30th Annual International IEEE EMBS Conference
Vancouver, British Columbia, Canada, August 20-24, 2008
ICA Based Multiple Brain Sources Localization
Yongjian Chen, Masatake Akutagawa, Masato Katayama, Qinyu Zhang and Yohsuke Kinouchi
Abstract—In this paper we describe an Independent
Component Analysis (ICA) method for computing the brain
signals of unknown source parameters for the inverse problem.
First, a method is applied to estimate the number of dipoles
beforehand and reduce dimensionality which can reduce the
ICA complexity and improve the unmixing accuracy. We apply
Blind Source Separation (BSS) for separating multichannel
EEG evoked by multiple dipoles into temporally independent
stationary sources. For every independent source, we are able
to determine the electrode potentials evoked by every dipole
separately by the projection of independent activation maps
back onto the electrode arrays. Then for every set of electrode
potentials, a source localization procedure is performed which
only involves searching for one dipole, thus dramatically
reducing the search complexity.
In the paper, it is explored that the possibility of applying
ICA for EEG multiple dipoles localization when the data are
corrupted by additive noise. Furthermore, we also give the
relationship of unmixing accuracy, distance between dipoles
and dipoles moment movements.
I. INTRODUCTION
SOURCE localization draws on a wide range of signal
processing techniques including digital filtering,
three-dimensional image analysis, array signal processing,
image modeling and reconstruction, and Blind Source
Separation (BSS) and phase synchrony estimation. By far,
some general methods (e.g. Neural Network) are difficult to
solve 3 or more sources localization problem, and other
methods (e.g. MUSIC) are time consuming. In this paper, we
consider a spatio-temporal model, and apply it to the entire
EEG time series to reduce the extent of the dipoles
configuration space. So in this work we apply Independent
Component Analysis (ICA) method.
In the Inverse Problem, we need to estimate the number of
unknown brain sources before locating the dipoles. It is
usually determined by iteratively adding one dipole at a time
until a reasonable fit to the data has been found. Even when
restricting the location of the dipole to the lattice sites, the
configuration space is factorially large. So it is a bottleneck
Manuscript received April 7, 2008.
Yongjian Chen is with Graduate School of Advanced Technology and
Science, The University of Tokushima, Tokushima, Japan (phone:
81-090-6287-8886; e-mail: cyj622@ee.tokushima-u.ac.jp).
Masatake Akutagawa is with The University of Tokushima, Tokushima,
Japan (phone: 81-088-656-7475; e-mail: makutaga@ee.tokushima-u.ac.jp).
Masato Katayama is with Graduate School of Advanced Technology and
Science, The University of Tokushima, Tokushima, Japan (phone:
81-088-656-7476; e-mail: zmm76933@ee.tokushima-u.ac.jp).
Qinyu Zhang is with Harbin Institute of Technology, Shenzhen, China
(phone: 86-755-2603-3786; e-mail: zqy@hit.edu.cn).
Yohsuke Kinouchi is with The University of Tokushima, Tokushima,
Japan (phone: 81-088-656-7475; e-mail: kinouchi@ee.tokushima-u.ac.jp).
978-1-4244-1815-2/08/$25.00 ©2008 IEEE. 1879
of many localization procedures [1,2]. In the paper, before
ICA processing, we apply a method to estimate the number
of dipoles that can reduce the ICA complexity and improve
the unmixing accuracy. In Section B, the analysis of the
singular value is used to estimate the number of sources
beforehand.
Principal Component Analysis (PCA) is a well-known
linear transformation that can minimize the reconstruction
error. In Section C, PCA is applied to pretreat the
experimental data correctly. Especially, based on the number
of sources, dimensionality reduction is performed in order
that ICA can obtain independent sources correctly.
BSS is a technique for recovering unknown source signals
from known observed mixtures, without knowledge on the
mixing process. ICA is a method used for solving the BSS
problem, based on the assumption that the sources are
statistically independent and the mixing is time-invariant and
linear [3-5]. The aim of the present work is the application of
the ICA technique in order to test the technique’s ability in
multiple brain sources localization. In Section D, an ICA
algorithm based on the Infomax principle [3, 6] is briefly
introduced, and applied to EEG recordings. In Section E, we
estimate the electrode potentials due to each dipole
separately by the projection of independent activation maps
back onto the electrode arrays. Then for every set of
electrode potentials, we only need to search for one dipole,
thus dramatically reducing the configuration space. For each
electrode potential due to single dipole, we choose an instant
in time from the signal and employ a downhill simplex
search method [7] to determine the dipole.
II. METHODS
We assume that the sources can be represented by n
equivalent current dipoles of unknown location and moment,
and EEG consists of simultaneous recordings at m electrodes
for k time instances. Solving the inverse problem is to find
the parameters of the dipole sources that correspond to a
certain potential distribution on the scalp. We will discuss
these methods in the next section.
A. Calculation of simulation EEG: Forward Problem
A current dipole positioned within the head is
characterized by six parameters: location coordinates (x, y, z),
orientation (θ, φ) and time-dependent dipole strength P(t).
The forward problem in dipole localization consists of the
calculation of the potentials at the scalp electrodes, caused
by a known dipole source within the brain. In this problem,
