Lecture 8: Advanced Analysis in fMRI Flashcards
What are the different advanced analysis in fMRI? - (4)
- exploratory (intersubject correlation, independent component analysis)
- functional connectivity
- multivariate techniques (MVPA, pattern classification)
- voxel modelling
Advanced analysis in fMRI is beyond simple yes-no hypothesis but towards more
open quesiton about “how the brain thinks”
What is inter subject correlations? - (2)
Functional MRI time courses that are shared by different individuals while performing same experimental tasks or experiencing the same stimuli
looking at common pattern of activaitons across partiicpants
In intersubject correlaiton they propose that
if a group of individuals all show the same activation, regardless of the experimental hypothesis, that shared activation likely reflects common mental processing.
What does inter subject correlation process involve? - (2)
involves looking for regular patterns in the data and then interpreting those regularities based on knowledge of the experimental paradigm.
But unlike the approaches discussed elsewhere in the chapter, those regular patterns will be between subjects rather than within individual subjects.
An example of intersubject correlation is
study by Hasson and colleagues
What questions did Hasson et al 2004 wanted to ask - (3)
What happens in the brain when subjects watch a film?
What parts of the movie reliably correlate with increased BOLD signal (reverse correlation - what makes that BOLD signal fire in unconstrained way)? -
To what extent are these correlations regionally selective?
Hasoon et al., 2004 had no - (2)
No model, very limited prior assumptions
Did not bring complex GLM and no idea experimental conditions started or finished as showing pps a movie clip
What is Hassoon et al’s 2004 study interested in?
interested in understanding the brain processes that underlie perception in open-ended, natural settings, such
as when people freely perceive and attend to distinct parts of the complex,
changing world.
What is the methodology of Hassoon et al’s 2004 study? - (4)
researchers showed five subjects a 30-minute excerpt from a classic western movie (Sergio Leone’s The Good, the Bad, and the Ugly).
Subjects were not instructed to perform any particular task while watching the movie, but to view it normally and describe the plot at the end of the experiment.
Then, using standard protocols for data collection, the researchers simply collected a time series of fMRI data while their subjects watched the movie.
For a control condition, they repeated the experiment with a separate group of
subjects who were lying in the scanner with their eyes closed,
What did Hassoon et al’s 2004 study found? - basic findings - (5)
NON-SELECTIVE
They found significant across-subject correlations in about 30% of the cortical surface, including many parts of the visual cortex and the visual processing streams, as well as in regions within the frontal cortex in experimental condition
Large parts of cortex show correlated activity during the film across pps correlation
Within participants brain activity across large parts of brain is correlated with itself
For control condition resulted in only chance correlations between participants
These parts of brain doing something similiar during the movie
What did Hassoon et al 2004 did in their data analysis? - (3)
The time course of the non-selective activity is similar in all subjects and across brain regions in the ventral occipital temporal cortex.
This pattern can be used as a regressor in a general linear model.
The unexplained (residual) variation can then be investigated for patterns of regional selectivity…
What does this diagram show? - (5)
Five different participants and BOLD signal chance during the movie
Their peaks in similar places in different people
Overall average in red show consistent deviations from the average in experiment
Use red line as regressor and put in GLM and should see parts of the brain that is not explained by this regressor must be doing something different in each pp or each voxel
Question is whether those distinct things can be explained by non-selective activity and do they correlate between people? If they do then correlations telling about what part of brain is doing in two different people at same time
What did they find after accounting for non-selective activity in large parts of cortex across different people Hassoon et al 2004? - (3)
Remaining residual activity that is coherent and correlated across different individuals (inter-subject correlation)
Having different time course for different regions – how do the regional time courses relate to function?
Strong voxels showing strong correlations between participants in a certain point of a movie
What did Hassoon et al’s 2004 study found?
Intersubject correlations resulted from two different effects - (4)
First, activation throughout much of the brain,
including most visual regions, rose and fell in a similar pattern across all
subjects.
