www.elsevier.
com/locate/ynimg
Evaluation of different cortical source localization methods using
simulated and experimental EEG data
Jun Yaoa and Julius P.A. Dewalda,b,c,d,*
a
Department of Physical Therapy and Human Movement Sciences, Northwestern University, Chicago, IL 60611, USA
b
Department of Physical Med and Rehab., Northwestern University, Chicago, IL 60611, USA
c
Department of Biomedical Engineering, Northwestern University, Evanston, IL 60201, USA
d
Sensory Motor Performance Program, Rehabilitation Institute of Chicago, Chicago, IL 60611, USA
Received 8 February 2004; revised 23 July 2004; accepted 29 November 2004
Different cortical source localization methods have been developed to
directly link the scalp potentials with the cortical activities. Up to now,
these methods are the only possible solution to noninvasively inves-
tigate cortical activities with both high spatial and time resolutions.
However, the application of these methods is hindered by the fact that
they have not been rigorously evaluated nor compared. In this paper,
the performances of several source localization methods (moving
dipoles, minimum Lp norm, and low resolution tomography (LRT)
with Lp norm, p equal to 1, 1.5, and 2) were evaluated by using
simulated scalp EEG data, scalp somatosensory evoked potentials
(SEPs), and upper limb motor-related potentials (MRPs) obtained on
human subjects (all with 163 scalp electrodes). By using simulated EEG
data, we first evaluated the source localization ability of the above
methods quantitatively. Subsequently, the performance of the various
methods was evaluated qualitatively by using experimental SEPs and
MRPs. Our results show that the overall LRT Lp norm method with p
equal to 1 has a better source localization ability than any of the other
investigated methods and provides physiologically meaningful recon-
struction results. Our evaluation results provide useful information for
choosing cortical source localization approaches for future EEG/MEG
studies.
D 2004 Elsevier Inc. All rights reserved.
Keywords: Brain imaging; Source localization; Dipole fit; Current density
reconstruction; Somatosensory evoked potentials; Event-related potentials
Introduction
Brain activity has both spatial and temporal characteristics.
Conventional scalp EEG recordings are noninvasive global
* Corresponding author. Department of Physical Therapy and Human
Movement Sciences, Biomedical Engineering, Northwestern University,
645 N. Michigan Ave, Suite 1100, Room 1159, Chicago, IL 60611, USA.
Fax: +1 312 908 0741.
E-mail address: j-dewald@northwestern.edu (J.P.A. Dewald).
Available online on ScienceDirect (www.sciencedirect.com).
1053-8119/$ - see front matter D 2004 Elsevier Inc. All rights reserved.
doi:10.1016/j.neuroimage.2004.11.036
NeuroImage 25 (2005) 369 – 382
measurements of cortical activity with a millisecond temporal
resolution. However, the spatial resolution of EEG recordings is
limited, which is mainly caused by the blurring effect of the head
volume conductor. In addition, the effects of electrode density, i.e.,
spatial sampling, signal-to-noise ratio (SNR) of the recordings, and
the reference electrode placement, all contribute to the level of
spatial resolution obtained by inverse calculations with EEG
(Nunez, 1981). To improve the spatial resolution of scalp EEG,
tremendous efforts have been made to use cortical source local-
ization techniques to de-blur the scalp EEG recordings by de-
convolving the measured scalp EEG distribution into an electrical
activity (current or potential) over the cortical surface. These
techniques are expected to improve the spatial resolution from
centimeter scale on the scalp to millimeter scale on the cortex, and
thus make it possible to show cortical spatiotemporal activities
with a significant improvement in spatial resolution in a non-
invasive way.
In order to de-convolve the scalp EEG, a head volume con-
ductor model and a source model are required. As for the head
volume conductor model, a homogeneous sphere head volume
conductor model, a boundary element method (BEM) model, or a
finite element method (FEM) model can be used. Since the BEM
model is a compromise between oversimplifying sphere head
volume model and the large computational costs of a FEM model,
the subject specific BEM model is currently most widely used
(Fuchs et al., 1998). As for the source model, two commonly used
models are equivalent current dipole models (Scherg, 1990; Scherg
and Buchner, 1993; Williamson and Kaufman, 1981) and current
distributed source models (Hamalainen and Ilmoniemi, 1994;
Ilmoniemi, 1993; Nenonen et al., 1994; Pascual-Marqui and
Biscay-Lirio, 1993; Pascual-Marqui et al., 1994; Wagner et al.,
2000; Wang et al., 1992, 1993).
When using any source models, regularized solutions are
commonly used to solve the ill-posed inverse problem. With the
combination of different head models, source models, and
associated regulation methods, there are currently dozens of
different cortical source localization methods available. The
370 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382
advantages and disadvantages of these methods were analyzed in
principle and described qualitatively by Wagner et al. (2000).
Comparisons between the ability of methods to detect source
location(s) have been reported in the literature using experimental
EEG data (Cincotti et al., 2004; Haan et al., 2000; Komssi et al.,
2004; Pascual-Marqui et al., 1994; Phillips et al., 2002; Stenbacka
et al., 2002; Waberski et al., 2000) and simulated EEG data
(Babiloni et al., 2000b, 2004; Baillet and Garnero, 1997; He et al.,
2002a) with different number of EEG electrodes (ranging from 19
to 128), head models, and regulation parameters involved. All
these previous studies incorporated comparisons of two or at most
three methods with fixed regulation settings. The purpose of this
paper is to expand comparisons to a large number of source
reconstruction methods, including dipole methods (single and dual
moving dipole), minimum Lp norm methods ( p equals to 1, 1.5,
and 2), and low resolution electromagnetic brain tomography
(LRT) methods with various Lp norms ( p equals to 1, 1.5, and 2).
In an effort to estimate the accuracy of these source localization
methods, we first created simulated EEG data with the obvious
advantage that the exact source locations/regions were generated
by the investigator. The shortcomings of using simulated EEG data
are (1) the data are not realistic and (2) since the forward and
backward calculations are based on the same head model, the
results obtained based on simulated EEG cannot reflect errors
introduced by the head model but only show the accuracy of source
model and associated regulation methods. In addition to using the
simulated EEG data, we also qualitatively tested the performance
of the various source localization methods using experimental EEG
data, including somatosensory evoked potentials (SEPs) from
electrical stimulation of the wrist, elbow, and shoulder, and event-
related potentials (ERPs) resulting from the generation of isometric
shoulder abduction torques. The advantage of using SEPs and
ERPs is that the location of sensory-motor cortical activity is well
described in the literature ((Babiloni et al., 2000a; Druschky et al.,
2003; Emerson and Pedley, 1984; Hari and Forss, 1999;
Hashimoto, 2000; He et al., 2002b; Recuero, 1999; Sonoo, 2000;
Towle et al., 2003; Valeriani et al., 2000) for SEPs and (Penfield
and Rasmussen, 1950; Rao et al., 1995) for ERPs), which
permitted us to determine whether solutions obtained with the
various source localization methods were reasonable or not.
In the following parts of this paper, a BEM realistic head model
was used for both EEG data simulation and inverse calculation. All
the computations were performed using the Curry V4.5 (Neuroscan,
TX) software package, which is a widely used commercial source
localization software package. Other source localization methods,
such as cortical potential imaging method (He et al., 2002b) and the
three-dimensional resultant vector method (Ricamato et al., 2003),
are not included in this study. Parts of this work have been
published as a conference preceding (Yao and Dewald, 2003).
Materials and methods
Simulation EEG data
By placing equal strength (1 AAmm) dipoles on the cortex, we
simulated three EEG data sets, i.e., (I) single dipole area with a fixed
center and an increasing radius ranging from 5 mm to 10 mm, (II)
two equal-size areas of dipoles with fixed centers and increasing
radiuses ranging from 5 mm to 10 mm, and (III) two dipoles with
increasing distance between them ranging from 4 mm to 20 mm.
(For the convenience of representation, in the later parts of this
paper, we call these three simulated EEG data sets as dsimulated
data sets I/II/IIIT.) A single dipole area (simulated data set I) was
used to simulate EEG data generated from a concentrated source.
Two equal-size areas of dipoles with increasing area size and two
dipoles with increasing distance (simulated data sets II and III) were
used to simulate EEG data generated from multiple sources.
Each EEG data set is comprised of a total of 163 EEG signals,