the quasistatic Maxwell equations can be used, which means
that time delays between source and measurements are
negligible [8]. Therefore, the potentials at the electrodes
reflect the instantaneous parameters of the dipole.
In this work the source model is multiple current dipoles,
and has 18 electrode positions that were placed according to
the standard 10–20 system as shown in Fig.1.
Fig. 1. Electrodes placement of 18 channels EEG
Besides we consider a three-concentric-shell for the head,
consisting of a central sphere for modeling the brain and two
concentric spherical shells for the skull and scalp, having
conductivities of 0.33Ω/m, 0.0042Ω/m and 0.33Ω/m
respectively. The outer boundaries of the brain, skull and
scalp have radii of 11.14cm, 10.29cm, and 9.687cm
respectively [9]. The advantage of using a spherical model is
the availability of an analytical expression for the lead field
matrix [9].
B. Estimation of the number of sources
EEG measurement at m electrodes location may be
expressed as
E = GM + N (1)
where E represents the scalp potentials at m electrode
locations, and E=(e1(tj), e2(tj),…, em(tj))T, j=1,2,…,k. The
matrix M is the moments of n current dipoles, and M=(m1(tj),
m2(tj),…,mn(tj))T. Each current dipole has three moments,
mi(tj)=(m1i(tj),m2i(tj),m3i(tj))T, i=1,2,…,n. The matrix N
represents the noise at m electrode locations. We refer to the
m×3n matrix G as the dipole “gain matrix” which maps a
dipole into a set of measurements.
We assume that each dipole moment component is seen as
an independent time course whose movement is different.
By building up the model in this way, the rank of the signal E
subspace is always equal to the number of independent time
courses which include noises [10].
C. Principal Components Analysis (PCA)
ICA need use PCA to pretreat the experimental data and
make it prewhitening correctly. The PCA finds a linear
transformation [11]
V = QE (2)
where E is the m (channels number)×k (samples number)
EEG data matrix and V is the m×k transformation matrix.
The row vectors of Q correspond to the normalized
orthogonal eigenvectors of the data covariance matrix.
According to the number of brain dipoles, dimensionality
reduction is performed in order that ICA can just obtain
independent sources correctly. We choose n (dipoles number)
dominant vectors (eigenvectors) in Q that have the largest
1880
eigenvalues to construct the n×k transformation matrix V
which minimizes the reconstruction error.
D. Blind Source Separation Based on ICA
In the inverse problem, we want to estimate the unknown
sources of mixing EEG data. It is the BSS problem. Firstly,
the linear generative model is expressed as
E = AS (3)
where S=[s1,s2,…,sn]T is a n (dipoles number)×k (samples
number) source matrix of real-valued random variables, each
row of S represents one dipole source, A is a nonsingular
m×n “mixing matrix”, and E is the EEG data matrix.
The goal of BSS is to reconstruct the original source
signals using the observations. For this purpose, after PCA
processing, we adjust an n×n matrix W, called the
“separating matrix”, such that the output
Ŝ = WV (4)
is a scaled and/or permuted version of the source S. Our task
is to discover the source samples Ŝ and separating matrix W.
To find the optimal separating matrix W, Bell, Sejnowski
and Lee have proposed an algorithm based on the extended
infomax principle [3, 6]. The algorithm is given by
ΔW = [I − K tanh(Sˆ)SˆT − SS
ˆ ˆT ]W
⎧ ki = 1: Sup − Gauss ⎫
⎨ ⎬ (5)
⎩ki = −1: Sub − Gauss ⎭
where ki are elements of the N-dimensional diagonal matrix
K. The ki can be derived from the generic stability analysis of
separating solutions.
E. Reconstruction of EEG due to each source
Considering the time series ei(tk) in the ith electrode, we
denote the contribution of independent component sˆ j to the
reconstruction of ei as
eij (tk ) = Pij−1sˆj (tk ), 1≤ i ≤ m,1< j < n, tk =1,2,"k , (6)
⎡W −1 ⎤
P -1 = Q -1 ⎢ ⎥ (7)
⎣ 0 ⎦
where Pij-1 denotes the (i, j)th element of the matrix P-1, and
eij(tk) is the scalp potentials in the ith electrode due to just the
jth source. Thus, ej(tk) is the set of scalp potentials in all
electrodes due to just the jth source which consist of eij(tk). It
is the projection of independent activation maps back onto
the electrode arrays. Then, the reconstruction of ei by using
the m independent components is given by
n
ei (tk ) = ∑eij (tk ), 1≤ i ≤ m,1< j < n, tk =1,2,"k (8)
j =1
That is to say, ICA allows us to reconstruct surface potentials
that would exist due to each dipole as if it was the only
source. For example, if the brain source estimation gives
four sources, that means we will be looking for only four
dipoles (including noise). Projecting each activation map on
the scalp electrodes gives us four different maps, each with a
time sequence of values. For each activation map, we choose
an instant in time from the time sequence (fixed point in
time), and then use them to localize one dipole by the our metric for comparing the potentials at the electrodes
downhill simplex method which finds the minimum of the [12].
multidimensional cost function. In an N dimensional space, The ICA error can thus be cumulatively estimated over all
the simplex is a geometrical figure that consists of (N+1) electrodes over the entire time sequence, by evaluating the
interconnected vertices [7]. Based on the source model, the mean square root error of ρi (the ith channel) from 1 to all
dipole has 6 parameters, so that the simplex has 7 vertices. channels:
The downhill simplex method minimizes a function by m
taking a series of steps, each time moving the point in the
simplex away from where the function is the largest.
error = ∑ (ρ
i =1
i − 1) 2 / m (9)

