EEG-MEG source reconstruction 2

EEG-MEG source reconstruction
Jean Daunizeau
Wellcome Trust Centre for Neuroimaging
23 / 10 / 2009
1
EEG/MEG data
sensor locations
structural MRI
• anatomical templates
• data convert
• epoching
• BEM forward modelling
• spatial denormalisation
gain matrix
trials
• baseline correction
• averaging over trials
• low pass filter (20Hz)
individual
meshes
evoked
responses
• inverse modelling
• 1st level contrast
cortical
sources
• standard SPM analysis
2
EEG/MEG data
sensor locations
structural MRI
• anatomical templates
• data convert
• epoching
• BEM forward modelling
• spatial denormalisation
gain matrix
trials
• baseline correction
• averaging over trials
• low pass filter (20Hz)
individual
meshes
evoked
responses
• inverse modelling
• 1st level contrast
cortical
sources
• standard SPM analysis
3
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
1. Introduction
2. Forward problem
3. Inverse problem
4. Bayesian inference applied to distributed source reconstruction
5. SPM variants of the EEG/MEG inverse problem
6. Conclusion
4
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Forward and inverse problems: definitions
 Forward problem = modelling
 Inverse problem = estimation of the model parameters
5
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Physical model of bioelectrical activity
current dipole
6
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Fields propagation through head tissues
noise
dipoles
gain matrix
measurements
Y = KJ + E1
7
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
An ill-posed problem
Jacques Hadamard (1865-1963)
1. Existence
2. Unicity
3. Stability
8
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
An ill-posed problem
Jacques Hadamard (1865-1963)
1. Existence
2. Unicity
3. Stability
9
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Imaging solution: cortically distributed dipoles
10
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Imaging solution: cortically distributed dipoles
11
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Regularization
Spatial and
temporal
constraints
Adequacy with
other
modalities
Data fit
data fit
constraint
(regularization term)
W = I : minimum norm method
W = Δ : LORETA (maximum smoothness)
12
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Priors and posterior
likelihood
posterior
priors
model evidence
13
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Hierarchical generative model
sensor level
source level
Q : (known) variance components
(λ,μ) : (unknown) hyperparameters
14
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Hierarchical generative model: graph
λ1
λq
J
μ1
Y
μq
15
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Variational Bayesian inversion (VB, EM, ReML)
ln p  y m   ln p   , y m 
q

 S  q   D K L q   ; p  y , m 

free energy : functional of q
 approximate (marginal) posterior distribution:
 q    , q   
1
2
p  1 ,  2 y , m 
p  1 or 2 y , m 
2
q  1 o r 2 
1
16
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Imaging source reconstruction in SPM
IID
COH
generative model M
ARD/GS
prior covariance structure
17
Introduction
Forward
Inverse
Bayes
SPM
Source reconstruction
for group studies
Conclusion
Group studies
canonical meshes!
18
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Equivalent Current Dipoles (ECD)
Somesthesic stimulation (evoked potential)
soft symmetry constraints!
ECD moments
prior precision
ECD positions
prior precision
ECD
moments
EEG/MEG data
ECD
positions
measurement noise
precision
19
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Dynamic Causal Modelling (DCM)
macro-scale
meso-scale
Golgi
micro-scale
Nissl
external granular
layer
EI
external pyramidal
layer
PC
internal granular
layer
action potentials
generation zone
internal pyramidal
layer
II
firing rate
membrane potential (mV)
synapses
membrane potential (mV)
time (s)
20
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
21
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
• EEG/MEG source reconstruction:
1. forward problem;
2. inverse problem (ill-posed).
• Prior information is mandatory
to solve the inverse problem.
• Bayesian inference is well suited for:
1. introducing such prior information…
2. … and estimating their weight wrt the data
3. providing us with a quantitative feedback
on the adequacy of the model.
22
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
individual reconstructions in MRI template space
L
R
2nd level group analysis
R
L
RFX analysis
p < 0.01 uncorrected
23
Introduction
Forward
Inverse
Bayes
SPM
Conclusion
Many thanks to
Karl Friston, Stephan Kiebel, Jeremie Mattout and Vladimir Litvak
24