which are the forward calculation results obtained at the electrodes
on the scalp with a reference at Oz. The forward calculation was
performed using a realistic BEM model, which consists of three
shells: skin, skull, and liquor. The BEM model is developed based
on the MRI data (with a resolution of 1.0
1.0
1.0 mm) of a
normal subject. The parameters of BEM model are listed in Table 1.
Since real EEG recordings are usually noisy, we further added
noise to the simulated EEG data. Noise was obtained from a real
EEG measurement in which the subject was not performing any
task and with his eyes open and looking at a fixed point. One
hundred epochs were averaged to achieve the noise levels
comparable to the typical event-related potentials (ERPs). This
corresponds to a SNR of approximately 9–10. Subsequently, the
data with noise were low-pass filtered by a 9th order zero-lag filter
with the cut-off frequency of 50 Hz.
Experimental EEG data
Each of the inverse methods was also applied to two types of
experimental EEG data, i.e., somatosensory evoked potentials
(SEPs) and event-related potentials (ERPs).
Recording and processing of somatosensory evoked potentials
SEPs resulting from electrical stimulation of right wrist, elbow,
or shoulder of one subject were recorded by a BioSemi Active II
system (BioSemi, Amsterdam, Netherlands) with 163 active
electrodes and a reference on the right ear lobe. Constant current
square-wave pulses of 0.3 ms duration were delivered to the
median nerve at the wrist, or lateral epicondyle elbow, or the
acromion of the shoulder through disposable neurology electrodes
(Ambu USA, Linthicum, MD) with 1 cm between anode and
cathode using a Compex2 system (Compex, Annecy-le-Vieux,
France). An interstimulus interval of 500 ms was used for a total of
1000 stimuli for each stimulus location. The intensity of the pulses
was set to the highest amplitude that the subject could tolerate
without discomfort. The SEP data were sampled at 2 kHz. The
positions of each EEG electrode and three fiducial landmarks
(Nasion, left preauricular point and right preauricular point) were
marked by a Polhemus (Polhemus, Colchester, VT) three-dimen-
sional magnetic digitizer. The three fiducial landmarks were used
to co-register the 163 EEG electrodes with the subject’s MRI data.
The co-registration was repeated by adjusting the fiducial land-
marks in MRI data until all the 163 electrodes were co-registered
well with the subject’s scalp.
Table 1
Parameters of the BEM model
Shell Surface Side length Conductivity
triangle no. (mm) (S/m)
Skin 2692 10.0 0.25
Skull 2418 9.0 0.017
Liquor 3306 7.0 1.79
 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382
The SEPs were read into Brain Vision Analyzer V1.04 software
(Brain Products GmbH, Mqnchen, Germany), filtered by a 1–
500 Hz (48 dB/oct) band-pass filter with a notch at 60 Hz. Then the
intervals that contained eye blink signals were eliminated from
further analysis. The data over 40 ms pre- and 60 ms post-stimulus
period were baseline corrected and averaged.
Recording and processing of event-related potentials
Electroencephalographic potentials related to isometric shoul-
der abduction torque generated by one subject were recorded by
the BioSemi Active II system (BioSemi, Amsterdam, Netherlands).
During the experiment, the subject was casted at the wrist and
secured to a six-degree-of-freedom (DOF) load cell with the
shoulder at a 758 abduction angle and a 408 flexion angle. The tip
of the hand was located at a distance from the body corresponding
to an elbow angle of 908 with 08 representing full extension of the
elbow. In order to minimize the activation of trunk muscles,
subjects were seated in a Biodex chair with the trunk secured and
shoulders strapped to the back of the chair. A monitor was placed
in front of the subject to provide visual feedback of torque
generation in the SABD direction.
The forces and torques generated at the wrist were measured
using a six-degree-of-freedom load cell (JR3 Inc., Woodland, CA)
and converted online to torques at the elbow (flexion/extension)
and shoulder (flexion/extension, abduction/adduction, and exter-
nal/internal rotation) based on a free body analysis of the upper
limb (Beer et al., 1995). Online visual feedback of the SABD
level was provided during the training section. Maximum
voluntary torque (MVT) in shoulder abduction direction was first
recorded for the subject. The subject was then trained to generate
a self-initiated SABD torque at 25% of his MVT and maintain it
for a period of 0.3 s. After the subject felt comfortable with the
task, EEG data were collected. During the data collection, the
subject was asked to repeat the same task 100 times in blocks of
20 trials with 10 s rest periods between individual trials and 10
min rest periods between blocks to avoid fatigue. In order to
reduce brain activity at the visual cortex, no visual feedback was
Fig. 1. Settings for the various source localization methods.
371
provided during the data collection section; instead, the subject
was required to look at a fixed point without eye movement
during the whole period of a trail (totally about 8 s including 5–6
s before moving and 2 s after moving). Torque generation
performance was reported to the subject at the end of each trial.
The EEG data were sampled at 2 kHz. Similarly, co-registration of
EEG electrode positions with the subject’s scalp was performed
off-line.
The event-related potentials (MRPs) were read into Brain
Vision Analyzer V1.04 software. The intervals that contained eye
blink signals were eliminated from further analysis. The data over
2000 ms pre- and 500 ms post-torque onset were baseline corrected
and averaged.
Source reconstruction
A realistic BEM head model and two different source models,
dipole and distributed current source models, were used to
reconstruct the cortical source of both simulated and experimental
EEG data. Each inverse method was applied to simulated EEG data
with 10 different time points. Among these 10 points, the noise is
random. Regularization methods and parameters used in this paper
are listed in Fig. 1. The meaning of each of the parameters is
described in the Curry reference guide (NeuroScan, 1999). For
convenience of representation, in the rest of this paper, we refer to
all the source localization methods with different source model,
different regularization methods, or parameters as methods 1–8.
The exact makeup of each of the methods is provided by the upper
indices in the most right column of Fig. 1 with abbreviations listed
beside the indices.
Inverse methods using the dipole source model operated on the
assumption that only a discrete number of generators (usually
dipoles) were active at a given time or over a time period. In this
paper, we investigated single and dual moving dipole methods
(MDP1 and MDP2). These methods model cortical currents with
equivalent current dipoles that are described by their three-
dimensional locations, fixed orientation, and variable amplitude
372 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382
(Williamson and Kaufman, 1981). The position of the dipoles was
found using a least-squares search that minimizes the difference
between the estimated and measured data. Dipole models require a
priori knowledge of the number of sources, which is usually
unknown.
Conversely, when dealing with current distributed-source
reconstruction (CDR) methods, no a priori knowledge of the
number of sources is required. Since the neurophysiological
potentials follow Ohm’s law, using CDR methods to reconstruct
the source is a linear underdetermined inverse problem. Solution to
this problem uses a regulation operator. More specifically, the
regulation used in minimum Lp norm methods minimizes the
variance D 2 which reads D 2 = tLx  mtp + ktWxtp , where x is
the vector of source currents; L is the lead field matrix; m is the
EEG measurements; W is a diagonal location weighting matrix; k
is the regulation parameter; and p (1 V p V 2) is the measure in
Banach space. When W is equal to identity matrix (as used in this
paper), it is referred to as a regular minimum Lp norm method;
otherwise, it is a weighted minimum Lp norm method. By
changing the weighting matrix W to BW, where B is the Laplacian
coupling matrix, it becomes a LORETA (LRT) method. The effect
of matrix B is to constrain the neighboring sources to similar
strengths and thus smooth the source-current distributions. In this
paper, minimum Lp norm and LRT Lp norm methods with p equal
to 1, 1.5, and 2 were investigated. Furthermore, solutions of CDR
methods investigated in this paper were constrained on a cortex
overlay with 5 mm between each point and the directions of
sources were rotated.
Evaluation methods
Evaluation of the results on simulated EEG data
The accuracy of the inverse solution using the simulated EEG
data was evaluated by three indices: (1) the error distance (ED),
which is the distance between the locations of real and estimated
sources, (2) the percentage of undetected source number (PUS),
and (3) the percentage of falsely-detected source number (PFS).
The error distance (ED) between the real and the estimated
source locations were defined as
1 X
NI   1 XNJ
 

