Automatic Generation of Volume Conductor Models of the Human Head for EEG Source Analysis Benjamin Lanfer WWU M¨ unster / BESA GmbH BaCI 2015
Automatic Generation of VolumeConductor Models of the HumanHead for EEG Source Analysis
Benjamin LanferWWU Munster / BESA GmbH
BaCI 2015
EEG Source Analysis and Volume Conductor Models
FEM (BEM, FDM, ...) solution ofquasi-static Maxwell equations
Volume conductormodel
Inverse Problem
Forward Problem
B. Lanfer ([email protected]) 1 / 13
Generation of Individual, Realistic Head Models
B. Lanfer ([email protected]) 2 / 13
What Do We Know About the Segmentation?
A-priori Knowledge
I Exploiting a-priori knowledge about . . .I . . . arrangment of head tissuesI . . . occurence of tissues at locations relative to (anatomical) reference
surfaces
B. Lanfer ([email protected]) 3 / 13
Segmentation in a Bayesian Framework
I Bayesian a-posteriori probability as measure for good segmentation
P (x |y) ∝ l(y |x)︸ ︷︷ ︸Likelihood
P (x)︸ ︷︷ ︸A-priori Probability
I Likelihood: how well does the current segmentation explain theobserved image?
l(yi |xi, λc) =1√
(2π)k|Σc|exp
{−1
2(yi − µc)
ᵀΣ−1c (yi − µc)
}
B. Lanfer ([email protected]) 4 / 13
The Markov Random Field Model
B. Lanfer ([email protected]) 5 / 13
The Markov Random Field Model (cont.)
I Markov Random Field (MRF)
P(xi |xS\{i}
)= P (xi |xNi) Markovianity
I Each MRF is equivalent to a Gibbs Random Field
P (x) =2
Zexp
(− 1
TU(x)
)Gibbs distribution
B. Lanfer ([email protected]) 6 / 13
The Markov Random Field Model (cont.)
I Gibbs energy function as a sum of single-site and pairwise cliquepotentials V1, resp., V2
U(x) =∑i∈S
V1(i, xi) +∑i∈S
∑i′∈Ni
V2(i, i′, xi, xi′
)+ . . .
I Definition of pairwise clique potentials using pseudo transitionprobabilities Pxi,xi′ (i, i
′)
V2(i, i′, xi, xi′) = − ln
(Pxi,xi′ (i, i
′))
BGSkin /Muscle
SCT
CoB
CaB
CSF
Dura
GM WM
B. Lanfer ([email protected]) 7 / 13
The Atlas-Based A-priori Probability
B. Lanfer ([email protected]) 8 / 13
Atlas Generation and Projection
I 20 labeled, template images from BrainWeb database1
I Averaging and normalization of local tissue histograms
I Atlas: local tissue probability mass functions depending on distancesto reference surfaces
I Projection to individual reference surfaces
Template image Local histogram
1Aubert-Broche et al., NeuroImage, 2006
Atlas Generation and Projection
I 20 labeled, template images from BrainWeb database1
I Averaging and normalization of local tissue histograms
I Atlas: local tissue probability mass functions depending on distancesto reference surfaces
I Projection to individual reference surfaces
Individual MRI and reference surfaces Local probability mass function
1Aubert-Broche et al., NeuroImage, 2006
The Segmentation Algorithm
Validation vs. Manual Raters
... ... ......
Proposedapproach
Majority
vote
Averaging
B. Lanfer ([email protected]) 11 / 13
Results Validation vs. Manual Raters
B. Lanfer ([email protected]) 12 / 13
Summary and Outlook
Summary
I Accurate segmentation of the four most relevant tissues for EEGsource analysis
I Low effort enables wider application of individual, realistically shapedFEM models in EEG source analysis, tDCS simulations, . . .
Outlook
I Improved segmentation of the skull base and the facial skull, e.g.,using templates
I Important for high-density electrode caps, temporal lobe activity
I Treatment of pathological anatomies (lesions, skull trepanation holes)
B. Lanfer ([email protected]) 13 / 13
Thank you!
WWU / IBB
I PD Dr. Carsten H. Wolters
I Prof. Dr. Martin Burger
I Prof. Dr. Christo Pantev
I Umit Aydin
I Felix Lucka
I Johannes Vorwerk
I Sven Wagner
I Dr. Harald Kugel
BESA GmbH
I Dr. Michael Scherg
I Dr. Tobias Scherg
I Theo Scherg