Please enable JavaScript.
Coggle requires JavaScript to display documents.
fMRI: Introduction to Experimental Design & Analysis - Coggle Diagram
fMRI: Introduction to Experimental Design & Analysis
Neuronal Activity
The Problem
Want to understand how brain functions
To do so need to understand how neurons function
fMRI only measure BOLD signal
Mostly measure blood flow over large (3x3x3) volume
Understanding How Neurons Function
What respond to
How respond
When respond
Neurone to Measure
Neurones respond rapidly to stimulus
Blood flow (vascular) changes much slower
Causes BOLD signal to be delayed & blurred version of underlying neuronal activity
Amount of delay & blurring described by HRF
BOLD signal a delayed & temporally smoothed of underlying neuronal (synaptic activity)
Delay & smoothing described as HRF
Blood Oxygen-Level Dependent Signal (BOLD)
Synaptic transmission consume energy & O2 to release glutamate transmitter
Glutamate release causes blood flow to increase over & above what required
To supply increase O2 consumption by neural activity
Causes concentration of deoxyhaemoglobin to decrease where blood flow increase
As deoxyhaemoglobin reduces MRI signal, increase in local neuronal activity causes increase in MRI signal
BOLD Signal
Reflects local synaptic activity more than local spiking (action potentials)
Therefore more closely reflects input & local processing than output spikes
BOLD & spiking not directly correspond
Is problem when comparing w. physiological data
But spiking & local processing normally closely correlated
Examining empirical good correspondence between spikes & BOLD
Local Field Potentials
LFP reflect pooled local synaptic activity
Similar to EEG
Averaged electrical activity of part of brain
Reflect average activity of synapses in that area
Haemodynamic Impulse Response Function (HRF)
The predicted BOLD response to instantaneous neuronal signal
Shape & delay varies somewhat between individuals, across brain areas & w. alertness
Sluggishness of HRF limits temporal resolution of fMRI to s rather than ms
But if know shape of HRF can estimate underlying neuronal response
But need to make assumptions to do so
Assumptions of Linear Systems Analysis
State that measures should reflect what trying to measure
BOLD response proportional to neural activity
BOLD responses sum linearly over time & space
Relationship between neural activity & BOLD responses not change over time
BOLD Response Proportional to Neural Activity
So size of BOLD response reflects amount of neural activity
Important as want to say that if BOLD increase by factor that neural signal also increase by same factor
Want to understand what neurones do, not the vascular system
Study: Proportional (Boynton et al., 1996)
A: BOLD signal at diff contrasts & diff times (3, 6, 12, 24s)
M: Group of Ps sit in scanner & fixate on point, show flickering patterns to give strong signal in V1, measure at diff contrasts & see whether BOLD response change in proportion to contrast
R: BOLD response change anatomically & in rough proportion to contrast
Study: Proportional (Heeger & Ress, 2002)
A: How contrast reflect changes in neural activity
M: Data from V1 area of brain in macaques, plot fMRI activity & add average curve of responses from macaques as overlay, compare line and own findings
Visual abilities similar so assume have similar V1
R: Responses from fMRI similar to those from activity measures
Study: Correlated (Logothesis & Wandell, 2004)
A: Correlation of BOLD w. local field potentials (LFP) & multi-unit activity (MUA)
P: Monkey
M: Measure monkeys in scanner & at same time measure w. electrodes
R: BOLD signal closely correlated w. LFP but not MUA
BOLD signal matches LFP filtered by (convolved w.) haemodynamic impulse response function HRF
LFP reflects pooled local synaptic activity similar to EEG
Study: Disproving
Large component of BOLD response not related to neuronal activity (Sirotin & Das, 2009)
Only stimulus-driven component is correlated
Stimulus-driven component well correlated w. spiking (neuronal) output activity (Lima et al., 2014)
Explanation
Likely due to BOLD being driven by glutamate release
Not postsynaptic activity
Explanation of Differences
Due to anaesthesia
Large fluctuations in BOLD during anaesthesia
Logothetis' results may have been dominated by task-/non-neural signals
Implication
Evidence BOLD signal pick up on regional synaptic activity
Spiking activity (no. action potentials at one time) not what drives activity but is rather the synaptic activity
Generated by local & input activity into synapses
BOLD Responses Sum Linearly Over Time & Space
So BOLD response to many stimuli is sum of all responses (neural activities)