min kd˜i  sj k þ
min ksj  dd̃i k : ð1Þ
NI i
j NJ ja J
i Hari and Forss, 1999). Direct recording of cortical potentials in
By using this definition, one can measure the averaged error
between the inverse calculation and the actual source location as a
function of distance. In this equation, s i is the real dipole location
of the simulated EEG data; d̃ i is the source location detected by
each inverse calculation method; i and j are the indices of locations
of estimated and the actual sources; and N I and N J are the total
numbers of estimated and the undetected sources, respectively. In
the case that CDR methods were used, a source location is defined
as the location where the current strength is larger than the
threshold. This threshold is set as the value corresponding to 98%
confident level of a Weibull distribution function (Weibull, 1939)
that fits well to the estimated cortical currents. The first term of Eq.
(1) calculates the mean of the distance from each estimated source
to its closest real source; and the corresponding real source is then
marked as a detected source. All the undetected real sources made
up the elements of data set J. And the second term of Eq. (1)
calculates the mean of the distance from each of the undetected
sources to the closest estimated sources.
The percentages of undetected source number (PUS) and of
falsely-detected source number (PFS) were defined as
Nun Nfalse
PUS ¼ ; and PFS ¼ ; ð2Þ
Nreal Nestimated
where N un, N false, N real, and N estimated are the numbers of
undetected, falsely-detected, real, and estimated sources, respec-
tively. When calculating PUS and PFS, an undetected source is
defined as a real source whose location to its closest estimated
source is larger than 0.6 times the unit distance of the source
layer. (Since we constrained all the solutions on a grid of cortex
with 5 mm between each point, the unit distance of the source
layer is equal to 5 mm.) Similarly, a falsely-detected source is
defined as an estimated source whose location to its closest real
source is larger than 0.6 times the unit distance of the source
layer.
Evaluation of the results on real EEG data
Since the exact locations of SEPs and ERPs are unknown,
it is hard to quantitatively evaluate the results obtained by
different inverse methods. Qualitative evaluation based on current
physiologic knowledge of expected active-region was therefore
performed.
EEG source localization associated with SEP data is relatively
clean and has been well documented in the literature (Babiloni
et al., 2000a; Emerson and Pedley, 1984; Hashimoto, 2000;
Recuero, 1999; Sonoo, 2000; Valeriani et al., 2000). The most
widely researched and clinical applied SEPs are elicited by
stimulation of the median nerve at the wrist (Emerson and Pedley,
1984; Sonoo, 2000; Valeriani et al., 2000). More recently, SEP
data resulting from the stimulation of fingers or other sites have
also been reported (Dowman and Schell, 1999; Druschky et al.,
2003; Restuccia et al., 2002; Schaefer et al., 2002; Waberski et al.,
1999). Different imaging modalities, such as fMRI, PET, and
source localization methods based on EEG/MEG, showed that
Brodmann areas 2 and 3a, which receive deep joint and muscle
receptor inputs, are believed to be the generator sites for the early
components of proprioception-related evoked responses, in
contrast to area 3b activation following cutaneous mechanical
and electrical stimulation, which with clearly somatotopical
representations of different body parts (Druschky et al., 2003;
the human further confirmed the above findings (Towle et al.,
2003).
Event-related cortical activity has mostly been studied using
cortical stimulation or extracellular recordings. Early stimulation
experiments in monkeys and chimpanzees (Ferrier, 1875; Hines,
1940; Sherrington, 1947; Woolsey, 1958) demonstrated that
stimulation of the primary motor cortex (M1) resulted in movement
in the contralateral side of the body. Penfield and Rasmussen (1950)
performed cortical stimulation experiments in humans and reported
results similar to those found in animal work. Furthermore,
intracortical microstimulation results in the monkey (Asanuma
and Rosen, 1972; Schieber, 2001) and from other studies, including
single neuron recordings in monkeys (Schieber and Hibbard, 1993),
fMRI recordings in humans (Sanes et al., 1995), and MEGs in
humans (Cheyne et al., 1991; Salenius et al., 1997), all suggested
that M1 is the source generator of movement-related potentials.
The previously established knowledge of the source for SEPs
and ERPs was used to evaluate the performance of different source
localization methods qualitatively.
 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382
Statistical analysis method
The three evaluation indices (i.e., ED, PUS, and PFS) were
subjected to separate multivariate analysis of variance (MAN-
OVA). The two dependent variables of the MANOVA were the
mean and standard deviation of each tested index, and the three
factors were the inverse method, the type of simulated EEG data
set, and the radius of the dipole area (for data set I/II) or the
distance between the two dipoles (for data set III). Up to 2-way
interactions between the factors were considered when performing
the analysis. The post hoc analysis was implemented using the
Bonferroni test at the P = 0.05 level of significance.
Results
Source localization results obtained on simulated EEG data
Example results obtained using each of the methods on
simulated data sets I/II/III (with a radius that equals 8 mm for
data sets I and II, and with a distance between the two dipoles that
equals 18 mm for data set III) are shown in Figs. 2A–C,
respectively. Results are shown in a brain coordinate system with
x- and y-axis representing anterior (negative x)/posterior (positive
x) and left (positive y)/right (negative y) in mm. In this figure, the
first and second rows showed the results of minimum Lp norm ( P =
1, 1.5, 2) and LRT Lp norm ( P = 1, 1.5, 2), respectively. Color bars
on the right side of each subplot show the averaged current strength
over the 10 time points on the cortex in AAmm. On the last row, the
purple dots on the first two subplots are the mean of the estimated
dipole location over the 10 time points obtained by using MDP1
and MDP2 methods. The horizontal, vertical, diagonal lines
represent the standard deviation of estimated dipole position on
x, y, and z directions. Finally, the purple dots in the last subplot
represent the real source dipole positions. Quantitative evaluation
results and statistical analysis results were performed and will be
discussed next.
Location accuracy
The location error distances (EDs) of each inverse method
using simulated data sets I/II/III are shown in Fig. 3. In this figure,
the gray bars on each of the data points represent the standard error
of corresponding results. Results clearly show that the LRT1
method provides the best estimation of source location for all three
simulated EEG data sets. When applying CDR methods on
simulated data sets I and II, the EDs of LRT1 have general
decreasing trends when the sizes of dipole-areas increase. How-
ever, such a decreasing trend is not so obvious for the other
methods on both data sets I and II (see Figs. 3a and b). Results of
EDs obtained by applying LRT with Lp ( P = 1, 1.5, 2) norm
methods on simulated data set III fluctuate when the distance
between the two source dipoles is about 12–16 mm (see Fig. 3c).
The reason for this fluctuation is believed to be related to
constraints imposed on source location by the 5 mm grid on the
cortex (see Materials and methods).
Although the moving dipole methods (MDP1 and MPD2) show
relatively low EDs, there is an obvious posterior and medial shift
shown on the results of all the three simulated EEG data sets, while
such a shift is not found in the results obtained by using CRD
methods (see Fig. 2). Even worse, locations estimated by the
MDP2 method over 10 time points show a large variance (see
373
Fig. 3). It is worth to note that when applying MDP2 to simulated
data set III, there is a decreasing trend in the location error when
the distance between the two real dipoles increases. This trend is
not obvious for any of the other methods. This result suggested that
the a priori knowledge of the number of dipoles could improve the
inverse results when using dipole source models. Unfortunately,
even when the correct number of dipoles is implemented, the
MDP2 method still produces posterior and medial shifts, which
result in substantial location errors.
The results of statistical analysis using a three-way MANOVA
and Bonferroni post hoc tests on the mean and standard deviation
of location errors are shown in Table 2a. In this table, values in the
upper and lower triangle are P values on the mean and standard
deviation of different indices, respectively. Results in Table 2a
show that location estimated by LRT1 method is significantly
more accurate than any other methods tested in this paper (see
statistical results on ED in the top triangle of Table 2a). Further-
more, the variability (estimated by standard deviation) of the LRT1
method on estimating the location of source is either significantly
less than or similar to other methods (see statistical results on the
standard deviations in the bottom triangle of Table 2a). The
location error obtained by using LRT15 is significantly smaller
than results obtained by using any other methods except LRT1 and
MDP1.
The percentage of undetected source number
The percentages of undetected source number (PUS) of each
source localization method using simulated data sets I/II/III are
shown in Fig. 4. Since dipole methods are not able to detect the
area of the source, the PUSs for dipole methods are close to 100%.
When using simulated data sets I and II, PUSs of all CDR methods
have a tendency to increase when the size of dipole area increases.
This tendency is slower for the LRT1 method than that for any of
the other tested CDR methods. It is worth noting that when using
simulated data set II to evaluate the CDR methods, there is an
obvious drop of PUS when the radiuses of the two source dipole-
areas increase from 5 mm to 6 mm. When referring to results
obtained with simulated data set II with the radiuses of the two
dipole areas equal to 5 mm in Fig. 2d, one finds that all the CDR
methods could not detect the laterally located dipole area; instead,
they detected a single enlarged and merged dipole area located
closer to the medial dipole area (see Fig. 2d). When increasing the
radiuses from 5 mm to 6 mm, all CDR methods show an un-
separated bimodal area, which is related to the drop of PUS
observed in Fig. 4b. This result shows that a 5 mm source grid does
not allow for the detection of two separated sources each with a 5
mm radius and with 29 mm between their centers.
The simulated data set III consists of two dipoles. For this type
of EEG data, all CDR methods are able to detect the two dipoles
when the distance between the 2 source dipoles is less than 12 mm.
However, most of CDR methods except LRT1 are unable to detect
one of the two dipoles when the distance between the 2 source
dipoles is larger than 12 mm. In all the circumstances, LRT1 has
the lowest PUS from all CDR methods.
The results of statistical analysis using three-way MANOVA
and Bonferroni post hoc tests on the mean and standard deviation
of the percentages of undetected source number (PUSs) are shown
in Table 2b. Results show that LRT1 has a significantly smaller and
less variable PUS than all the other methods tested in this paper
(except that the mean and standard deviation on PUSs for LRT1
and LRT15 are similar).
374 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382