Evaluated with the above formula, the four activation

III. RESULTS projections restore the original (unmixed) potential
We prepare the simulated data generated by forward distribution with RMS errors of 0.71%, 1.19%, 2.42% and
problem described in the previous sections, and assume that 2.87% respectively. It takes about 0.1 second for the
the dipoles are spatial stationary and temporal independent. simulation to run on a Pentium 4 2.8G CPU and 2G memory
The data are corrupted by additive spatially white noise with computer. When the data are corrupted by different additive
5dB SNR. Fig.2 shows the time dependent EEG of 128 ms white noise, the errors are shown in TABLE I.
for all 18 electrodes. By estimating the number of sources we In order to know the relationship of unmixing error,
know that there are four dipole sources within the brain that distance between dipoles and dipole moment movements, in
are correct. Then ICA procedure is performed on the given the following simulation we assume there are two dipoles in
time dependent data, and the activation maps are shown in the brain, and calculate simulated data in the different
Fig.3. We notice that there are four different activation maps parameter configuration by forward problem. By ICA
presented in the result, the other is noise. Projecting the first processing, the errors in all kinds of configurations of
activation map on all 18 channels, we get the signal eˆi , dipoles parameter are shown in TABLE II.
In the configuration I, the dipoles are spatial stationary
which are the potentials due to the single temporal lobe and temporal independent. And the unmixing accuracy is
dipole. Now we can check the accuracy of the ICA satisfying. In the configuration II, although 2 dipoles are in
decomposition by comparing ICA recovered time series eˆi the same location, dipole moment movements are temporal
to the result ei of the forward problem simulation that run independent component 1
with three of the four dipoles “turned off”. Because ICA 5

