlates with BOLD:
(Logothetis et al., 2001)
Main Goal of an fMRI experiment
Design fMRI experiments that are sensitive to a specific
hypothesis (i.e, statistical planned comparison or "contrast").
– Example: Is brain region X more involved in processing syntax or semantics of language?
– Translation in real life: Design study that maximizes the difference in BOLD signal between
the experimental (‘Task on’ and control condition (‘Task off’). This is another way of saying
that you want maximize ‘design efficiency’
Subtraction Logic and the importance of the baseline
• An fMRI study is like any other experiment in establishing the
relationship between two variables
-Explanatory / independent variable : psychological task
-Outcome / dependent variable : brain activity
We are evaluating the brain’s response to an experimental
manipulation consisting of an experimental condition (sometimes
called ‘activation task’ and a control condition (sometimes called
‘baseline task’).
- fMRI brain response is arbitrary and has no meaning independent of the
-Experimental Condition or Activation Task: involves the psychological
process of interest (e.g., face-processing)
-Control Condition or Baseline Task: should be as similar as possible to
the activation task but should lack the psychological process of interest
(e.g. object processing)
Activation for PsychProcess = Experimental – Control
Cognitive subtraction
Pure Insertion: Psychological tasks can be elaborated on by inserting additional
psychological processes.
The validity of this assumption depends on the psychological model being
tested. This may work well for low-level psychological processes (vision), where
certain it is evident what is added or subtracted from one another. But what
about higher level cognitive processes, whereby subtracting out one process
cannot be easily achieved without altering the other psychological process that
is leftover? Can we really know what is psychologically ‘leftover’ from a
cognitive subtraction? Careful thought is needed when designing your
experiment... Your experimental condition is only as good as what your control
condition is.
What is being subtracted?
Objects > Textures
object shapes, irregular shapes, familiarity,
visual features (e.g., brightness,
contrast, etc.), actability, attention...
Word > NonWord
phonology, semantics, lexical
properties, imagery, animacy...
- Expense
- Depending on academic/industry study, can be up to £1,200 per 90 minute
- Assuming 40 volunteers, each scanned twice, costs can be £100,000+ for a
single study
- Think carefully – do we need fMRI?
- What extra information does imaging provide?
- Is it useful to know where in the brain something happens?
- Can we design experiments to go beyond localisation?
- Time
- Maximum approx 90 mins in the scanner
- Includes structural scans (approx 20 mins) and rest time
- Perhaps 60 mins for task performance
- Multiple tasks?
- Alertness/attention
- tradeoff between lots of data and good data
Environment/Equipment limits what we can do
Behavioural Performance
– Affects interpretation of results
– E.g. Reduced activation + worse performance in patients (vs controls) could indicate
a) Neuronal dysfunction causing worse performance or
b) Worse performance causing reduction in neuronal signal.
- Match behavioural performance
- Difficult to interpret differences in activation
- Increase in activation in one group may indicate better neuronal function or less
efficient neuronal function
fMRI data relies on correlation, not causation
Importance of converging evidence from neuropsychology
Aron et al – Motor inhibition performance associated with volume of damage to right inferior frontal cortex
Reduced activation in this region in ADHD patients relative to controls
 Patients
– Parkinson’s: tremor
– Alzheimer’s: difficulty remembering task instructions
– Obesity: difficulty fitting in the scanner
– Ketamine users: ulcerative cystitis
– ADHD: turning up
– Tailor paradigms to take into account specific patient requirements
– Shorter, easier tasks with regular breaks are often more appropriate
 Pharmacological fMRI
– Practice effects/neural adaptation – pay close attention to repeatability of
task (in within-subjects designs)
– Side effects: Nausea can be a serious concern in the scanner
– Timing of scanning – where do you want Tmax?
– Controlling for vascular effects
–Control task
–Arterial Spin Labelling (ASL)
– Behavioural performance – difficult to interpret neural effects in the
absence of behavioural effect
- Artifacts
Artifacts (aka ‘noise’) known not to be related to neural activity are
everywhere (e.g., field distortions, subject movement, respiration,
low-frequency scanner drift) and reducing the contributions of
these variables is always a challenge. We don’t also want bad
experimental design as another contributor of ‘noise’ to our data.
Designing a good fMRI experiment is largely an exercise in
maximising signal to noise ratio
-Minimise (and correct for) subject motion
- Pay close attention to frequency of events in design
BOLD signal
fMRI signal is an indirect measure of neural activity and folds
out "sluggishly" over time - making the design of an ‘efficient’
fMRI experiment a not-so-straightforward exercise.
Blocked designs
Canonical HRF
fMRI Data
Implications (Neural)
Best for detecting amplitude differences between conditions (i.e. most efficient
design when the aim is to detect differences in BOLD signal between conditions)
Not good at isolating responses to single events within blocks
Cannot estimate shape of HRF effectively, but are robust when there is uncertainty
about the timing and shape of HRF
Can acquire more trials in less time than event-related design, because you don’t
have to worry about spacing trials apart to get an estimate for each individual event
Blocked designs
Canonical HRF
fMRI Data
Limitations (Psychological)
- Highly predictable occurrence of stimuli: subjects know what is
coming and may alter strategies accordingly (not always a pro)
- Inflexible for more complex tasks: impact of oddball stimuli? or
stimuli or events that occur uncontrollably?
- Ecological validity. Does blocking trials change the psychological
process you are interested in?
Event-related designs
In freeing us from the necessity of block designs, event-related fMRI enables
us to design more complex and novel experiments.
- Spaced single trial design: Present trial, wait for HRF to pan out, then present next trial.