The percentage of falsely-detected source number are shown in Fig. 5. Results show that when using simulated data
The percentages of falsely-detected source number (PFSs) of sets I and II, PFSs of LRT1 and LRT15 methods have a tendency to
each source localization method using simulated data sets I/II/III decrease when the size of dipole area increases, while this trend is

Fig. 2. Results obtained using each of the investigated inverse methods for simulated EEG data with (A) single dipole area with a radius equal to 8 mm, two areas
of dipoles with a distance between two centers equal to 29 mm and radiuses equal to (D) 5 mm and (B) 8 mm, and (C) two single source dipoles with an inter-
dipole distance of 18 mm. The order of the inverse methods is corresponding to the methods 1–8 in Fig. 1. Results were shown in brain coordinates with x- and y-
axis representing anterior (negative x)/posterior (positive x) and left (positive y)/right (negative y) in mm. Color bars on the right side of subplots in the first two
rows show the averaged current strength over the 10 time points on the cortex in AAmm. In the last row, the purple dots on the first two subplots are the mean of
the estimated dipole locations over the 10 time points obtained by using MDP1 and MDP2 methods. The horizontal, vertical, and diagonal lines represent the
standard deviation of estimated dipole position on x, y, and z directions. Finally, the purple dots in the last subplot represent the source dipole positions.
 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382 375

Fig. 2 (continued).

not apparent for the other CDR methods. PFSs of MDP1 for
simulated data sets I/II/III are always equal to 100%, which shows
that this method is not able to detect the source location within the
source area (caused by the posterior and medial shift). The PFSs of
MDP2 decrease with the increase of the radius of source area using
simulated data set I, which clearly suggests that adding a second
dipole is beneficial in reducing the false detection ratio. However,
MDP2 has a significant larger standard deviation compared to
other methods (see Table 2c).
The results of statistical analysis using three-way MANOVA
and Bonferroni post hoc tests on the mean and standard deviation
of the percentages of PFS are shown in Table 2c. Results showed
that LRT1 has a significantly smaller PFS than any of the other
methods.
376 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382

the sources of SEPs are located on the primary sensory cortex (S1
area). Results obtained by using CDR methods have a similar dhot
spotT located on S1 area, while multiple other active sites besides S1
area also show up. Among all the CDR methods, results of methods
L1, L15, and LRT1 are more focused on S1 area; furthermore,
methods L15 and LRT1 show relatively stronger activity on S1 area.
Results of ERPs related to self-initiated isometric shoulder
abduction during a time window from 110 to 90 ms before the
onset of torque are shown in Fig. 7. The majority of activities prior
to shoulder abduction obtained by all the CDR methods are located
on the contralateral hemisphere. Results by MDP1 and MDP2 are
located either at prefrontal lobe or parietal lobe. Activities on
prefrontal and parietal lobes are also found on the results obtained
by CDR methods. However, one observes that the result obtained
by LRT1 is the most focused on the medial pre-central gyrus with
very limited activity anywhere else. The L1 CDR method
performed only slightly worse than LRT1. All other CDR methods
show activity in many regions other than M1. The results obtained
with LRT1 are consistent with those reported in the previous

Table 2
Results of Bonferroni post hoc tests on the mean and standard deviation of
the ED, PUS, and PFS indices

Fig. 3. The error distance for the three simulated EEG data sets using the
eight inverse methods. The gray error bars on each of the data points
represent the standard error.

