1. Imaging Source Reconstruction in the Bayesian Framework
Practical aspects of…
Imaging Source Reconstruction in the Bayesian Framework
Jason Taylor
MRC Cognition and Brain Sciences Unit (CBU)
Cambridge Centre for Ageing and Neuroscience (CamCAN)
19 January 2011 | Brussels | Thanks to Rik Henson & the CBU MEG community

2. Choices, choices
Selecting data (evoked/induced) & time-window (epoch/window)
Meshes and Forward Models (template? resolution?)
Inversion schemes (IID / MSP / other)
Group Inversion, Sensor Fusion, fMRI priors
Selecting time-window(s) for contrasts
Statistcs 2

3. Test Case: MEG Lexical Decision Data
Neuromag Mags/Grads (No EEG)
N=18
480 trials: 50% words, 50% pseudowords
Respond: Word/Non-word with button press, hand counterbalanced
Taylor & Henson, submitted 3

4. Mags
Grads
(RMS)

5. (Taylor & Henson, submitted)
The Entire
Analysis
Pipeline
(Taylor & Henson, submitted)

6. Selecting data to invert
SPM allows you to invert either trials (epochs) or averages
->Decide whether interested in evoked or induced (total)
->for induced: covariance is accumulated over trials
->allows selection of frequency window of interest
-- note: trial inversion can by memory-intensive if many trials 6

7. Meshes and Forward Models
(see Christoph’s talk this morning)
->For the present data, used inverse-normalised template cortical mesh (~7000 points) + individually defined inner-skull & scalp mesh BEM
IID
MSP
Can(ctx)
+Ind(skull)
Individual
(all)
Henson et al, 2009, NImage 7

8. Inversion scheme (MSP, IID, …)
Model comparison approach: Don’t need to decide a priori
->But consider:
- Expect focal or distributed sources?
(MSP better captures focal; IID more appropriate for distributed?)
- Individual accuracy vs. group consistency?
(Maximising individual accuracy may come at the expense of consistency over subjects – if localisation/anatomy is variable)
- Distributional assumptions of stats?
(Sparse solutions tend not to be Gaussian - recall stats talk) 8

9. Group Inversion, Sensor Fusion, etc.
Model comparison approach also works for Individual vs. Group, and for the addition of priors from e.g. fMRI
To compare Separate Sensor vs Fusion inversions, however, the data have changed, so model comparison does not apply
Group: Litvak & Friston, 2008, Nimage
Fusion: Henson et al, 2009, NImage 9

10. Step 1: Individual subject/sensor inversions
Taylor & Henson, submitted

11. Step 2: Fusion of Sensor Types
Taylor & Henson, submitted

12. Taylor & Henson, submitted
fMRI
priors
Taylor & Henson, submitted

13. Taylor & Henson, submitted
MSP >> IID (note difference in scales between two plots)
Group inversion doesn’t affect IID (nothing to optimise);
Group inversion decreases (though n.s.) MSP model evidence (at individual level); trade-off of individual accuracy and group consistency
(3) fMRI priors improve IID but not MSP (presumably fMRI blobs already covered by patches in MSP)
Taylor & Henson, submitted

14. Selecting time-windows for contrasts
Our approach: Use sensor stats to constrain/inform source analysis
->Identify time-windows of interest
->divide into sub-windows based on hierarchical cluster analysis 14

15. Taylor & Henson, submitted

16. Taylor & Henson, submitted

17. Taylor & Henson, submitted

18. Statistics
Discussed yesterday: Sparse source images tend not to be Gaussian (e.g., MSP)
SPMs, PPMs, SnPMs
Trade-off?: individual accuracy vs. group consistency 18

19. Taylor & Henson, submitted
fMRI priors
When several fMRI (or other) priors are entered separately, each may be up- or down-weighted
Different priors may be endorsed for different subjects
Group optimisation reduces these inter-subject differences
Taylor & Henson, submitted

20. Statistics
Discussed yesterday: Sparse source images tend not to be Gaussian (e.g., MSP)
Also yesterday: SPMs, PPMs, SnPMs
Trade-off?: individual accuracy vs. group consistency
Factorise time: allows inferences about emergence/disappearance of effects 20

21. Taylor & Henson, submitted
Condition X
Time-Window
Interactions
Taylor & Henson, submitted

22. - The End -
Thanks!