does not preserve scale, we use correlation coefficients ρ as

S c a le

0.2 0.5 0.2 -5

p o t e n t ia l(m v )

p o t e n t ia l(m v )

p o t e n t ia l(m v )

0 20 40 60 80 100 120 140

0 0 sample number

-0.2 0 -0.2 independent component 2

0 50 100 150 0 50 100 150 0 50 100 150
time(ms) time(ms) time(ms) 5
S c a le

1 0.2 0.4 0
p o t e n t ia l(m v )

p o t e n t ia l(m v )

p o t e n t ia l(m v )

0.5 0 0.2 -5
0 20 40 60 80 100 120 140
0 -0.2 0 sample number
0 50 100 150 0 50 100 150 0 50 100 150
time(ms) time(ms) time(ms)
0.2 0.2 0.2 independent component 3
p o t e n t ia l(m v )

p o t e n t ia l(m v )

p o t e n t ia l(m v )

5
0 0 0
S c a le

0
-0.2 -0.2 -0.2
0 50 100 150 0 50 100 150 0 50 100 150 -5
time(ms) time(ms) time(ms) 0 20 40 60 80 100 120 140
sample number
0.2 0.2 1
p o t e n t ia l ( m v )

p o t e n t ia l ( m v )

p o t e n t ia l ( m v )

0 0 0.5 independent component 4

-0.2 -0.2 0
5
0 50 100 150 0 50 100 150 0 50 100 150
S c a le

time(ms) time(ms) time(ms) 0

0.5 0.2 0.2 -5

p o t e n t ia l ( m v )

p o t e n t ia l ( m v )

p o t e n t ia l ( m v )

0 20 40 60 80 100 120 140

0 0 0 sample number
-0.5 -0.2 -0.2
0 50 100 150 0 50 100 150 0 50 100 150 independent component 5
time(ms) time(ms) time(ms) 5
0.2 0.2 0.2
p o t e n t i a l( m v )

p o t e n t i a l( m v )

p o t e n t i a l( m v )

S c a le

0
0 0 0
-5
-0.2 -0.2 -0.2
0 20 40 60 80 100 120 140
0 50 100 150 0 50 100 150 0 50 100 150 sample number
time(ms) time(ms) time(ms)
Fig.3 Independent components obtained by unmixing the input signals.
Fig.2 Simulated scalp potentials evoked by four dipole sources in 18 The anterior four ones are due to 4 dipoles, and the last one is noise
channels. The SNR is 5 dB. which autocorrelation coefficient is about 0.