Source localization results on real EEG data

The results of different inverse methods on SEPs resulting from

electrical stimulation of the subject’s wrist during the time window
from 20 to 25 ms after the stimulus are represented in Fig. 6. As
shown in Fig. 6, the majority of activities obtained by all the
methods stay on the contralateral hemisphere. Inverse results on
SEPs from shoulder, elbow, and wrist shift from medial to lateral
along the central sulcus. (SEP results of simulating elbow and
shoulder are not represented here.) Qualitatively, one can observe
In this table, values in the upper (white cells) and lower triangle (gray cells)
that the inverse results obtained by using method MDP1 or MDP2 are P values of the mean and standard deviation, respectively. The table
have a large variance even during a short time window (5 ms). And should be read from row to column, e.g., the element in first row and third
the mean locations obtained by MDP1 and MDP2 are located at pre- column of (c) shows that the mean of PFS for the L1 method is significantly
central gyrus (i.e., motor cortex). These results are not consistent larger (+) than the mean for the L2 method, where +/P b 0.05, ++/P b
with results published in the previous literature, which suggests that 0.01, +++/P b 0.001, ++++/P b 0.0001.
 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382 377

Although the evaluation on the source localization ability of two or

three different methods was already reported previously, this is the
first paper that compared these many inverse methods using both
simulated and experimental data sets. Our evaluation results
provide useful information for choosing the appropriate cortical
source localization techniques for scientific and clinical purposes.

Evaluation methods

In order to evaluate the localization ability of the different

source localization methods, we designed three types of simu-
lated EEG data sets. Results obtained with each of the simulated
EEG data sets represented various facets of the accuracy of the
inverse methods. Simulated data sets I and II allowed for the

Fig. 4. The percentage of undetected source number for the three simulated
EEG data sets using the eight inverse methods. The gray error bars on each
of the data points represent the standard error.

literature and demonstrated that M1 is the primary source of

activity related to arm activity.
In general, inverse results obtained by applying LRT1 to
SEPs and ERPs are qualitatively better than other methods,
because it provides results that mirror those obtained with direct
cortical recordings or with other noninvasive neuroimaging
modalities.

Discussions and Conclusions

In order to evaluate different cortical source localization Fig. 5. The percentage of false-detected source number for the three
methods, we applied 8 different inverse methods available in simulated EEG data sets using the eight inverse methods. The gray bars on
CURRY software to both simulated and experimental EEG data. each of the data points represent the standard error.
378 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382

Fig. 6. Inverse calculation results for SEPs resulting from electrical stimulation of the medial nerve at wrist in a subject. The figure legends are the same as
those in Fig. 2. The green line in each of the subplots is the central sulcus.

evaluation of source localization ability; and data sets II and III
allowed for the evaluation of the spatial resolution of the various
techniques.
Besides simulated EEG data, we also evaluated the different
inverse methods using experimental EEG data. Two typical
experimental EEG data sets (i.e., SEPs and ERPs) were used to
qualitatively evaluate results obtained with each of the inverse
methods. Our assessments are based on our current knowledge of
regions of cortical sensory motor activity associated with sensory
evoked or motor-related potentials.

Fig. 7. Inverse calculation results for ERPs related to isometric shoulder abduction torque generated by a subject. The figure legends are the same as those in
Fig. 2. The green line in each of the subplots is the central sulcus.
 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382
In case of applying inverse methods to simulated EEG data,
where the real source locations are known, three different
indices, i.e., the error distance (ED), percentage of undetected
source number (PUS), and percentage of falsely-detected source
number (PFS), were used to quantify the performance of the
different inverse methods. A method with lower ED, PUS, and
PFS is able to detect the source location more accurately. Based
on the definition of ED, there is a positive correlation between
ED and PUS/PFS, i.e., a larger PUS or PFS will cause an increase
in ED. The correlation coefficients between ED and PUS/PFS are
shown in Table 3. Because of the highly significant positive
correlation between ED and PUS/PFS, one can only refer to ED
for the evaluation of the source localization ability of an inverse
method.
Factors that may influence our evaluation methods
Several factors that may influence the performance of inverse
methods were not studied in this paper. First of all, previous reports
showed that results of inverse calculation are sensitive to the
conductivity of head (Benar and Gotman, 2002; Gencer and Acar,
2004; van Burik and Peters, 2000; van den Broek et al., 1998; Wen
and Li, 2001). Among conductivities of brain, skull, and scalp, the
inverse result is most sensitive to the conductivity of skull (Ferree
et al., 2000). However, the real conductivity of skull in vivo is still
unknown. Direct measurements showed a large variance on this
value ranging from 15 mS/m to 80 mS/m (Akhtari et al., 2002;
Ferree et al., 2000; Hoekema et al., 2003). The conductivity values
used in this paper were listed in Table 1 and corresponded to a
skull-scalp-ratio equal to 1:15. Inverse results were also obtained
using skull-scalp-ratios equal to 1:26 and 1:80. In all of these cases,
LRT1 provides the overall best results.
Secondly, the regulation parameter, k, is crucial to the
performance of inverse methods. As a general rule, the degree
of regularization (k) should increase with the level of noise in the
data. Based on this rule, k is often determined using v 2 and L-
curve methods. Unfortunately, these two methods are not
available for the minimum Lp norm and LRT Lp norm methods
in CURRY. In this paper, a fixed value (k = 1) was used for all
the methods. In order to limit the potential influence caused by
using a non-optimized regulation parameter, we controlled the
signal to noise ratio (SNR) of all the EEG data, including both the
simulated and experimental data, to a very consistent level,
ranging from 9 to 10. Regardless, evaluation of the different
source localization methods by using an optimized regulation
parameter remains desirable.
The shape of simulated EEG data may be another factor that
will affect the evaluation results. All the three simulated EEG data
sets are composed of equal strength dipoles (1 AAmm). No decay
Table 3
Pearson product moment correlations between the three evaluation indices
and P values obtained using ANOVA for linear regression
Simulated EED Simulated EED Simulated EED
data I data II data III
PUS PFS PUS PFS PUS PFS
ED 0.704 0.629 0.864 0.718 0.362 0.245
(b0.0001) (b0.0001) (b0.0001) (0.0386) (0.02) (0.350)
In this table, P values and R values are shown inside and outside the
parentheses, correspondingly.
379
of dipole strength was included. The square-wave-shaped sources
used here are sub-optimal for LRT Lp norm methods, which are
designed on the assumption that the source activities are distributed
smoothly on the source layer. We postulate that using simulated
EEG data with decays will further improve the performance of
LRT Lp norm methods.
When evaluating the CDR methods, a threshold, which was set
at a 98% confidence level, was used to define a source. It is
obvious that the choice of threshold also affects our evaluation
results. Since currents estimated by different CDR methods have
different distributions, we choose a data-specific threshold for each
of the inverse results (i.e., we first fitted the estimated cortical
currents using a Weibull distribution; then we set the threshold at
the 98% confidence level). Even with such a threshold, the
residuals of the fitting procedure are slightly different depending
on results obtained by different localization methods. Therefore,
there possibly is a small bias for ED, PUS, and PFS depending on
the choice of source localization method.
Using a data-specific threshold at 98% confidence level, all
CDR methods tend to provide a relatively large PFS; however,
different methods have a different distribution of the falsely-
detected sources. For example, when using LRT1 method, the
falsely-detected sources are distributed around the real source;
while when using other CDR methods, more falsely-detected
sources are distributed around the edge of the cortex (see Fig. 8,
where the purple crosses represent the detected source positions).
By increasing the threshold, we can effectively increase the PUS
and decrease the PFS of CRD methods. Since a positive correlation
existed both between ED and PUS as well as between ED and PFS,
the impact of an increased threshold on ED is not straightforward.
Other factors, such as the number of electrodes, the depth of the
source, and the SNR of signals, also have impact on the
performance of the various inverse methods. The impact of these
parameters were not included as part of the current investigation.
Accuracy of the different source localization methods
According to the results obtained using 3 different simulated
EEG data sets, the LRT L1 norm method showed the overall
highest accuracy. We further summarized the comparison on the
mean and the standard deviation (inside the parentheses) of the
three evaluation indices between LRT1 to other inverse methods in
Table 4. (Note, only significant results were represented.) As
shown in Table 4, the LRT1 method has either significant better or
similar performance when compared to the other investigated
methods for all of the three indices, which showed an overall
superior ability to detect the location of the source. When applying
the various methods to experimental EEG data, including both
SEPs and ERPs, the LRT1 method also provides the best estimates.
These findings further confirm that among all the tested methods,
LRT1 has the most accurate and least variable source localization
abilities.
Our results further indicate that a weighted minimum norm
method, such as LRT-Lp, has better source localization ability than
the regular minimum norm methods. These results are in
accordance with those reported in previous literature (Baillet and
Garnero, 1997; Babiloni et al., 2000b, 2004). Furthermore, we also
demonstrate that CDR methods provide physiologically meaningful
results while MDP1 and MDP2 solutions failed in many situations,
which is also in agreement with previous reported results (He et al.,
2002a; Pascual-Marqui et al., 1994). Within the same CDR
380 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382