Always have mountable neurones doing same things in same area
Important as look at diff neural systems & make assumption about how active & function
Can assume the signal measure is proportional to sum of neural activity
Study: Sum Linearly (Boynton et al., 1996)
A: Response to & prediction from response to pulse stimulus
R: Actual & predicted responses mapped line up & differences explained by how neurones act to longer stimuli
Reasonably well for shorter times (6 & 12s)
Very well for longer times (24s)
Relationship Between Neural Activity & BOLD Responses Not Change Over Time
So BOLD response to given stimuli always the same
Stable over time
Same response whatever time point it occur
Important as show response is fixed
Limitation of Assumptions
BOLD primarily reflect synaptic activity
Synapses can be inhibitory & excitatory
Inhibitory can lead to reduction in spikes but increase in BOLD
Cortical neurones organised into micro-circuits containing excitatory & inhibatory neurones
Precise relationship between neural activity & BOLD depend on balance of excitatory & inhibitory neurones in cortex
Balance can change
Potentially change relationship between BOLD & neural activity
Cortical Microcircuits & BOLD (Logothetis, 2008)
Input -> excitatory neurones & inhibitory neurones -> output
Weak inhibition = BOLD - spike coupling ++
Result in proportional output to input
Strong inhibition - BOLD - spike coupling +? -?
Even when strong input that may be weaker output
GABA
Can vary w. menstrual cycle, caffeine, age
Having coffee before scan means peakier response
Study: GABA & BOLD (Donahue et al., 2010)
R: More GABA have lower BOLD response amplitude
Study: Correlation w. GABA & BOLD (Harris et al., 2015)
R: No correlation
Study: GABA & BOLD (Muthukumaraswamy et al., 2011)
M: MRS
R: More GABA have weaker & more sluggish HRF
What Change Relationship Between BOLD & Spikes
Changes in excitatory/inhibitory balance
Adaptation
Neuromodulation
Top-down feedback
Attention
Vascular mechanisms
Issue
Want to avoid inadvertently changing in experiment
Most have no control over
Experiment design means can account for them
Should consider them when analysing papers
How implicitly account for them (as explicit accounting requires additional studies)
BOLD & Spikes Implications for Analyses of fMRI Data
Can only measure BOLD not spiking
But if relationship between BOLD spiking is linear (constant & proportional) can use as method for measure
Measure HRF & use to recover underlying neuronal activity from BOLD response
Use BOLD response as indirect but roughly proportional measure of relative neuronal response
Keeping relationship constant is crucial
Control the measures
Remember BOLD measures population not single unit activity
Limitation of HRF
Restrict temporal resolution of fMRI
Quick occurring changes blur out and appear as single signal
Resolution
Temporal
Spatial
Temporal Resolution of fMRI
Temporal resolution determined by factors
Temporal precision can be much less than second
Delay between neuronal activity & BOLD response can be relatively stable
Factors
How often can take measurements given TR typically 1-3s
Limited by scanner gradient magnet speed & how large volume want to measure over
New parallel imaging methods can achieve TR of ~.5s
Intrinsic temporal resolution of blood flow response
Few s
Spatial Resolution of fMRI
W. standard scanners, voxels can be as small as 1mm or less
Blood flow changes more widespread than underlying neuronal activity
By few mm
Much of signal comes from draining veins away from site of neuronal activity
In practice factors limit spatial resolution of fMRI to few mm
High-field scanners (7T & more) & special MRI techniques can allow better resolution
Study: Spatial Resolution Limits (Logothetis & Wandell, 2004)
Logic of Experimental Design
Experimental & Statistical Strategies
Model-driven experiments
Model-free experiments
Model-Driven Experiments
Measure how brain activity varies as result of experimental manipulation
In all voxels/within specific part of brain
Process
Present stimulus/have subject perform task
Measure BOLD fMRI signal
Look for changes in BOLD signal (neural activity) in response to task/stimulus manipulation
Paradigms
Subtraction
Adaption
Classification
Subtraction
Look for increase/decrease in BOLD response as function of stimulus/task manipulation
Relative to control state
Logic
Neurones selective for stimulus/task will increase activity when presented with stimulus/perform task
Cause increase in BOLD signal
Voxel/region showing increase in fMRI response when presented w. particular stimulus/task relative to some control condition is evidence of neurones selective for stimulus/task
Found by subtracting response of control from response of manipulation
