METHODS

published: 09 June 2021

doi: 10.3389/fnins.2021.569918

ConvDip: A Convolutional Neural

Network for Better EEG Source
Imaging
Lukas Hecker 1,2,3,4*, Rebekka Rupprecht 5 , Ludger Tebartz Van Elst 1,2 and
Jürgen Kornmeier 1,2,3
1
Department of Psychiatry and Psychotherapy, University of Freiburg Medical Center, Freiburg, Germany, 2 Faculty of
Medicine, University of Freiburg, Freiburg, Germany, 3 Institute for Frontier Areas of Psychology and Mental Health (IGPP),
Freiburg, Germany, 4 Faculty of Biology, University of Freiburg, Freiburg, Germany, 5 Machine Learning Lab, University of
Freiburg, Freiburg, Germany

The electroencephalography (EEG) is a well-established non-invasive method in

neuroscientific research and clinical diagnostics. It provides a high temporal but low
spatial resolution of brain activity. To gain insight about the spatial dynamics of the EEG,
one has to solve the inverse problem, i.e., finding the neural sources that give rise to the
recorded EEG activity. The inverse problem is ill-posed, which means that more than one
configuration of neural sources can evoke one and the same distribution of EEG activity
on the scalp. Artificial neural networks have been previously used successfully to find
Edited by:
either one or two dipole sources. These approaches, however, have never solved the
Alard Roebroeck, inverse problem in a distributed dipole model with more than two dipole sources. We
Maastricht University, Netherlands present ConvDip, a novel convolutional neural network (CNN) architecture, that solves
Reviewed by: the EEG inverse problem in a distributed dipole model based on simulated EEG data.
Annalisa Pascarella,
Istituto per le Applicazioni del Calcolo We show that (1) ConvDip learned to produce inverse solutions from a single time point
“Mauro Picone” (IAC), Italy of EEG data and (2) outperforms state-of-the-art methods on all focused performance
Miguel Castelo-Branco,
Coimbra Institute for Biomedical
measures. (3) It is more flexible when dealing with varying number of sources, produces
Imaging and Translational Research less ghost sources and misses less real sources than the comparison methods. It
(CIBIT), Portugal produces plausible inverse solutions for real EEG recordings from human participants. (4)
*Correspondence: The trained network needs <40 ms for a single prediction. Our results qualify ConvDip
Lukas Hecker
[email protected] as an efficient and easy-to-apply novel method for source localization in EEG data, with
high relevance for clinical applications, e.g., in epileptology and real-time applications.
Specialty section:
This article was submitted to Keywords: EEG-electroencephalogram, artificial neural networks, convolutional neural networks (CNN), inverse
Brain Imaging Methods, problem, machine learning, electrical source imaging
a section of the journal
Frontiers in Neuroscience

Received: 05 June 2020

1. INTRODUCTION
Accepted: 14 April 2021
1.1. The EEG and the Inverse Problem
Published: 09 June 2021
Electroencephalography (EEG) is among the most used imaging techniques for noninvasive
Citation: measurements of electromagnetic brain activity. Its main advantage over other methods (e.g.,
Hecker L, Rupprecht R, Tebartz Van
functional magnetic resonance imaging [fMRI]) is the high temporal resolution, which comes at
Elst L and Kornmeier J (2021)
ConvDip: A Convolutional Neural
the cost of a considerably low spatial resolution (Luck, 2014). As a consequence, EEG was mainly
Network for Better EEG Source used to study temporal brain dynamics at fine time scales. In the past decades, however, there
Imaging. Front. Neurosci. 15:569918. has been a steady growth of interest in the neural sources underlying the EEG signal (Koles,
doi: 10.3389/fnins.2021.569918 1998; Pascual-Marqui, 1999; Grech et al., 2008; He et al., 2018; Michel and Brunet, 2019). The