Fig. 8. Inverse calculation results for simulated EEG data set II with radius of the two dipole areas equal to 8 mm. Figure legends were the same as those in
Fig. 2. The purple dots represent the estimated and real source dipole positions.

regulation method (minimum norm or weighted minimum norm) a the method is defined as the minimum radius of dipole area that a
lower Lp value will improve the localization results. method can distinguish between the two areas. Simulated EEG data
set III consists of two moving dipoles with the distance between
Spatial resolution of different methods them increasing from 4 mm to 20 mm. We defined the spatial
resolution of an inverse method applied to this type of data as the
The purpose of applying inverse methods is to improve the minimum distance between two moving dipoles that can be
spatial resolution of EEG measurement. However, to date, no distinguished.
quantitative examination has been performed to provide the spatial Results on simulated EEG data II showed that when the radiuses
resolution of an inverse method. For this reason, we designed of the two dipole areas are equal to 5 mm, none of the CDR methods
simulated EEG data sets II and III to quantify the spatial resolution could separate the two areas (i.e., only one mode is detected). When
of different inverse methods for 163-channel EEG signals with a increasing the radiuses from 5 mm to 6 mm, L15, L1, LRT15, and
SNR around 9–10. Simulated EEG data set II has two dipole-areas LRT1 methods showed an un-separated area, however, with a
with increasing radiuses (ranging from 5 mm to 10 mm) and a bimodal distribution. Furthermore, two separated areas can be found
fixed distance (29 mm) between their centers. When applying an when using a higher threshold. These results suggest that the spatial
inverse method to simulated EEG data II, the spatial resolution of resolution for these methods is around 6 mm. Further increasing the
radiuses of the two dipole areas did not cause the merge of the two
Table 4 dipole areas. This is a very interesting result. Considering that by
Comparison of the three evaluation indices between LRT1 and the other increasing the radiuses of the two dipole areas while keeping the
methods distance between their centers the same, the two source regions were
Methods LRT1 getting closer, which we thought would require a better spatial
resolution to separate them. However, this is not correct. A source
ED PUS PFS
image with a smaller radius has higher frequency components. In
L1 ++++ (+++) ++++ (++++) ++ (+++) the extreme case, the frequency component of a source image with
L15 ++++ ++++ (+++) ++++ only a single dipole is the highest. A source containing higher
L2 ++++ ++++ (++++) ++++
frequency components requires a higher spatial resolution to detect.
LRT15 ++++ ++
LRT2 ++++ (+) ++++ (++++) ++++
Therefore, our results showed a better ability to separate two areas
MDP1 ++++ ++++ ++++ when the radiuses of the two source dipole-areas are increasing.
MDP2 ++++ (++++) ++++ ++++ (++++) The effect of the frequency component of the image would also
explain the results obtained using simulated EEG data set III which
The table should be read from row (method) to column (index), e.g., the
element in first row and second column shows that the mean of ED for L1
showed that when increasing the distance between the two dipoles,
method is significantly larger than the mean of ED for LRT1 method, where no obvious change in performance of the CDR methods was
+/P b 0.05, ++/P b 0.01, +++/P b 0.001, ++++/ P b observed ( P values of post hoc analysis on the effect of radius are
0.0001. The results with and without parentheses represent the P value of larger than 0.05). Obviously, changing the distance between the
the standard deviation and mean, respectively. two dipoles does not change the frequency component of source
 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382
image enough to improve the performance of the CDR source
localization methods.
Another aspect that may have influenced our results is the use
of the 5 mm grid of source layer to constrain the source locations.
We believe that by using a finer grid the spatial resolution for L15,
L1, LRT15, and LRT1 methods would further improve the
performance of these CDR methods.
Practical implications for the application of the various source
localization methods
Dipole source models are more widely used in the previous
literature to estimate centers of cortical activity. We believe that
they are only suitable for the situation where a single source is
presented. In order to model multiple sources, multiple dipole
methods were developed. However, our results showed that even
using the correct number of dipoles, multiple dipole methods could
not determine the accurate location of all the sources.
CDR methods have the ability to determine the current
distribution of brain activity. Among all the CDR methods
investigated in this paper, the LRT1 method showed the highest
accuracy and, therefore, is the most promising method for source
localization estimation especially when different cortical regions
are expected to be simultaneously active during a particular
experimental paradigm.
Acknowledgments
This work was supported by a National Institutes of Health
Grant 1R03HD39804-01A1 and by a donation from the RIC
Women’s Board. The authors want to acknowledge the construc-
tive suggestions made by Drs. Leo Towle, Wim van Drongelen,
Aryre Nehorai, and the reviews of NeuroImage during the
preparation and revision of the manuscript.
References
Akhtari, M., Bryant, H.C., Mamelak, A.N., Flynn, E.R., Heller, L., Shih,
J.J., Mandelkern, M., Matlachov, A., Ranken, D.M., Best, E.D.,
DiMauro, M.A., Lee, R.R., Sutherling, W.W., 2002. Conductivities of