Study: Subtraction Paradigm (McGugin et al., 2012)
A: Responses in fusiform face area & how reflects expertise
R: Some voxels respond better to certain objects
Response reflects diff in response to manipulation vs control
Example: Contrast & Orientation in Primary Visual Cortex
Neurones in V1 sensitive to stimulus contrast
Neurones in V1 sensitive to particular orientation
To test contrast selectivity could measure fMRI response at high & low contrast to see if responses differ
Work as all neurones respond in same way to contrast
To test orientation need to select a stimulus w. contrast and unoriented stimulus as control
But contrast would measure contrast
But unaligned still have response to weak orientations
So would appear V1 have no sensitivity to orientation
Limitations
Pure insertion
Based off assumption that only difference between stimulus/task condition is parameter of interest
Measured state is control state plus the stimulus
In reality is difficult to accomplish
Other neurones in same voxel/area may be selective for control condition but not task/stimulus condition
Addition of task/stimulus component to control condition may change neuronal processing non-linearly
Assumes neurones not selective for stimulus/task will respond equally strongly for both task & control
But if not selective neurones respond differently results may be uninterpretable/misleading
Adaptation
Look for changes (decreases) in BOLD response following adaptation
Evidence of stimulus-selective adaptation
Most neurones reduce responsiveness after prolonged exposure to preferred stimulus
After adaptation neurones will respond more weakly to stimuli
Neurones will only adapt strongly to preferred stimuli
If observe reduced response to stimulus after prolonged exposure but not reduced response to other stimuli is evidence of neurones selective for stimulus
Logic
Only neurones selective for stimulus will respond more weakly to adapted stimulus
Difference in responses to adaptation dependent of the stimuli
If equal/no response to stimuli suggests are not selective
Example: Disparity- & Velocity-Based Signals (Rokers et al., 2009)
A: Disparity- & velocity-based signals for three-dimensional motion perception in human MT+
R: Stimuli that adapt to movement in depth there is no evidence of adaptation in V1 but are in MT
Show selective adaptation to motion stimulus in that part of brain
Advantage
Adapted & unadapted stimuli can be closely matched
Both stimuli can be used as adapters & result averaged
Removes potential absolute response differences between stimuli prior to adaptation
Disadvantages
Adaptation may change local excitatory-inhibitory balance
Lead to change in relationship between neural activity & BOLD
Risk making results uninterpretable
Adaptation may primarily affect spiking output w. no effect on synaptic input
Make neurones less sensitive to synaptic afferent stimulation
Can fail to detect adaptation in output of area if input to area not adapt
If observe adaptation could be synaptic input for other area that has adapted
Issue of Inherited Adaptation (Larsson & Harrison, 2015)
Area 1 -> adaptation -> Area 2 -> adaptation <- Area 3
Area 2 & 3 adapt each other
fMRI response in Area 2 may reflect other mechanisms
Other Mechanisms
Neuronal adaptation in Area 2
Neuronal adaptation in Area 1 resulting in reduced feedforward synaptic input to Area 2
Appear as reduced activity/no adaptation
Neuronal adaptation in Area 3 resulting in reduced synaptic input to Area 2
Appear as reduced activity
Issue of Adaptation & Expectation
When repeated (adapted) stimuli are unexpected BOLD responses are larger than when are expected (Summerfield et al., 2008)
Adaptation may reflect expectation/stimulus novelty rather than neuronal adaptation
Effect disappears when attention controlled (Larsson & Smith, 2012)
When Ps don't attend to stimuli the effect of expectation goes away but adaptation remains
Classification
Look for changes in distributed (multivariate) BOLD responses across multiple voxels
That can predict some property of stimulus/task
Changes in pattern of activity rather than simple in/decreases
Look for patterns that represent stimulus & use as evidence that part of brain is processing stimulus
Multivariate Pattern Classification Analysis
Cortical neurones organised in columns
Column defined by having similar response properties throughout column but diff from nearby columns
Columns too small to be resolved w. fMRI
However each voxel samples slightly diff proportions of columns selective for diff features
Responses of individual voxels may be weakly biased for particular feature
Logic
Voxels contain slightly diff proportions of neurones selective for diff stimuli