Frontiers in Neuroscience | www.frontiersin.org 1 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

non-invasive estimations of neural generators, based on their variance (LCMV) beamforming approach is a spatial filter
projections to the scalp electrodes/sensors, constitutes an inverse that assumes that neural sources are uncorrelated and in
problem. Without further constraints, it is ill posed because it which portions of the data that do not belong to the signal are
lacks an unique solution, since multiple configurations of neural suppressed (Van Veen et al., 1997). Drawbacks of the LCMV
sources can produce identical topographies of electrical activity approach are the susceptibility to imprecisions in the forward
at the scalp (see e.g., Nunez and Srinivasan, 2006). model and that correlated sources are often not found.
Invasive multimodal methods (e.g., using combined EEG Growing interest toward a Bayesian notation of the inverse
and electrocorticography) help to bridge the gap between scalp problem could be observed in the past two decades (Friston et al.,
recordings and neural generators, and thus in handling the 2008; Wipf and Nagarajan, 2009; Chowdhury et al., 2013). One
inverse problem in this constellation. However, access to these prominent Bayesian approach is the maximum entropy on the
methods is limited and the conclusions that can be drawn mean (MEM) method, which aims to make the least assumptions
are constrained by various factors such as placement of the on the current distribution by maximizing entropy (Amblard
cortical electrodes or coverage of brain areas that project to et al., 2004; Grova et al., 2006; Chowdhury et al., 2013).
the scalp electrodes. Combined EEG-fMRI has also been shown
as a useful tool in providing insight into the spatiotemporal
dynamics of the EEG (Ritter and Villringer, 2006). However, the 1.3. Artificial Neural Networks and Inverse
costs of this technique are considerably high and the relation Solutions
between electromagnetic and metabolic dynamics is yet not Artificial neural networks (ANN)-based inverse solutions follow
fully understood. a data-driven approach and were of increasing interest in the past
years (Awan et al., 2019). A large number of simulated EEG data
samples is used to train an ANN to correctly map electrode-space
1.2. Classical Approaches to the EEG signals to source-space locations (Jin et al., 2017). Long training
Inverse Problem periods are usually required for an ANN to generalize beyond the
Introducing some constraints on the solution, one can solve training set. After successful training, it is capable of predicting
the inverse problem or at least reduce the number of possible the coordinates and orientations of a dipole correctly, given only
solutions. One approach is the equivalent current dipole model, the measurements at the scalp electrodes without further priors
which is based on the assumption that the source of a signal, (Zhang et al., 1998; Abeyratne et al., 2001).
measured with the EEG, can be modeled by a single (or Robert et al. (2002) reviewed the literature on ANN-based
sometimes few) dipole(s) (Kavanagk et al., 1978; Scherg, 1990; methods to solve the inverse problem of single dipoles and
Delorme et al., 2012). Although single-dipole sized sources are found that all reports achieved localization errors of <5%.
too small to generate detectable scalp potentials (Nunez and The low computational time (once trained) and robustness to
Srinivasan, 2006), they can produce reasonable results (Ebersole, measurement noise was highlighted when compared to classical
1994; Lantz et al., 1996; Willemse et al., 2016; Sharma et al., 2018). approaches. However, the classical approaches were capable of
A more physiologically realistic approach is the distributed achieving zero localization error in single-dipole simulations
dipole model in which activity is expected to extend over larger under zero-noise conditions (Hoey et al., 2000).
areas of the brain (as opposed to tiny dipoles). Distributed In an effort to predict source locations of two dipoles, Yuasa
dipole models aim to find the 3D distribution of neural activity et al. (1998) presented a feed-forward network with two hidden
underlying the EEG measurement (Hämäläinen and Ilmoniemi, layers and achieved localization errors in a range between 3
1994; Pascual-Marqui et al., 1994). A distributed dipole model and 9%. They found that if multiple simulated dipoles have a
proposes that sources of the EEG are better modeled using sufficient distance among each other the localization success of
hundreds to many thousand dipoles and therefore aim to find a ANNs even for multiple simulated sources is equal to classical
distributed activity that can explain the EEG data. This model can approaches, making a strong case for the feasibility of ANNs for
be viewed opposed to single- to few-dipole models, which assume the EEG inverse problem, already in 1998. Although ANNs as
that EEG data can be sufficiently modeled using point sources. such gained popularity in the past decade on other areas (e.g.,
A popular family of distributed dipole solutions is the image classification or natural language processing), the idea of
Minimum Norm Estimates (MNE; Ioannides et al., 1990; ANN-based solutions to the EEG inverse problem received little
Hämäläinen and Ilmoniemi, 1994), which aim to find the further attention.
source configuration that minimizes the required power to Only very recently several studies about ANN-based solutions
generate a given potential at the scalp electrodes. Low to the inverse problem have been published. Cui et al. (2019)
Resolution Electromagnetic Tomography (LORETA) is a famous showed that a neural network can be trained to reconstruct the
proponent of the MNE-family that assumes sources to be position and time course of a single source using a long-short
smoothly distributed (Pascual-Marqui et al., 1994). In the term memory (LSTM) recurrent neural network architecture
most sophisticated version, exact LORETA (eLORETA, Pascual- (Hochreiter and Schmidhuber, 1997). LSTMs allow to not only
Marqui, 2007) showed zero localization error when localizing use single time instances of (e.g., EEG) data but instead learn
single sources in simulated data (Pascual-Marqui, 2007). from temporally lagged information.
Another popular family of inverse solutions is the In a recent work, Razorenova et al. (2020) showed that the
beamforming approach. The linear constrained minimum inverse problem for cortical potential imaging could be solved