three-layer live human skull. Brain Topogr. 14, 151 – 167.
Asanuma, H., Rosen, I., 1972. Topographical organization of cortical
efferent zones projecting to distal forelimb muscles in the monkey. Exp.
Brain Res. 14, 243 – 256.
Babiloni, F., Babiloni, C., Locche, L., Cincotti, F., Rossini, P., Carducci, F.,
2000a. High-resolution electro-encephalogram: source estimates of
Laplacian-transformed somatosensory-evoked potentials using a real-
istic subject head model constructed from magnetic resonance images.
Med. Biol. Eng. Comput. 38 (5), 512 – 519.
Babiloni, F., Babiloni, C., Locche, L., Cincotti, F., Rossini, P.M., Carducci,
F., 2000b. High-resolution electro-encephalogram: source estimates of
Laplacian-transformed somatosensory-evoked potentials using a real-
istic subject head model constructed from magnetic resonance images.
Med. Biol. Eng. Comput. 38, 512 – 519.
Babiloni, F., Babiloni, C., Carducci, F., Romani, G.L., Rossini, P.M.,
Angelone, L.M., Cincotti, F., 2004. Multimodal integration of EEG and
MEG data: a simulation study with variable signal-to-noise ratio and
number of sensors. Hum. Brain Mapp. 22, 52 – 62.
Baillet, S., Garnero, L., 1997. A Bayesian approach to introducing
anatomo-functional priors in the EEG/MEG inverse problem. IEEE
Trans. Biomed. Eng. 44, 374 – 385.
381
Beer, R.F., Johnston, E.R., Eisenberg, E., 1995. Vector Mechanics for
Engineers: Statistics. McGraw-Hill, New York.
Benar, C.G., Gotman, J., 2002. Modeling of post-surgical brain and skull
defects in the EEG inverse problem with the boundary element method.
Clin. Neurophysiol. 113, 48 – 56.
Cheyne, D., Kristeva, R., Deecke, L., 1991. Homuncular organization of
human motor cortex as indicated by neuromagnetic recordings. Neuro-
sci. Lett. 122, 17 – 20.
Cincotti, F., Babiloni, C., Miniussi, C., Carducci, F., Moretti, D., Salinari,
S., Pascual-Marqui, R., Rossini, P.M., Babiloni, F., 2004. EEG
deblurring techniques in a clinical context. Methods Inf. Med. 43,
114 – 117.
Dowman, R., Schell, S., 1999. Innocuous-related sural nerve-evoked and
finger-evoked potentials generated in the primary somatosensory and
supplementary motor cortices. Clin. Neurophysiol. 110, 2104 – 2116.
Druschky, K., Kaltenhauser, M., Hummel, C., Druschky, A., Huk, W.,
Neundorfer, B., Stefan, H., 2003. Somatosensory evoked magnetic
fields following passive movement compared with tactile stimulation of
the index finger. Exp. Brain Res. 148, 186 – 195.
Emerson, R., Pedley, T., 1984. Generator sources of median somatosensory
evoked potentials. J. Clin. Neurophysiol. 1, 203 – 218.
Ferree, T.C., Eriksen, K.J., Tucker, D.M., 2000. Regional head tissue
conductivity estimation for improved EEG analysis. IEEE Trans.
Biomed. Eng. 47, 1584 – 1592.
Ferrier, D., 1875. Experiments on the brain of monkeys. Proc. R. Soc.
London, B: Biol. Sci. 24, 409 – 432.
Fuchs, M., Drenckhahn, R., Wischmann, H.-A., Wager, M., 1998. An
improved boundary element method for realistic volume-conductor
modeling. IEEE Trans. Biol. Med. Eng. 45, 980 – 997.
Gencer, N.G., Acar, C.E., 2004. Sensitivity of EEG and MEG measure-
ments to tissue conductivity. Phys. Med. Biol. 49, 701 – 717.
Haan, H., Streb, J., Bien, S., Rosler, F., 2000. Individual cortical current
density reconstructions of the semantic N400 effect: using a generalized
minimum norm model with different constraints (L1 and L2 norm).
Hum. Brain Mapp. 11, 178 – 192.
Hamalainen, M.S., Ilmoniemi, R.J., 1994. Interpreting magnetic fields of the
brain: minimum norm estimates. Med. Biol. Eng. Comput. 32, 35 – 42.
Hari, R., Forss, N., 1999. Magnetoencephalography in the study of human
somatosensory cortical processing. Philos. Trans. R. Soc. London, B:
Biol. Sci. 354, 1145 – 1154.
Hashimoto, I., 2000. High-frequency oscillations of somatosensory evoked
potentials and fields. J. Clin. Neurophysiol. 17, 309 – 320.
He, B., Yao, D., Lian, J., Wu, D., 2002a. An equivalent current source
model and Laplacian weighted minimum norm current estimates of
brain electrical activity. IEEE Trans. Biomed. Eng. 49, 277 – 288.
He, B., Zhang, X., Lian, J., Sasaki, H., Wu, D., Towle, V.L., 2002b.
Boundary element method-based cortical potential imaging of somato-
sensory evoked potentials using subjects’ magnetic resonance images.
NeuroImage 16, 564 – 576.
Hines, M., 1940. Movements elicited from precentral gyrus of adult
chimpanzees by stimulation with sine wave currents. J. Neurophysiol.
3, 442 – 466.
Hoekema, R., Wieneke, G.H., Leijten, F.S., van Veelen, C.W., van Rijen,
P.C., Huiskamp, G.J., Ansems, J., van Huffelen, A.C., 2003. Measure-
ment of the conductivity of skull, temporarily removed during epilepsy
surgery. Brain Topogr. 16, 29 – 38.
Ilmoniemi, R.J., 1993. Models of source currents in the brain. Brain Topogr.
5, 331 – 336.
Komssi, S., Huttunen, J., Aronen, H.J., Ilmoniemi, R.J., 2004. EEG
minimum-norm estimation compared with MEG dipole fitting in the
localization of somatosensory sources at S1. Clin. Neurophysiol. 115,
534 – 542.
Nenonen, J.T., Hamalainen, M.S., Ilmoniemi, R.J., 1994. Minimum-norm
estimation in a boundary-element torso model. Med. Biol. Eng.
Comput. 32, 43 – 48.
NeuroScan, L., 1999. Curry Reference Guide. pp. 101 – 109.
Nunez, P., 1981. Quantitative states of neocortex. In: Nunez, P. (Ed.),
382 J. Yao, J.P.A. Dewald / NeuroImage 25 (2005) 369–382
Neocortical Dynamics and Human EEG Rhythms. Oxford University
Press, Inc., New York, pp. 3 – 67.
Pascual-Marqui, R.D., Biscay-Lirio, R., 1993. Spatial resolution of neuro-
nal generators based on EEG and MEG measurements. Int. J. Neurosci.
68, 93 – 105.
Pascual-Marqui, R.D., Michel, C.M., Lehmann, D., 1994. Low resolution
electromagnetic tomography: a new method for localizing electrical
activity in the brain. Int. J. Psychophysiol. 18, 49 – 65.
