1. EEG/MEG source reconstruction in SPM5
Jérémie Mattout / Christophe Phillips / Karl Friston
With thanks to
John Ashburner, Guillaume Flandin, Rik Henson, Stefan Kiebel

2. Outline Introduction - EEG/MEG inverse problem I - Source model
- 3D reconstruction in SPM5
I - Source model
II - Data registration
III - Head model and forward computation
IV - Inverse estimation
Demo

3. Introduction - EEG/MEG inverse problem

4. Introduction - EEG/MEG inverse problem
Jacques Hadamard ( )
Existence
Unicity
Stability
“Will it ever happen that mathematicians will know enough about the physiology of the brain, and neurophysiologists enough of mathematical discovery, for efficient cooperation to be possible?”

5. Introduction - EEG/MEG inverse problem
Forward problem (well-posed)
Y = K(J) + E
Inverse problem (ill-posed)
Data Y
Current density J
Bayesian framework
incorporate multiple constraints/prior information
estimate the optimal contribution of those priors
evaluate the relevance of the priors/model
Parametric empirical Bayes
Bayesian model comparison

6. Introduction - 3D Reconstruction in SPM5
Preprocessing
Projection
SPM5-engine
EEG/MEG Raw data
2D - scalp
SPM{t}
SPM{F}
Mass univariate
analysis
Single Trials
- epoching
- artefacts
- filtering
- averagin
3D - brain
DCM
spm_eeg_inv_*.m

7. Introduction - 3D Reconstruction in SPM5
Sources
MEG data
3D Projection
‘Equivalent Current Dipoles’ (ECD)
‘Imaging’
EEG data

8. Introduction - 3D Reconstruction in SPM5
ECD
Imaging
(1) Source model
(3) Forward model
(2) Registration
Data
(4) Inverse method
Anatomy

9. Introduction - 3D Reconstruction in SPM5
data: [151x2188x5 spm_file_array]
channels: [1x1 struct]
scale: [1x1 struct]
filter: [1x1 struct]
events: [1x1 struct]
reref: []
descrip: []
datatype: 'int16'
fname: 'fmbe_emer01_TCS.mat'
fnamedat: 'fmbe_emer01.dat'
Nchannels: 151
Nevents: 5
Nsamples: 2188
Radc: 625
path: [1x76 char]
inv: {1x7 cell}
modality: 'MEG'
D = spm_eeg_ldata;
Data structure
D.inv{1} =
method: 'Imaging'
mesh: [1x1 struct]
datareg: [1x1 struct]
forward: [1x1 struct]
inverse: [1x1 struct]
comment: {'MN + Smoothness'}
date: [2x11 char]

10. Outline Introduction - EEG/MEG inverse problem I - Source model
- 3D reconstruction in SPM5
I - Source model
II - Data registration
III - Head model and forward computation
IV - Inverse estimation
Demo

11. I - Source Model (Meshes)
Compute transformation T
Individual MRI
wmeshTemplate_3004d.mat
- wmeshTemplate_4004d.mat
- wmeshTemplate_5004d.mat
- wmeshTemplate_7004d.mat
Templates
Apply inverse transformation T-1
Individual mesh
input
functions
output
Individual MRI
Template mesh
spatial normalization into MNI template1
inverted transformation applied to the template mesh2
inner-skull and scalp binary masks
cortical mesh
inner-skull mesh
scalp mesh
1Unified segmentation, J. Ashburner and K.J. Friston, NeuroImage, 2005.
2Canonical source reconstruction for EEG & MEG, J. Mattout and K.J. Friston, in preparation.

12. I - Source Model (Meshes)
D.inv{1} =
method: 'Imaging'
mesh: [1x1 struct]
datareg: [1x1 struct]
forward: [1x1 struct]
inverse: [1x1 struct]
comment: {'MN + Smoothness'}
date: [2x11 char]
D.inv{1}.mesh =
sMRI: [1x87 char]
nobias: [1x86 char]
def: [1x94 char]
invdef: [1x98 char]
msk_iskull: [1x92 char]
msk_scalp: [1x91 char]
msk_flags: ''
tess_ctx: [1x95 char]
Ctx_Nv: 4004
Ctx_Nf: 8000
tess_iskull: [1x108 char]
Iskull_Nv: 2002
Iskull_Nf: 4000
tess_scalp: [1x106 char]
Scalp_Nv: 2002
Scalp_Nf: 4000
CtxGeoDist: [1x101 char]

13. Outline Introduction - EEG/MEG inverse problem I - Source model
- 3D reconstruction in SPM5
I - Source model
II - Data registration
III - Head model and forward computation
IV - Inverse estimation
Demo

14. Rigid transformation (R,t)
II - Data Registration
fiducials
fiducials
Rigid transformation (R,t)
Landmarks (MEG/EEG)
ICP Surface matching (EEG)
EEG/MEG
sensor space
MRI space
input
sensor locations
fiducial locations
(in sensor & MRI space)
structural MRI
(scalp mesh)
output
functions
registered data
transformation matrix
registration of the EEG/MEG data into MRI space3
3A method for registration of 3d-shapes, P.J. Besl and N.D. McKay, IEEE Trans. Pat. Anal. And Mach. Intel., 1992.