Frontiers in Neuroscience | www.frontiersin.org 2 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

using a deep U-Net architecture, once more providing a strong serves the purpose to avoid the inverse crime. The inverse crime
case for the feasibility of ANNs to solve the EEG inverse problem is committed when the same GM is used to both simulate
(Fedorov et al., 2020). data and calculate inverse solutions, leading to overoptimistic
Tankelevich (2019) showed that a deep feed-forward network results (Wirgin, 2004). First, electrode positions were changed
can find the correct set of source clusters that produced a given by adding random normal distributed noise N (0, 2 mm) to the
scalp signal. To our knowledge, this was the first ANN approach X, Y, and Z coordinate of each electrode. This resulted in an
to calculate distributed dipole solutions. average displacement per electrode of≈ 3.7 mm. Furthermore,
In the very recent years, convolutional neural networks we changed the tissue conductivity in the BEM solution to
(CNNs) have proven to be a useful tool in a steadily increasing 0.332 S/m2 for brain and scalp tissue and 0.0113 S/m2 for skull
number of domains, like image classification (Krizhevsky et al., tissue (values were adapted from Wolters et al., 2010). This
2012), natural language processing (Kim, 2014), and decoding of step was done in order to introduce some alterations in the
single trial EEG (Schirrmeister et al., 2017). CNNs are capable volume conduction of the brain-electric signals, which are never
of learning patterns in data with a preserved temporal (e.g., time known precisely a priori. Importantly, this changes the skull-to-
sequences) or spatial (e.g., images) structure by optimizing filter brain conductivity ratio from 1:50 to 1:25. Last, we changed the
kernels that are convolved with a given input. Two famous CNNs spacing of the dipoles along the cortical sheet from icosahedral to
are AlexNet (Krizhevsky et al., 2012) and VGG16 (Simonyan octahedral with a higher resolution. This resulted in 8,196 dipoles
and Zisserman, 2014) that won the ImageNet classification in the source model of the AGM, all of which were placed at
competition in 2012 and 2014, respectively. slightly different locations on the cortex compared to the source
In the current study, we explored the feasibility of model of the GM. The idea of increasing the resolution of the
CNNs to solve the EEG inverse problem. Specifically, we source model to avoid the inverse crime was adapted from Kaipio
constructed a CNN, named ConvDip, that is capable of detecting and Somersalo (2006). All simulations in the evaluation section
multiple sources using training data with biologically plausible were carried out using this AGM, whereas the calculation of the
constraints. ConvDip solves the EEG inverse problem using a inverse solutions of cMEM, eLORETA, and LCMV beamformer
distributed dipole solution in a data-driven approach. ConvDip was based on the GM. Likewise, data used for training ConvDip
was trained to work on single time instances of EEG data and was simulated using only the GM.
predicts the position of sources from potentials measured with
scalp electrodes.
2.2. Simulations
2. METHODS Training data for ConvDip were created using the GM, whereas