Brief stimuli every 16s: HRF rises (2s), peaks (4-6s) and falls back to baseline (10-14s).
- Rapid single trial design: Individual trials spaced closer together and decompose
overlapping HRFs through sophisticated jittering of time in between trials or randomization.
How it works:
Present events far enough apart so you can estimate the HRF for each individual
OR randomize the order of presentation of events or jitter (i.e. randomly space out)
the stimuli with variable times between each stimuli. Randomization or jittering
allows the interval between two stimuli to be on average separated far enough
apart to enable estimation of the HRF to (averaged) trials containing a single
event type
Event-related designs
Canonical HRF
fMRI Data
Implications (Neural)
Best for estimating the shape of HRF and looking for differences in timing.
Also allows one to estimate activation in response to single events, allowing much
more flexibility in the paradigms that can be used and questions that can be tested.
Need many more trials per condition compared to block design in order to increase
detection power. Trial averaging helps get rid of noise in estimating response to
single events, but more trials makes averaging more robust.
Because of the need for more trials and the jittering of time between stimuli, eventrelated designs come at the cost of increased scanning time.
Event-related designs
Canonical HRF
fMRI Data
Implications (Psychological)
Allows researcher flexibility to design task with multiple parts and you can estimate
HRF to each part of the task: E.g., Cue - + - Target - + Feedback.
Allows opportunity for trial sorting. Researchers can re-conditionalize their paradigm
based on behavioral responses. E.g. Comparing events that were subsequently
remembered vs. forgotten, accurate vs. inaccurate judgments, etc.
Importance of anatomical hypotheses
 Multiplicity problem
X 36 slices
= 147,456 voxels
- High chance of
false positive results
Importance of anatomical hypotheses
Anatomical hypotheses enable us to reduce the multiplicity
problem and to ask more focused, meaningful questions
about brain function than mere localisation, e.g.
– How does region X support process Y?
– What are the response properties of region X?
– What computational processes are carried out by region X?
– How is activation in region X modulated by drug Y?
Region of Interest (ROI) Approach
 Restrict our analysis to a single region, or set of regions.
 Define regions anatomically, e.g. the hippocampus, or functionally, e.g. on the
basis of areas activated in our experiment
 Voxelwise comparisons with ROI or average activation across voxels
Importance of anatomical hypotheses
 Weak anatomical hypotheses lead to reverse inference...
The Importance of Strong Anatomical
Hypotheses in fMRI
•Reverse inference in ‘neuromarketing’:.
“…most striking of all was the flurry of activation in the insular cortex of the
brain, which is associated with feelings of love and compassion. The
subjects’ brains responded to the sound of their phones as they would
respond to the presence or proximity of a girlfriend, boyfriend or family
member… In short, the subjects didn’t demonstrate the classic brainbased signs of addiction. Instead, they loved their iPhones.”
Caveat: The insula is activated in 1/3 of all brain imaging studies, not just
those involving love. In fact, the insula has been shown to be involved in
negative emotion (e.g., disgust). We could have said with equal
confidence “when subject heard their iPhones, they were disgusted…
No anatomical hypothesis
“Shopping list” of areas
What does this tell us?
Certainly not going
beyond localisation
Strong anatomical
Good baseline
Beyond localisation
Converging evidence
Efficient experimental design
Scan for as long as possible.
Power depends on df which depend on no of independent observations (scans).
*Sample size has more impact on power than # of scans though…
Minimise “dead time” whenever possible– e.g., reduce ITI.
Avoid breaks in scanning
- disrupt the spin equilibrium (i.e, require extra dummy scans at the beginning of the run)
- reduce efficiency of any temporal filtering during analysis (data no longer a single timeseries)
- introduce potential "session" effects
Don’t contrast trials that are far apart in time
- low frequency noise (scanner drift), high pass filtering tries to get rid of this, but will also
get rid of experimental variability if contrasting trials very far apart.
- in block designs, avoid using blocks that are longer than 50s
Randomise trial order or stimulus onset asynchrony (SOA) – essential.
- i.e., AABBAABB with fixed ITI of 2 sec is very inefficient.
Add null events (e.g., fixation cross) to allow another ‘low level baseline’ to compare
experimental and control conditions to.
Example: Measuring the effects of drug on
hedonic brain responses to alcohol in the
Expense – do we need fMRI? What will it add?
Time – do we have enough?
Environment/equipment – can we do this?
Artifacts – possible sources of noise?
Sluggish BOLD signal – an issue?
Event-related/block design?
Strong anatomical hypothesis?
Going beyond localisation?
fMRI – Recognition Memory Task
- Indoor or Outdoor?
- Old or New?
= Met carriers (Encode – Rest) – Val carriers (Encode – Rest)
Condition of
not recognised
No effect of polymorphism on activation
during successful memory encoding
 fMRI is not a ‘magic bullet’
– Researchers often think “I’ve got this task, let’s run it with fMRI and see what happens.”
– Usually leads to uninterpretable results
– reverse inference – i.e. long list of activated regions and an attempt to explain them by referring
to previous studies
– Simplistic claims based on localisation – We have discovered the neural basis of X
– Need to design fMRI experiments with a clear hypothesis and ask what information we
can get from fMRI that we can’t get from a less expensive and less time-taking
– Designing an fMRI experiment involves a series of decisions about trade-offs and
– e.g. collect as much data as possible but not at the expense of tiring out subjects and collecting
low quality data
– Design an interesting study but not try to ask too many questions – keep as simple as possible
– Always remember, the scanner is a noisy, unfamiliar, uncomfortable environment.
There is a limit to what the subject will put up with, and this will be (much) lower than
what they will tolerate outside of the scanner.
Property of GlaxoSmithKline

similar documents