15. II - Data Registration D.inv{1} = method: 'Imaging' mesh: [1x1 struct]
datareg: [1x1 struct]
forward: [1x1 struct]
inverse: [1x1 struct]
comment: {'MN + Smoothness'}
date: [2x11 char]
D.inv{1}.datareg =
sens: [1x98 char]
fid: [1x94 char]
fidmri: [1x94 char]
hsp: ''
scalpvert: ''
sens_coreg: [1x104 char]
fid_coreg: [1x100 char]
hsp_coreg: ''
eeg2mri: [1x87 char]

16. Outline Introduction - EEG/MEG inverse problem I - Source model
- 3D reconstruction in SPM5
I - Source model
II - Data registration
III - Head model and forward computation
IV - Inverse estimation
Demo

17. III - Head model & Forward computation
Compute for
each dipole p + n
Forward operator
MRI space
Head model
functions
input
single sphere
three spheres
overlapping spheres
realistic spheres
output
sensor locations
cortical mesh
scalp mesh
forward operator
BrainSTorm

18. III - Head model & Forward computation
D.inv{1} =
method: 'Imaging'
mesh: [1x1 struct]
datareg: [1x1 struct]
forward: [1x1 struct]
inverse: [1x1 struct]
comment: {'MN + Smoothness'}
date: [2x11 char]
D.inv{1}.forward =
bst_options: [1x1 struct]
bst_channel: [1x100 char]
bst_tess: [1x97 char]
gainmat: [1x103 char]
pcagain: [1x107 char]

19. Outline Introduction - EEG/MEG inverse problem I - Source model
- 3D reconstruction in SPM5
I - Source model
II - Data registration
III - Head model and forward computation
IV - Inverse estimation
Demo

20. IV - Parametric Empirical Bayes (Inverse) 2-level hierarchical model
Single trial
Gaussian variables
with unknown variance
Sensors
Sources
Linear parameterization of the variances
Q: variance components
: hyperparameters

21. + + IV - Parametric Empirical Bayes (Inverse)
Bayesian inference on model parameters + +
Model M
Maximizing the log-evidence
data fit
priors
Expectation-Maximization (EM)
E-step: maximizing F wrt J
MAP estimate
M-step: maximizing of F wrt
ReML estimate
Bayesian Model Comparison ?
Log(Bayes factor) = F1-F21
Inference
4Comparing dynamic causal models, W.D. Penny, K.E. Stephan, A. Mechelli, K. Friston, NeuroImage, 2004.

22. IV - Parametric Empirical Bayes (Inverse)
Evoked and induced activity
Events s  t t
Synchronized oscillations in time,
but not in phase with the stimulation FT
Average
Evoked resp.
Induced resp. - =

23. IV - Parametric Empirical Bayes (Inverse)
data & constraints
Multiple trials
evoked energy
induced energy

24. IV - Parametric Empirical Bayes (Inverse)
Example
100
200
300
400
time (ms)
Right temporal evoked signal
faces
scrambled
M170
Energy changes (Faces - Scrambled, p<0.01)
0.1
0.2
0.4
0.6
0.8
time (s) 10 20 30 40 35 45 15 25
0.7
0.5
0.3
-0.1 1 2 3 -2 -3 -1
frequency (Hz)
Time-frequency subspace
200
time (ms)
400
MEG experiment
of Face perception4
4Electrophysiology and haemodynamic correlates of face perception, recognition and priming, R.N. Henson, Y. Goshen-Gottstein, T. Ganel, L.J. Otten, A. Quayle, M.D. Rugg, Cereb. Cortex, 2003.

25. IV - Parametric Empirical Bayes (Inverse)
Example

26. IV - Parametric Empirical Bayes (Inverse)
Example

27. IV - Parametric Empirical Bayes (Inverse)
input
functions
output
preprocessed data
- forward operator
mesh
constraints
- compute the MAP estimate of J1
compute the ReML estimate of 1
model evidence2,4
source dynamic1,2
power3
1An empirical Bayesian solution to the source reconstruction problem in EEG, C. Phillips, J. Mattout, M.D. Rugg, P. Maquet and K.J. Friston, NeuroImage, 2005.
2MEG source localization under multiple constraints: an extended Bayesian framework, J. Mattout, C. Phillips, M.D. Rugg and K.J. Friston, NeuroImage (in press).
3Bayesian estimation of evoked and induced responses, K.J. Friston, R.N. Henson, C. Phillips and J. Mattout, Hum. Brain Mapp. (in press).
4Variational free energy and the Laplace approximation, K.J. Friston, J. Mattout, N. Trujillo-Barreto, J. Ashburner and W. Penny (in preparation).

28. IV - Parametric Empirical Bayes (Inverse)
D.inv{1} =
method: 'Imaging'
mesh: [1x1 struct]
datareg: [1x1 struct]
forward: [1x1 struct]
inverse: [1x1 struct]
comment: {'MN + Smoothness'}
date: [2x11 char]
D.inv{1}.inverse =
activity: 'evoked'
contrast: [ ]
woi: [ ]
priors: [1x1 struct]
dim: 4004
resfile: 'fmbe_emer01_TCS_remlmat_150_190ms_evoked_11H3.mat'
LogEv: e+003