To interpret this activation, the researchers used a reverse correlation
approach: they identified time points when this collection of regions showed
maximum activation and then evaluated what was happening in the movie
at those times.
They found that this non-spatially-selective component tended to have highest amplitude during the most surprising and evocative points in the movie (e.g., gunshots, explosions, or unexpected plot twists); thus, it could reflect a broad increase in arousal.
Second, there were spatially selective in-
tersubject correlations that had a unique time course in each of several brain
regions.
What does this figure show of Hasson et al 2004 study?
common activation in the fusiform gyrus tended to increase in response to movie scenes that involved a close view of a face,
Essentially what Hassoon et al 2004 found that functional anatomy revealed that
some regions are specifically active during qualitatively different shots
What did Hassoon et al 2004 finding some regions are specifically active during qualiatively different shots using reverse correaltion?
This technique of reverse correlation might have some validity for showing what is going on in brain without bringing specific experimental design and assumptions
What is reverse correlation based on - (2)
Single-unit studies with monkeys
Monkey shown a lot of stimuli and researchers asking what makes that neuron fire?
What does this diagram show of Hasoon et al 2004 study? - (3)
Peaks in activity in posterior collateral sulcus (adjacent to parahippocampal “place area”) correlate with interior scenes/landscapes.
It’s worth bearing in mind that we are only seeing the shots where these regions do respond
Some subjectivity in interpretation of what regions do since movie is not conrolled
What does this diagram show of Hasoon et al 2004 study - (4)
Peaks in activity in the mid post-central sulcus in somatosensory cortex seem to correspond to scenes involving manipulation with their fingers.
A new finding? “Mirror neurons”?
As you are watching someone touching something in movie generating same neural activity if touching the same thing
Because of subjective element it is hard to be sure (had to put arrows)
What is advantages of exploratory approaches (ISC)? - (3)
Few assumptions = not biasing what we can discover
Tasks less constrained by design considerations; allows more “ecologically valid” tasks
Rich data = every single voxel was scanned and telling us we have not guessed
What is disadvantages of exploratory approaches like ISC? - (3)
Extraneous variables poorly controlled e.g., selective activity there is big face or explosion so not controlled , no attempt to separate diff influences on brain activity –> anything can be affected
Subjective element in interpreting/highlighting results
Requires analytical innovation = requires whole new data analysis to analyse natural viewing
Typical fMRI data analyses (e.g., using FSL) are based on fitting a
complex GLM model to the data.
In many studies the GLM model is complex and may include - (2)
untested and/or implicit) assumptions like linearity but way different tasks are fractioned and combined into different thought processes
Forcing ourselves an answer we are expecting to see
The implementation of the GLM in typical fMRI data analysis relies on many
assumptions about the way neural activity leads to a characteristic pattern of change in blood flow and how these changes is add up to produce the observed BOLD signal.
Alternative models are not typically tested
against one another.
Independent component analysis provides a way of doing a
“model-free” exploratory analysis
ICA is more systematic exploratory approach than - (2)
ISC ,
has a certain toolbox but avoid fitting complex GLM model to data
Independent componen analysis introduced by McKoewn assumes that - (2)
fMRI data consists of spatially overlapping components, each with independent spatial pattern and unique time course
that is the components contribute differently to the overall four-dimensional time course at different points in time
What does this diagram show of ICA?
Schematic example, thicker lines indicate when specific components contribute relatively heavily to specific time points
What does the term ‘ independent’ means in ICA? - (2)
algorithm minimizes the overlap between the components;
Each ICA component is independent of all other components
What does ICA offer?
a principled mathematical technique for identifying unknown and statistical independent signals (in a mixture)
Explaining ICA in terms of paints - (2):
If we had a collection of mixed brown paints and knew made by mixing some other paints
Try out and work what colours went in mixing the brown
What does ICA assumes
that some distinct signals have been mixed together to make the observed data
In ICA, to find out what must have gone into mixtures, we reverse the process by
changing the mixing matrix until unknown sources are distinct as possible
For a large number of datapoints (e.g., time series) in ICA, it uses mutual information which
measures how much knowing the value of x tells you about the corresponding value of y
In ICA, the “sliders” in diagram are moved until the
mutual information between x and y (the putative sources) is minimized.