the data for the evaluation of the model were created using the
A Python package to create a forward model, simulate data, train AGM. Since AGM and GM have different source models, we
an ANN, and perform predictions is available at https://github. implemented a function which transforms a source vector jAGM
com/LukeTheHecker/ESINet. to a source vector jGM . This was achieved by K nearest-neighbor
interpolation, i.e., each value in the target space was assigned the
2.1. Forward Model average value of the K nearest neighbors in the initial space. K
To simulate EEG data realistically, one has to solve the was set to 5 neighbors. Since the source model of AGM contained
forward problem and construct a generative model (GM). An almost twice as many dipoles as that of the GM, a source vector
anatomical template was used as provided by the Freesurfer translated from AGM to GM will effectively loose total source
image analysis suite (http://surfer.nmr.mgh.harvard.edu/) called power. We therefore normalized the resulting source vector jGM
fsaverage (Fischl et al., 1999). to have equal energy (sum of all dipole moments) as jAGM .
We calculated the three shell boundary element method In summary, different sets of simulated EEG samples were
(BEM, Fuchs et al., 2002) head model with 5120 vertices per shell generated to (1) create training data for ConvDip and (2) for
using the python package MNE (v 19, Gramfort et al., 2013) and evaluation of ConvDip and other inverse solutions.
the functions make_bem_model and make_bem_solution therein. Each simulation contained at least one dipole cluster, which
The conductivity, measured by Siemens per square meter, was set can be considered as a smooth region of brain activity. A
to 0.3 S/m2 for brain and scalp tissue and 0.06 S/m2 for the skull. dipole cluster was generated by selecting a random dipole in
The source model was created with p = 5124 dipoles along the the cortical source model and then adapting a region growing
cortical surface provided by Freesurfer with icosahedral spacing. approach as described in Chowdhury et al. (2013). In brief,
We chose q = 31 electrodes based on the 10–20 system (Fp1, F3, we recursively included all surrounding neighbors starting from
F7, FT9, FC5, FC1, C3, T7, TP9, CP5, CP1, Pz, P3, P7, O1, Oz, O2, a single seeding location, thereby creating a larger source
P4, P8, TP10, CP6, CP2, Cz, C4, T8, FT10, FC6, FC2, F4, F8, and extent with each iteration. The number of iterations define the
Fp2). The leadfield K ∈ Rq×p was then calculated using the head neighborhood order s, where the first order s1 entails only the
model with dipole orientations fixed orthogonal to the cortical single selected dipole. Each seed location was assigned a dipole
surface (see e.g., Michel and Brunet, 2019 for explanation). moment between 5 and 10 Nano-Ampere Meter (nAm). The
Additionally, we created a second forward model that we neighboring dipoles were assigned attenuated moments based
refer to as the alternative generative model (AGM). The AGM on the distance to seeding location. The attenuation followed a
will be used to simulate data for the model evaluation and Gaussian distribution with a mean of the seeding dipole moment

Frontiers in Neuroscience | www.frontiersin.org 3 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