Voxels have slight biases for some stimuli over others
Biases in individual voxels too small to be reliable detectable & cancel out when averaging across voxels
Machine learning methods look at combined (not averaged) responses across voxels w/in, region to find consistent patterns of biases
If patterns found can conclude that are neurones in region that are selective for particular feature/stimulus
Machine Learning in MVPA
Look for pattern by looking at activity in multiple voxels at same time
Study: MVPA (Kamitani & Tong, 2005)
A: Decode visual & subjective contents of human brain
M: Showed diff orientations & measured responses
R: Diff voxels in visual cortex had diff preference for orientation that varies between Ps but is consistent w./in P
Can be used to predict out what stimulus P had seen at better than chance level
Disadvantages
Not sure the origin & spatial scale of voxel stimulus biases
Probably vascular in origin as blood vessels collect preferentially from columns w. particular stimulus preference
Recent studies suggest biases large scale not columnar
Only works if is underlyin unevent distribution of neuronal preferences
Failure to classify may be because of lack of sensitivity for feature or because neurones selective for feature are mixed spatially rather than clustered in columns
So null result is inconclusive
Not quantitative
Cannot be used to compare neuronal selectivity between areas/conditions meaningfully
Origin of Voxel Biases (Freeman et al., 2011)
Orientation classification in V1 reflects large-scale orientation preferences
Not columnar organisation
Issue of Interpretation
Successful classification means is info in data to make classification
Does not necessarily imply that brain uses this info
MVPA primarily useful when stimulus selectivity of neurones already known
Example: Brain Not Using Info
Pattern of responses of retinal photoreceptors differ between faces & non-faces
Mean is info about faces in retina
But does not mean photoreceptors of face selective or represent faces
Model-Free Experiments
Identify patterns in brain activity that change over time/space
Not assume how changes should look
Practice of Experimental Design
Process
Planning experiment
MRI data collection
Stimulus & task presentation
Planning Experiment
Single/multiple subjects
Region-of-interest/whole-brain analysis
Choose appropriate data analysis method
MRI Data Collection
Slice prescription & length of runs
Number & length of runs
Choose imaging parameters
TR
TE
Flip angle
Stimulus & Task Presentation
How & when to present stimuli/tasks
Choose appropriate stat methods
fMRI Data
One volume per timepoint
Each volume made up of stacks of slices
Each slice made up of voxels
Image elements
Value of each voxel gives BOLD signal at that point
Approx. 3x3x3 mm
Selecting Receiver Coil
Head coil
Surface coil
Head Coil
Wide coverage
Poor signal
Surface Coil
Low coverage
Good signal
Advantage
Fit average head so better scan
Selecting Slice Prescription
Choose which part of brain want to cover
Select & rotate box to select from rather than entire brain
Trade-off between coverage & TR
Number of slices & time to acquire each volume
Advantage
Less time so more samples per second can be taken
Limitation
Can't go below .5 s
Imaging Parameters
TR, TE, & flip angle the main parameters of MRI sequence
Determine image contrast
Each volume consist of fixed no. slices acquired in sequence every TR
Each slice contain fixed no. voxels
TR
Repetition time
How often new volume acquired
Typically 1-3s
Most involved alteration/setting as rest based on MRI physics
Long TR
More slices acquired
Shorter TR
More data obtained
Better response can be measured/resolved
Task/Stimulus Presentation Design for Model-Driven Experiment
Block design
Event-related design
Mixed design
Phase-encoded design
Block Design
Alternate between diff stimuli/tasks/controls in temporarily extended blocks
Typically every 20-30s
Measure fMRI response to each block
Event-Related Design
Most commonly used
Measure response to brief events
1-10s
Repeated many times
5-15s between events
Mixed Design
Combination of block & event-related design
Phase-Encoded/Cyclic/Periodic Design
Alternate between two stimuli/tasks in constant-period cycle
Analyse by looking for correlation w. sinusoid
Model-Free Analysis Methods
Principal component analysis
Independent component analysis
Method of Analysis
Divide data into spatial components each w. diff timecourse
Instead of one image per timepoint, returns one image per components w. associated timecourse that is same for all voxels in components
Advantage
Good way of identifying potential artifacts
Disadvantage
Model-free but hard to interpret results