Please enable JavaScript.
Coggle requires JavaScript to display documents.
fMRI: Data Analysis & Applications - Coggle Diagram
fMRI: Data Analysis & Applications
Data Analysis Theory
Simple Block Design fMRI Experiment
Intermittent stimulus presentation
fMRI measures volumes to create slices made of voxels
Issues w. What fMRI Data Look Like
Noisy
Poor spatial resolution
Arbitrary units
Enormous amounts
One subject data 1GB+
Noisy
Need to average/smooth data
Poor Spatial Resolution
Need to align w. anatomical MRI to localise w. any precision
Arbitrary Units
Only relative changes measured
Enormous Amounts
One time series (100 - 500 points) for each voxel
~100,000 per volume
Model-Driven fMRI Analysis
fMRI data is one measurement over time (time series) for each voxel
Most analyses based on performing standard statistical test independently & separately for each voxel
Model-Driven fMRI Analysis Strategies
Whole-brain analysis
Region-of-interest analysis
Region-of-Interest Analysis
Whare are responses of pre-defined part of brain/region of interest
Whole-Brain Analysis
Where in brain some process takes place
Brain mapping/voxel hunting
Statistical Test
Test how strongly the activity of voxel varies
W. stimulus/task
Testing each voxel known as mass univariate
Produce one statistical test result for each voxel
Statistical Tests
Commonly used are generalised linear models
Correlation/regression
ANOVA
t-test
ANCOVA
Principle
Stimulus covariate = Predicted voxel response corresponding to stimulus/task mapped onto graph
fMRI response in one voxel mapped onto graph
Produced result of statistical test that is overlaid on data
Statistical parametric map
Creating Covariates/Regressors
Stimulus * HRF = Stimulus covariate
Multiple Covariates/Design Matrix
Covariates for each stimulus/task
Interpreting Design Matrix
Visual & mathematical representation of experimental stimulus/design task
Used for actual statistical computations (essentially a type of multiple regression)
Each column (covariate) corresponds to one stimulus/task
Represents (models) the predicted response to stimulus/task
Each row is time point
Creating Covariates
Smooth (convolve) stimulus/task sequence w. haemodynamic response (HRF)
Additional columns represent head movement (noise covariate), behaviour measurements, physiological covariates ect
Results of fMRI Stat Analyses
Statistical parametric map created
1+ images containing test statistic & significance value for each voxel & task/stimulus covariate
Statistical parametric map indicates for each voxel how strongly that responded to stimulus/task
High statistic w. low p-value in voxel indicates neurones w./in voxel responded to stimulus/task
Looking at voxels w. high significance can find brain regions that responded to stimulus/task
Limitation
Standard stat techniques insufficient
Multiple comparisons issue
Spatial autocorrection
Temporal autocorrection
Multiple Comparison Problem
Conventional stat thresholds would expect false positives everywhere
Single fMRI contain ~100,000 voxels
As significance p < .05 5,000 voxels significant by chance alone
Dealing With
Bonferroni correction
False discovery rate correction
Gaussian random field models
Stimulation/permutation methods
Clusters
ROI analysis
Bonferroni Correction
Divide alpha level by no. measurement
But too conservative
False Discovery Rate Correction
Less conservative
Gaussian Random Field Models
Estimate how many independent measurements have if know surrounding voxels may be correlated
Look at how random connections in space act if know they correlate w. each other
Requires smoothing images
Simulations/Permutation Methods
Estimate likelihood of false positive
Use Gaussian theory
Clusters
Search clusters activated voxels rather than single voxels
But still need methods to estimate significance
ROI Analysis
Independently identify small no. regions of interest
Perform standard stat analysis on data averages across voxels in each ROI
Advantage
Remove multiple comparisons problem
Only one data series
Increase stat power
Average across all voxels in ROI
ROIs hypothesis driven & can be defined anatomically/functionally
Spatial Autocorrection
Nearby voxels not stat independent
Temporal Autocorrection
Nearby time points not stat independent
Spatial & Temporal Correlations
Nearby voxels correlated
So cannot treat voxels as independent measures
Nearby points in time correlated
So cannot treat individual time points as independent
Rectify Spatial Issues
Techniques for multiple comparisons/averaging voxels for ROI analysis
Can make rough assumption that nearby voxels act in same way
Look for clusters of voxels
Activated regions tend to spread across multiple voxels
Rectify Temporal Issues
Remove temporal correlations by pre-processing
Permutation stats
Stats on averaged responses to single trials
Data Analysis Practical Steps
Steps to Analyse fMRI Data
Pre-processing
Whole-brain analysis
ROI analysis
Stat analysis
Visualisation/plotting results
Pre-Processing
Unwarping
Motion correction
Slice timing correction
Temporal filtering
Coregistration of fMRI & anatomical MRI data
Unwarping
Taking images v. quickly & way that data acquired introduces distortions
Particularly pronounced at edges of brain & where is other matter like ears & eyes
Inhomogeneities in static magnetic field cause distortions on EPI (BOLD) images
Cause misregistration w. anatomical images
Magnetic field measures using specific pulse sequences
Field map used to correct distortions
Motion Correction
Head motion the most common & sig source of noisy data
Intensity challenges due to motion can be much larger than task-related effects
Motion correction software shifts all volumes to same position
But best to avoid motion
Limitation
Issue is head movement correlated w. experiment
Slice Timing Correction
Each MR volume collected as series of slices
In sequence/interleaved
Diff in time between first and last slice the close to TR