and a standard deviation of half the radius of the dipole cluster, done in order to avoid the inverse crime (for details, see Methods
yielding smooth source patches. section 2.1).
After generating this spatial pattern, we added a temporal
dimension to the data as follows. We considered an epoch length
of 1 s at a sampling frequency of 100 Hz (i.e., t = 100 time 2.3. ConvDip
points). The time course was modeled by a central peak of 100 In this section, the architecture of ConvDip, a CNN which solves
ms temporal extension using half a period of a 5 Hz sinusoidal, the inverse problem, is described and a mathematical description
surrounded by zeros. Each dipole moment was then multiplied in the classical notation of inverse solutions is given.
by this time course. The simulated source J ∈ Rp×t of p = 5,124
dipoles was then projected to q = 31 EEG electrodes M ∈ Rq×t 2.3.1. I/O of ConvDip
by solving the forward problem using the leadfield K ∈ Rq×p : Since ConvDip was designed to operate on single time instances
of EEG data, we extracted from each simulated data sample
M = KJ (1) only the time point of the peak source activity. The EEG data
at this time frame was then interpolated on a 2D image of
size 7 × 11. This is necessary, given that the architecture of
To generate realistic training data, we added real noise from
ConvDip (Figure 1) requires a 2D matrix input to perform spatial
pre-existing EEG recordings, conducted with the same set of
convolutions on. Note that this interpolation procedure does not
electrodes as described above. Therefore, from a set of different
add information to the EEG data. The output of ConvDip is a
raw EEG recordings, we extracted 200 random segments. These
vector of size 5,124 corresponding to the dipoles in the source
EEG segments were filtered beforehand using a band-pass
model (see Figure 1).
between 0.1 and 30 Hz. Each segment contained 1 s of EEG
data. The sampling frequency was reduced from 1,000 to 100
Hz. Baseline correction was applied based on the first 100 ms, 2.3.2. ConvDip Architecture
and the resulting data were re-referenced to common average. The design and training of the neural network was accomplished
For each sample, we then created 20 EEG trials by adding a using the Tensorflow (Abadi et al., 2016) and Keras (Chollet,
randomly selected noise segment to the simulated EEG. Prior to 2015) libraries, which are based on Python 3. Training was
this, the noise was scaled to yield a signal-to-noise ratio (SNR) of partially accomplished on a Nvidia Titan V graphical processing
1 within single trials. The average of these 20 trials, i.e., the event- unit (GPU) and an Nvidia RTX 2070.
related The architecture of ConvDip (Figure 1) is inspired by CNNs
√ potential, should therefore exhibit a theoretical SNR of 4.5 for image classification, in which the input layer is forwarded
(1 × (20)) at the ERP peak.
The extent of the sources was defined between two (s2 ) and to a series of convolutional layers followed by a series of fully
five (s5 ) neighborhood orders, which corresponds to diameters connected layers followed by the output layer (e.g., AlexNet;
between ≈ 21 and ≈ 58 mm (or 19–91 dipoles). Krizhevsky et al., 2012). In typical CNNs, the convolution layers
are followed by pooling layers to reduce dimensionality and to
2.2.1. Training Data promote invariance to an object’s position Deru and Ndiaye
For training ConvDip, we simulated in total 100,000 samples (2019). Notice that in ConvDip no pooling layers were used
using the GM as described above. Each sample contained since spatial information would get lost or at least blurred.
between 1 and 5 source clusters of extent between 2 and 5 While invariance to the position of an object is desirable in
neighborhood orders. Since ConvDip operates on single time image classification tasks (what-task), it would be detrimental
instances of EEG or ERP data, we only used the EEG at the signal in the case of source localization problems (where-task). Finally,
peak as input. we decided for a moderately shallow architecture with a single
convolution layer followed by a fully connected layer and an
2.2.2. Evaluation Data output layer.
Two separate sets of simulations of 1,000 samples each were We decided to use Rectified Linear Units (ReLUs) as activation
created for the evaluation. The first set contained samples of functions after each layer (Nair and Hinton, 2010; Glorot et al.,
single source clusters with varying extents from two to seven 2011). ReLUs have shown to exhibit the best performance in
neighborhood orders, which we will refer to as single-source our preliminary tests compared to alternatives (e.g., sigmoid
set. The second set contained samples with varying numbers of function). After each of the inner layers, we added a batch
source clusters from 1 to 10 and with varying extents from 2 to normalization layer (Ioffe and Szegedy, 2015) in order to speed
7 neighborhood orders or diameters from ≈ 21 to ≈ 58 mm (or up training and to reduce internal covariance shifts. In the
16–113 dipoles in the AGM). We will refer to this as the multi- current version of ConvDip, the output layer consists of 5,124
source set. Source extents have different diameters for different neurons, which equals the number of voxels in the modeled
source models, especially in our case where the source model brain. ReLU activation functions were used in the output layer.
of GM has roughly half the number of dipoles compared to Typical CNNs for regression use linear activation functions,
AGM. This explains why s7 for the AGM has approximately as however, since predictions are by definition non-negative in our
many dipoles as s6 in the GM. Note that both sets were created application, ReLUs appeared to us as an appropriate alternative.
using simulations of brain-electric activity based on an AGM that ConvDip can be described as a function φ : M 7→ ĵ, M ∈
differs from that used in the inversion/training process. This was R7×11 , ĵ ∈ Rp , with p = 5124 that maps a single time instance