The components in ICA reflect
any systematic change in BOLD signal such as stimulus driven activation produced by internally generated rhoughts, physiological processes like head motion, respiration and other nusiance variation - figure below
What does this diagram show? - (5)
Extracting task-related and non-task-related
components using ICA.
In this session, the subject viewed photographs of faces and of photographs of houses in alter-
nating 40-s blocks.
A series of components, four of which are shown here, were extracted using ICA.
(A) Within voxels in the fusiform gyrus and the dorsal parietal cortex, there was a task-related modulation of activation.
Other components identified by ICA were associated with non-task-related vari- ability, such as a transient scanner artifact (B), high-frequency physiological noise around vessels in the base of the brain (C), and head motion (D), which was evident as characteristicchanges in activation along the edge of the brain.
In functional connectivity,
You take the activation of a specific part of the brain (the seed region) and extract BOLD signal and use BOLD signal as regressor in GLM and reveal those part of brain that show similar pattern of activation in seed region
Diagram of functional connectivity
A set of regions where brain activity is active during rest - (2)
Default mode network which parts of brain where activity is correlated with each other and regions often active while participants don’t have a task to do
At rest variation in visual networks
What is neurosynth? - (2)
Useful for functional connectivity and archives results from many studies
Can look at areas you are interested in like hippocampus using coordinates
What does this functional connectivity study by Smith et al 2009 show? - (4)
ICA yields similar components for resting state (left) and task-related (right) data sets.
These components can be loosely associated with the distinct computations being undertaken in the task related dataset.
When doing an experiment that explicility manipulates some cognitive process, find same subset of networks in operation than rest
Showing different parts of brain don’t act alone but act in concert - regions in back work together with regions in front
With classic mass univarate analysis they compare
responses to conditions in each voxel independently
With multi-variate pattern analysis (MVPA) it compares
patterns of responses of many voxels to different conditions
What does this diagram of univarate analysis - show? - (2)
Estimate the signal change during condition A and estimate signal change during condition B and identify set of voxels where activation is significantly greater in condition A and B
In order to calculate pattern of voxels more active in A than B , looking at each voxel in turn and do same for every voxel in brain
What does this MVPA diagram show? - (3)
red areas more active on average than blue areas in different conditions in experiment
Look at different patterns and compare with each other
Use correlation to measure similarity of responses in different conditions in same part of brain in same set of voxels
What does this diagram show in MVPA? - (10)
Participants viewing different categories of images like objects, faces,places
Look at pattern of activity elicited in each task like top could be faces and bottom could be pattern of activity in place task
We can compare patterns of activity both within and between categories
Between = compare activity we seen during face condition to place condition and look whether those patterns correlate with each other
Within = Look at 2 different sets of trials from face condition and compare those and see if they correlate
They divide experiment and analyse pattern of activity in even and odd runs
What is typically found is if we look at same category stimuli (within) , there is positive correlation as same response to same stimuli
In between with face vs place we see negative correlation as pattern of response different to different stimulus type
Not looking at overall activity (represented by number) but correlation between 2 diff patterns and their similarity
But we can use MVPA to find out what the categories are (instead of defining above) so look at particular part of brain and say can brain tell diff between two diff tasks
One of the first studies that introduced idea of MVPA is by
Haxby et al., (2001)
Haxby et al 2001 methods for faces and houses and results - (7)
Had different classes of stimuli like faces, places, cats, scissors, chairs
Looked at pattern similarity
For every class of stimulus they compare within and between category responses in terms of pattern of activity they elicted (not overall activity but similarity between patterns)
e.g., faces look similarity in responses to 2 different sets of faces r= 0.81 or look at correlation between set of faces and houses which is negative ( r = - 0.40)
Pattern more similar within category than between as strong positive correlation with faces and modest negative correlation with face and houses as well as with other categories of stimuli over whole area of occipital temporal cortex even when FFA is removed
Similar pattern for houses over whole area even when PPA is removed
Suggests that not accurate to think of modular selective regions
What does this diagram show in MVPA? - (4)
If we consider many exp conditions
We can create a matrix between different pair of experimental conditions and how brain processes different tasks and if it processes similarliy or differently
Looking at pattern of activation greater in one condition than other
condition F and B more correlated than condition F and A so brain may doing something similar in condition F and B almost same tasks
What is representational similarity analysis? - (4)
Ask participant to experience different stimuli or tasks like responses to faces, umbrellas etc..