1881
 TABLE I for Spatio Temporal Neuromagnetic Source Localization. Report:
THE UNMIXING ERROR WHEN THE DATA ARE CORRUPTED BY DIFFERENT LA-UR-96-2042.
ADDITIVE SPATIALLY WHITE NOISE [3] A. J. Bell and T. J. Sejnowski. An information-maximization approach
SNR 0dB 5dB 10dB No noise to blind separation and blind deconvolution. Neural Comput., vol. 7,
pp. 1129–1159, 1995.
0.63% 0.71% 0.92% 0.64% [4] J. F. Cardoso and B. H. Laheld. Equivariant adaptive source separation.
1.28% 1.19% 1.21% 0.92% IEEE Trans. Signal Processing, vol. 44, pp. 3017–3030, 1996.
Errors
1.98% 2.42% 1.80% 1.66% [5] A. Hyvärinen and E. Oja. A fast fixed-point algorithm for independent
3.68% 2.87% 2.12% 2.88% component analysis. Neural Comput., vol 9, pp. 1483-1492, 1997.
[6] T. W. Lee, M. Girolami, and T. J. Sejnowski. Independent component
independent, and the unmixing result can be obtained. In the analysis using an extended infomax algorithm for mixed sub-Gaussian
and super-Gaussian sources. Neural Comput., vol. 11, pp. 417–441,
configuration III, it is equivalent to only one dipole in the 1999.
brain, so we can only obtain a good result. In the [7] Nedler, J.A. and Mead, R. A simplex method for function
configuration IV, V and VI, dipoles moment movements are minimization. Compt. J 7, 308-313, 1965.
the same and locations are fixed. Whatever the distance is, it [8] Malmivuo J and Plonsey R. Principles and Applications of Bioelectric
and Biomagnetic Fields. New York, Oxford University Press, 1995.
seems there is only an independent source in the brain, and [9] Salu Y, Cohen L G, Rose D, Sato S, Kufta C and Hallett M. An
we only obtain a correct result. From above illustration, we improved method for localizing electric brain dipoles. IEEE Trans.
can see that the dipoles are spatial stationary and temporal Biomed. Eng. 37 699–705,1990
independent in our simulation, so the brain sources [10] John C. Mosher, and Richard M. Leahy. Recursive MUSIC: A
Framework for EEG and MEG Source Localization. IEEE
independence is mostly relative with the dipoles moment TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 45,
movement, not the distance between dipoles. The accuracy NO. 11, NOVEMBER 1998
of ICA is mostly relative with the independence of brain [11] T.P Jung, Colin Humphries, Te-Won Lee. Removing
sources, so the errors of localization lie on the movement of Electroencephalographic Artifacts: Comparison between ICA and
PCA. Computational Neurobiology Lab, The Salk Institute, San Diego,
dipoles moment mostly. CA 92186-5800
[12] Leonid Zhukov, David Weinstein and Chris Johnson. Statistical
IV. CONCLUSIONS Analysis for FEM EEG Source Localization in Realistic Head Models.
Department of Computer Science, University of Utah., 2000.
In the paper, it is explored that the possibility of applying
ICA for EEG multiple dipoles localization on a concentric
head model with simulated EEG data, when the data are
corrupted by additive noise. Some algorithms have been
obtained that reduce the complexity of multiple dipole
sources localization by exploiting the temporal dependence
of the data. It has been also shown that our algorithm is
capable of correctly predicting the number of independent
sources in the model and reconstructing scalp potentials
evoked by each source separately. Furthermore, we also give
the relationship of unmixing accuracy, distance between
dipoles and dipoles moment movements.

REFERENCES
[1] Supek, S. and Aine, C.J. Simulation studies of multiple dipole
neuromagnetic source localization: model order and limits of source
resolution. IEEE Trans. Biomed. Eng. 40, 354-361, 1993
[2] Harrison, Reid R. and Aine, Cheryl J. and Chen, Hai wen and Edward
R. Flynn, Edward R. and Huang Mingxiong. Investigation of Methods
TABLE II
THE ERRORS IN ALL KINDS OF CONFIGURATIONS OF DIPOLES PARAMETER
No. I II III IV V VI

Movements
Temporal Temporal
of 2 dipoles Same Same Same Same
independent independent
moment

Distance
between 3cm 0cm 0cm 0.1cm 1cm 3cm
2 dipoles
0.14% 0.08% 0.004% 0.19% 0.32% 0.089%
Error 0.12% 1.18% 98% 93.9% 92% 95.7%

1882