Frontiers in Neuroscience | www.frontiersin.org 4 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

FIGURE 1 | ConvDip architecture. The values from a single time point of EEG data were interpolated to get a 7 × 11 matrix as an input (see illustration on the bottom
left). The subsequent convolution layer has only 8 convolution kernels of size 3 × 3 pixels. The convolution layer is followed by a fully connected (FC) layer consisting
of 512 neurons. Finally, the output layer contains 5124 neurons that correspond to the voxels in the brain (plotted on a cortical surface on the right) (This diagram was
created using a web application at http://alexlenail.me/NN-SVG/).

of 2D-interpolated EEG data M to a source vector ĵ: appropriate since it operates pixel-wise, i.e., the error does not
provide information on how close a wrong prediction is to the
ĵ = φ(M) (2) true source. This is especially problematic when ConvDip is
designed to find focal sources. Various approaches exist to tackle
The architecture of ConvDip starts with a convolution layer with this issue, e.g., by solving an optimal transport problem (e.g.,
only 8 filters Fi , i = {1, . . . 8} of size 3 × 3. The weights of these using Wasserstein metric) or by calculating the distance between
filters are determined during the training period. In the forward two sets of coordinates (e.g., Hausdorff distance). We decided
pass, the padded input EEG data M are convolved with each filter for the weighted Hausdorff distance (WHD) as described by
Fi , resulting in one feature map Gi ∈ R7×11 . Ribera et al. (2019)1 .
The WHD has shown to perform well on image segmentation
Gi = M ∗ Fi (3) tasks and allowed for a fast convergence of ConvDip compared to
MSE. Note that WHD requires a normalization of the prediction
The feature maps Gi , i ∈ {1, . . . , 8} are stacked to a tensor G ∈ and target to ensure all values are between 0 and 1. Therefore,
R7×11×8 and reshaped to a vector g̃ ∈ R616 (also called flattening) only positional information is regarded in this loss. The rationale
to enable a connection to the next FC Layer. The flattened vector for this is to disentangle the problem of estimating absolute
g̃ consists of 616 output nodes. Each node is connected to every dipole moments and that of estimating dipole positions.
neuron of the following FC Layer. As mentioned above, ConvDip predicts source locations
For each of the 512 neurons in the hidden fully connected without correct global scaling (see also Equation 2). To obtain
layer, we transform its input g̃ using the weight vector w, bias the true amplitude of the sources, we used Brent’s method (Brent,
b, and the activation function h: 1971) to find a scalar ŝ that minimizes the mean squared error
between the forward solution (x̂) of the predicted sources (ĵ) and
z = h(wT g̃ + b) (4) the unscaled EEG vector x:
The hidden layer is finally connected to the output layer of
5,124 neurons. 1 X
m
ŝ = arg min (x̂j · s) − xj (5)
2.3.3. Optimization and Loss Function s m
j=1
Convolution filters, weights, and biases were optimized using
adaptive momentum estimation (ADAM, Kingma and Ba, 2014)
with default settings as suggested by the authors (learning_rate = The scalar ŝ can then be used to scale the prediction.
0.001, β1 = 0.9, β2 = 0.999, ǫ = 108 ).
We tried out various loss functions for the training of
ConvDip. The mean squared error loss is a classical loss for 1 Implementation was adapted from https://github.com/N0vel/weighted-
regression tasks. For EEG inverse solutions, however, it is not hausdorff-distance-tensorflow-keras-loss/.