Look at pattern of activation in set of voxels (in table) and those patterns will have a pattern of similarity/disimilarity
Tend to calculate disimilarity tend to be 1 - correlation
So for any pair of stimulus condition think about overall similarity of patterns in the different conditions which allows to produce matrix across diff task conditions and use matrix to decide what the brain is doing in different tasks with various models/explanations
In RSA, can create many matrices such as - (5)
Matrix based on similarity of behavioural responses in given task as how long RT to sort faces vs objects
Does behavioural matrix similar to brain matrix as if it does then maybe behaviour is being driven by this part of brain that has similar RSA
Do same with computational model and think whether this part of brain is face detector which would respond similar way to different faces and respond disimilar to other categories
Or look at whether pattern of brain activity could be explained in characteristics of stimuli like low level like how similar images are in underlying pixels and another matrix
Which explanation fits best the pattern of similarity we are seeing in experiment
What does this diagram show in pattern classification? - (10)
researchers want
to identify voxels whose activation predicts whether the subject is looking at photographs of animals or plants.
At the first
stage, feature selection, the researchers identify a subset of voxels for subsequent analyses.
A typical feature set consists
of the activation intensity for each voxel on each trial.
The
feature set splits into a training set, from which the pattern classifier will be derived, and a testing set that provides a novel test of the generalization of the classifier.
Shown here
is a simplified example of pattern classification using two features (i.e., two voxels) and two trial categories (A and B).
The activation values of those two voxels on each trial are
shown as a two-dimensional plot.
Note that fMRI pattern
classification involves many more dimensions and thus a
much higher-dimensional space.
In the common technique of
support vector machines, the pattern classification algorithm attempts to identify the surface that maximally distinguishes
the two categories.
Here, a linear classifier optimally sepa- rates the two stimulus categories.
Once a classifier has been
identified, it is tested on the novel training set to ensure that the classification rule can be generalized to untested data.
Example of pattern classification in Haynes et al 2007 - (4)
Has 2 phases of word ‘select’ and after certain period of time participants given 2 numbers and have to decide to add or subtract the new numbers in green phase
Then see 2 numbers and add/subtract and then after a period of time see 4 answers and press button which one is right one
A computer is trained to recognize patterns of activity that predict whether participants will add or subtract (Haynes et al., Current Biology, 2007).
Using activity in the green areas predictions are up to ~70% accurate –> used search light technique by looking at chunk of brain
What is voxel modelling? - (5)
Another advanced technique of what is happening in brain and trying to use that to predict and decode what patterns of activity mean (similar to pattern classification but different) but works on individual voxels
Carry data of pps looking at picture for long time and produce diff brain activity
Using pattern of activity we have time courses that we can attempt to explain by modelling diff characteristics of images they are looking at
So for any given voxel say if that pattern of activity in voxel responding to activity to loads of different properities like responding to at top side of screen or to top left? Was it responding to surfaces? Semantics to movies?
We can create a model and estimate how model works and get best fitting model so present picture able to predict pattern of activity in each voxel
Collect second data from pps looking at diff model and using model we produced and generating predictions of pattern of activity in each voxel
Example of study using voxel modelling - Gallant - (3)
Showing movie to pps
Then different movie to pps
When the participant watches a new video, the model decodes what they are seeing by finding video clips that would produce similar patterns of activity and combining them.
