Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin...
Transcript of Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin...
![Page 1: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/1.jpg)
Session III: Types of MVPA Analysis
Martin HebartLaboratory of Brain and Cognition
NIMH
![Page 2: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/2.jpg)
MVPA Workflow
Statistical Analysis
Classification
Type of Analysis
Preprocessing
Design and Data Aquisition
![Page 3: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/3.jpg)
Overview
Types of MVPA Analysis: OverviewWholebrain AnalysisRegion of Interest AnalysisSearchlight Analysis
Multivariate Generalization (aka Cross-Decoding)
![Page 4: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/4.jpg)
TYPES OF MVPA ANALYSIS
![Page 5: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/5.jpg)
Types of MVPA Analysis
Wholebrain Region of Interest
time-resolvedSearchlight
![Page 6: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/6.jpg)
Wholebrain Analysis
All voxels are included into the classification
< All available spatial information is provided to the classifier
Problem of overfitting=
No information about location of information possible
=
< No “multiple comparisons” problem
![Page 7: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/7.jpg)
Wholebrain Analysis: Problem of Overfitting
Reduction of accuracy with large number of voxels
# voxel
Acc
urac
y
![Page 8: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/8.jpg)
Wholebrain Analysis: Problem of Overfitting
Possible solution: Use sparse models (e.g. LASSO)But: Sparse models find only subset of voxels with
information à we miss a lot of relevant voxelsAlternative solution: Use prior to regularize smoothness
But: unclear if prior is correct
Michel et al. (2011) – IEEE Trans Med Imaging
TV λ = 0.01 TV λ = 0.05 TV λ = 0.1
R2 = 0.83 R2 = 0.84 R2 = 0.84
Which oneto pick?
![Page 9: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/9.jpg)
Wholebrain Analysis
Weight maps: Solution to inability to localize information?
Mourao-Miranda et al. (2006) – Neuroimage
![Page 10: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/10.jpg)
Wholebrain Analysis
Problem: Voxel weights reflect signal (= source) and unspecific noise
Haufe et al. (2014) – Neuroimage; Hebart & Baker (2017) – Neuroimage
Voxel 2, w = 1
Voxe
l 1, w
= 0
Discriminative pattern
Voxel 2, w = 0.5
Voxe
l 1, w
= 0
.5
Measured data
Voxel 2
Voxe
l 1
Correlated noise
Weights do not indicate source of information
But: source of information can be reconstructed
![Page 11: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/11.jpg)
Region of Interest Analysis
Voxels are selected using spatial criteria
< Relatively precise localization of information
Selection of regions required in advance (bias?)=
Information encoded only across the distance gets lost
=
< Small ”multiple comparisons” problem
Limited comparability between ROIs=
![Page 12: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/12.jpg)
Region of Interest Analysis
ROIs with very different size difficult to compare
# voxel
Acc
urac
y
ROI 1 ROI 2 ROI 3
Haushofer et al (2008) – PLoS biol; Schreiber & Krekelberg (2013) – PLoS One
![Page 13: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/13.jpg)
Searchlight Analysis
Sphere of voxels is moved through brain as ROI
< Very precise localization of information
“multiple comparisons” problem=Ideal shape and size of searchlight unknown=
Informative regions possibly inflated=
< Maximization of local information
< Good for group comparisons analogous to SPM
Information encoded only across the distance gets lost=
![Page 14: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/14.jpg)
Train
Searchlight Analysis
Voxel 2Vo
xel 1
Class ,Sample 1
… …Class ,Sample 2
Class ,Sample 1
Class ,Sample 2
… … … …
Output (z.B. Accuracy: 60%)
Step 2: Extract local pattern from functional images
Step 3: Perform classification
Voxel 2
Voxe
l 1
Step 4: Fill-in classification output at central voxel in result image
Step 1: Select current searchlight position
Test
![Page 15: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/15.jpg)
Searchlight Analysis
Etzel et al (2013) – Neuroimage
• Searchlight analysis acts as spatial smoothing
• Reason: Neighboring searchlights overlap a lot, i.e. have very similar information
• Extreme example: One highly informative voxel creates high accuracy in size of searchlight
Related “problem” in spatial smoothing
![Page 16: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/16.jpg)
Interim Summary
• There are three main types of analysis: Wholebrain, ROI, and searchlight
• Each type has specific advantages and disadvantages• For most disadvantages there are potential solutions
![Page 17: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/17.jpg)
Multivariate Generalization
Approach that is growing in popularity to test association of cognitive states
Cross correlationCross classification
TrainingCognition A
Category A
Test Cognition B
Category B
Category A Category B
vs
vs
Category A Category B
Category A Category B
directed approach
Cognition A
Cognition B
undirected approach
![Page 18: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/18.jpg)
Multivariate Generalization
Multivariate equivalent to conjunction analysis
Conjunction analysis Generalization (multivariate)
LaBar et al (1999) – Neuroimage
Results: Attention and working memory activate the
same regions
ATT > 0 WM > 0
Result: Attention and working memory content are encoded
similarly
Attention Working memory
Accuracy
![Page 19: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/19.jpg)
Cross-Classification
Idea: Cross-classification is more specific than conjunction analysis, because pattern needs to be similar in both cases
Problem: Still trivial univariate effects possible Class A Class B
Voxel 1 Voxel 2
ROI
Part with univariate
effect
Part without univariate
effect
Voxel 1 Voxel 2
Condition 1 Condition 2
![Page 20: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/20.jpg)
Cross-Classification
Why does cross-classification sometimes work better one way than the other?
Condition 1 Condition 2
AccuracyCondition 2 Condition 1
Accuracy
Cond 2 has some information of Cond 1 in same direction but more in orthogonal direction
Hebart & Baker (2017) – Neuroimage
cross-classify
68 % accuracy
Condition 1 Condition 2
cross-classify
50 % accuracy
Orthogonality more likely in higher dimensionality (i.e. more voxels)
Condition 2 Condition 1
![Page 21: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/21.jpg)
Summary
• Generalization is a good approach for testing associations• Cross classification is directed, cross-correlation in
undirected• Trivial explanations remain possible
![Page 22: Session III: Types of MVPA Analysis · 2018. 12. 14. · Session III: Types of MVPA Analysis Martin Hebart Laboratory of Brain and Cognition NIMH. MVPA Workflow Statistical Analysis](https://reader035.fdocuments.net/reader035/viewer/2022071404/60f95a8b89b93d42ca34db50/html5/thumbnails/22.jpg)
Study questions
You are interested in studying consumer preferences (i.e. whether you would prefer buying A or B) and have some a priori assumptions about where in the brain you would expect this to be encoded.
Question 1: Compare these two approaches:(a) Run an analysis combining these ROIs and (if significant) removing
individual ROIs (so called “virtual lesioning” approach)(b) Run individual ROI analyses
Question 2: What are the pros and cons of restricting yourself to grey matter voxels only?