Diff is problem when looking at responses to brief events
Shifts data in each slice to common time point
As if has all been acquired at same time
Co-Registration w. Anatomical Data
BOLD images have poor tissue contrast & coarse resolution
Make localisation difficult
Functional data overlaid on high-resolution anatomical images using software
Temporal Filtering
fMRI data noisy
Noise often 10x larger than signal
Types of Noise
Low frequency
High frequency
Low Frequency
Slow modulation
Use highpass filtering
Remove slow noise
Types of Low Frequncy
Physiological
Physical
Physiological
Breathing
Attention
Physical
Heating of gradient coils
High Freqnecy
Rapid modulation
Use lowpass filtering
Remove fast noise
Types of High Frequency Noise
Physiological
Physical
Physiological
Pulse
Breathing
Physical
Instrument noise
Whole-Brain Analysis
Spatial smoothing
Spatial normalisation
Spatial Smoothing/Filtering
Average voxels w. neighbours across space
Increase signal to noise at expense of resolution
Highly undesireable
But necessary for applying random field correction for multiple comparisons
But necessary for spatial averaging even if not recommended
Issue
Can introduce errors
If smoothing across sulci
Mainly legacy of older techniques (PET)
Should not be used unless necessary
Spatial Normalisation
Average data across subjects
Methods
Spatial normalisation & coregistration
Surface based
Spatial Normalisation & Co-Registration
Scale & warp individual brains to match each other
Assume brains differ only in overall shape & side
Not in finescale organisation of cortical areas
Issue
Assumption fundamentally incorrect
Most common approach but not recommended
Only works if images severely smoothed/blurred
Effective resolution of result is >1cm
Origin in PET where signal to noise so low that intersubject averaging necessary
Surface-Based Normalisation
Inflate cortical surface to sphere & register
So patterns of sulci & gyri align
Advantage
Respect cortical anatomy
No blurring across sulci
More accurate
Limitation
Assume sulcal pattern corresponds to function
Not necessarily true
ROI Analysis
Identify ROIs
Average Data Across Subjects
Identify ROIs in each subject & analyse data for each ROI across subjects
Use mapping to identify areas/identify brain regions responding to desired area in each subject
Average fMRI data across all voxels for each individually defined ROI
Run intersubject analysis on data using standard stat methods
Advantage
Encourage hypothesis-driven experiments rather than blind search for changes
Limitation
Require independent ways to identify ROIs
Stat Analysis
Voxel-by-voxel/each ROI
Visualisation/Plotting
On anatomical images
On anatomical surfaces
Can make organisation clearer
As graphs for ROI
Can be more informative than image
Whole-Brain fMRI Analysis Steps
Pre-processing
Stat analysis
Stat inference
Pre-Processing
Image-time series goes through motion correction & coregistration to remove artefacts from head motion
Filter goes through smoothing
Template, motion correction & coregistration goes through normalisation to bring to same standard spatial format
Motion correction & coregistration, & normalisation goes through smoothing
Stat Analysis
Design matrix & pre-processing go through stat analysis
Parameter estimates for each voxel
Visualise parameters for every voxel
Stat parametric map
Stat Inference
Gaussian random field theory/permutation/simulation
Decrease false positives
Estimate significance of results
Image generated w. colour indicating peaks & significance
ROI Analysis Steps
Pre-processing
Stat analysis
Stat inference
Pre-Processing
Image time-series goes through motion correction & coregistraton
Other fMRI data/anatomical info goes through identification of ROI
Motion correction & coregistration, & identification of ROIs goes through extract ROI data
Stat Analysis
Design matrix & preprocessing goes through stat analysis
Parameter estimates for each ROI
Plot parameter estimates for each ROI
Stat Inference
Standard parametric/non-parametric stats
Estimate significance of results
What Learned
Software for fMRI Analysis
Stat analysis
Visualisation
Stat Ana;ysis
Stat parametric mapping (FIL/UCL)
FSL (Oxford)
AFNI
BrainVoyager (commercial)
mrLoadRet/mrVISTA/mrTools
Visualisation
FreeSurfer
Caret
SurfRelax
fMRI the New Phrenology
fMRI widely used & growing fast
No sign of slowdown
Accused of being new phrenology/neurophrenology in pseudocolour
Often makes media headlines often w. dubious claims
Issue whether is meaningless or just used impoperly
Learnt w. fMRI in Human Visual Cortex
Prior to 1990 knowledge of human visual cortex structure & function v. limited
Believed to be scaled up version of macaque visual cortex
By 2010, 10-20 new visual areas discovered by fMRI
Many unique to humans
Function of areas increasingly understood
Importance of fMRI
No other technique could have made such developments in such short time
Many advances made possible by prior knowledge of macaque visual cortex function
How to Use fMRI
Difficult so learn to use right
Remember correlation not causation
Not just issue w. fMRI
Avoid testing black-box models of brain function
Remember limits of fMRI
Right Use
Require expertise in physics, neuroscience, psychology, & stats
Rare skill combination
Easy to run experiments but difficult to interpret & understand all processing steps fully
Design is crucial
Standard group analysis methods questionable
Black-Box Models
Cog models unlikely to map onto neural function in one-to-one fashion
Limits of fMRI
Population-based measure
May primarily measure synaptic activity
Input & local processing vs spiking output
Relationship between spikes/synaptic activity/BOLD not necessarily constant
Coarse temporal & spatial resolution
Need to understand limitations to use correctly
Accept many q's cannot be answered by fMRI
Be critical, especially of fMRI in media
Magnetic Resonance Spectroscopy