Penfield, W., Rasmussen, T., 1950. The cerebral cortex of man. A Clinical
Study of Localization of Function. Macmillan, New York.
Phillips, C., Rugg, M.D., Fristont, K.J., 2002. Systematic regularization of
linear inverse solutions of the EEG source localization problem.
NeuroImage 17, 287 – 301.
Rao, S.M., Binder, J.R., Hammeke, T.A., Bandettini, P.A., Bobholz, J.A.,
Frost, J.A., Myklebust, B.M., Jacobson, R.D., Hyde, J.S., 1995.
Somatotopic mapping of the human primary motor cortex with
functional magnetic resonance imaging. Neurology 45, 919 – 924.
Recuero, E., 1999. The use of somatosensory evoked potentials for cortical
localization during surgical operations. Rev. Neurol. 28, 588 – 590.
Restuccia, D., Valeriani, M., Insola, A., Lo Monaco, M., Grassi, E., Barba,
C., Le Pera, D., Mauguiere, F., 2002. Modality-related scalp responses
after electrical stimulation of cutaneous and muscular upper limb
afferents in humans. Muscle Nerve 26, 44 – 54.
Ricamato, A., Dhaher, Y., Dewald, J., 2003. Estimation of active cortical
current source regions using a vector representation scanning approach.
J. Clin. Neurophysiol. 20, 326 – 344.
Salenius, S., Portin, K., Kajola, M., Salmelin, R., Hari, R., 1997. Cortical
control of human motoneuron firing during isometric contraction. J.
Neurophysiol. 77, 3401 – 3405.
Sanes, J.N., Donoghue, J.P., Thangaraj, V., Edelman, R.R., Warach, S.,
1995. Shared neural substrates controlling hand movements in human
motor cortex. Science 268, 1775 – 1777.
Schaefer, M., Muhlnickel, W., Grusser, S., Flor, H., 2002. Reproducibility
and stability of neuroelectric source imaging in primary somatosensory
cortex. Brain Topogr. 14, 179 – 189.
Scherg, M., 1990. Fundamentals of dipole source potential analysis. In:
Gandori, F., Hoke, M., Romani, G.L. (Eds.), Auditory Evoked Magnetic
Fields and Potentials: Advances in Audiology, vol. 6. Karger, Basel,
pp. 40 – 70.
Scherg, M., Buchner, H., 1993. Somatosensory evoked potentials and
magnetic fields: separation of multiple source activities. Physiol. Meas.
14 (Suppl. 4A), A35 – A39.
Schieber, M.H., 2001. Constraints on somatotopic organization in the
primary motor cortex. J. Neurophysiol. 86, 2125 – 2143.
Schieber, M.H., Hibbard, L.S., 1993. How somatotopic is the motor cortex
hand area? Science 261, 489 – 492.
Sherrington, C.S., 1947. The Integrative Action of the Nervous System
(1906), 2nd ed. Cambridge University Press (reprinted 1973).
Sonoo, M., 2000. Anatomic origin and clinical application of the
widespread N18 potential in median nerve somatosensory evoked
potentials. J. Clin. Neurophysiol. 17, 258 – 268.
Stenbacka, L., Vanni, S., Uutela, K., Hari, R., 2002. Comparison of mini-
mum current estimate and dipole modeling in the analysis of simulated
activity in the human visual cortices. NeuroImage 16, 936 – 943.
Towle, V.L., Khorasani, L., Uftring, S., Pelizzari, C., Erickson, R.K., Spire,
J.P., Hoffmann, K., Chu, D., Scherg, M., 2003. Noninvasive identifica-
tion of human central sulcus: a comparison of gyral morphology, func-
tional MRI dipole localization, and direct cortical mapping. NeuroImage.
Valeriani, M., Le Pera, D., Niddam, D., Arendt-Nielsen, L., Chen, A., 2000.
Dipolar source modeling of somatosensory evoked potentials to painful
and nonpainful median nerve stimulation. Muscle Nerve 23, 1194 – 1203.
van Burik, M.J., Peters, M.J., 2000. Estimation of the electric conductivity
from scalp measurements: feasibility and application to source local-
ization. Clin. Neurophysiol. 111, 1514 – 1521.
van den Broek, S.P., Reinders, F., Donderwinkel, M., Peters, M.J., 1998.
Volume conduction effects in EEG and MEG. Electroencephalogr. Clin.
Neurophysiol. 106, 522 – 534.
Waberski, T.D., Buchner, H., Perkuhn, M., Gobbele, R., Wagner, M.,
Kucker, W., Silny, J., 1999. N30 and the effect of explorative finger
movements: a model of contribution of the motor cortex to early
somatosensory potentials. Clin. Neurophysiol. 110, 1589 – 1600.
Waberski, T.D., Gobbele, R., Herrendorf, G., Steinhoff, B.J., Kolle, R.,
Fuchs, M., Paulus, W., Buchner, H., 2000. Source reconstruction of
mesial-temporal epileptiform activity: comparison of inverse techni-
ques. Epilepsia 41, 1574 – 1583.
Wagner, M., Kohler, T., Fuchs, M., Kastner, J., 2000. An extended
source model for current density reconstructions. In: Nenonen, J.,
Ilmoniemi, R.J., Katila, T. (Eds.), Proc. 12th Int. Conf. Biomagnetism,
Biomag2000. Helsinki University of Technology, Espoo, Finland,
pp. 749 – 752.
Wang, J.Z., Williamson, S.J., Kaufman, L., 1992. Magnetic source images
determined by a lead-field analysis: the unique minimum-norm least-
squares estimation. IEEE Trans. Biomed. Eng. 39, 665 – 675.
Wang, J.Z., Kaufman, L., Wiliamson, S.J., 1993. Imaging regional changes
in the spontaneous activity of the brain: an extension of the minimum-
norm least-squares estimate. Electroencephalogr. Clin. Neurophysiol.
86, 36 – 50.
Weibull, W., 1939. A statistical theory of the strength of materials.
Ingeniors Vetenskaps Akademiens Handlingar, R. Swed. Inst. Eng.
Res., vol. 153. Stockholm, Sweden.
Wen, P., Li, Y., 2001. Comparison study of different head model structures
with homogeneous/inhomogeneous conductivity. Australas. Phys. Eng.
Sci. Med. 24, 31 – 36.
Williamson, S.J., Kaufman, L., 1981. Biomagnetism. J. Magn. Magn. Mat.
22, 129 – 201.
Woolsey, C.N., 1958. Organization of somatic sensory and motor areas of
the cerebral cortex. In: Warlow, H.F., Woolsey, C.N. (Eds.), The
Biological and Biochemical Bases of Behavior. University of Wisconsin
Press, Madison, pp. 63 – 81.
Yao, J., Dewald, J.P.A., 2003. Evaluation of different cortical potential
imaging methods using simulated EEG data. IEEE EMBS, Proc. 25th
Ann. Int. Conf. IEEE, Cancun, Mexico, pp. 2128 – 2131.