Frontiers in Neuroscience | www.frontiersin.org 5 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

2.4. Implementation of cMEM, eLORETA, therefore captures how well the source extent is captured.
and LCMV AUCfar on the other hand sampled the negatives from voxels
To evaluate ConvDip, we calculated inverse solutions on the far away from true sources, therefore capturing possible false
same set of simulations using eLORETA, LCMV beamformer, positives in the estimated source vector ĵ. Far negatives were
and cMEM, and compared the different methods with each other. sampled from the 50% of the farthest negatives to the next
eLORETA and LCMV were carried out by the active dipole. The overall AUC was then calculated by taking
implementations in the Python library MNE. Using the the average of AUCclose and AUCfar .
same head model as described in section 2.1, each inverse 2) Mean localization error (MLE): The Euclidean distance
algorithm was subjected to each sample of the evaluation set between the locations of the predicted source maximum and
as described in section 2.2. The first 400 ms of each trial were the target source maximum is a common measure of inverse
used to estimate the noise covariance matrices. Regularization solution accuracy as it captures the ability to localize the
of the noise covariance was established using the standard MNE source center accurately. MLE was calculated between the
procedure as described in Gramfort et al. (2014). We choose positions of the maxima of ĵ and j. This metric is only suitable
the regularization parameter for eLORETA inverse solutions at for calculating MLE when a single source patch is present.
λ2 = 91 and for LCMV beamforming we set the data covariance For multiple sources, we adapted the following procedure.
regularization to λ = 0.05, which both correspond to MNE’s First, we identified local maxima in both the true source
default parameters. Dipole orientations were restricted to be vector j and the estimated source vector ĵ. First, local maxima
fixed orthogonal to the cortical surface. were identified where a voxel value is larger than all of its
cMEM inverse solutions were calculated with Brainstorm neighboring voxels. This then yields many local maxima,
(Tadel et al., 2011), which is documented and freely available for which had to be filtered in order to achieve reasonable results.
download online under the GNU general public license (http:// First, we removed all maxima whose neighbors were not
neuroimage.usc.edu/brainstorm). The same template brain sufficiently active (< 20% of the maximum). This takes
(fsaverage) was used to calculate a forward model in Brainstorm. care of false positive maxima that do not constitute an active
Furthermore, the BEM solution was calculated using the same cluster of voxels. Then, we removed those maxima that had a
parameters as in MNE. cMEM inverse solutions were calculated larger maximum voxel in close neighborhood within a radius
using a neighborhood order of 4 with temporally stable clusters. of 30 mm. These procedures result in a list of coordinates
Precomputed noise covariance matrices were imported from of maxima for both j and ĵ. We then calculated the pairwise
MNE Python to ensure that inverse solutions were calculated Euclidean distances and between the two lists of coordinates
under the same conditions in Brainstorm and MNE. After of maxima. For each true source, we identified the closest
calculating all inverse solutions, the files were exported again for estimated source and calculated the MLE by averaging of
further evaluation in Python. these minimum distances. We further labeled those estimated
sources that were ≥ 30mm away from the next true maximum
as ghost sources. True maxima that did have an estimated
2.5. Evaluation Metrics
source within a radius of 30mm were labeled as found sources,
We calculated a number of performance metrics to assess the
whereas those true maxima that did not have an estimated
quality of an inverse solution. Note that for each sample in the
maximum within a radius of 30mm were labeled as missed
evaluation sets, we only analyzed the central peak activity.
sources. Finally, we calculated the percentage of found sources,
1) Area under the ROC curve: We calculated the area under the i.e., the ratio of the number of correctly identified sources and
receiver operator curve (ROC) to assess the ability to estimate the number of all true sources.
the correct spatial extent of sources. We adapted a similar 3) Mean squared error (MSE): The MSE is calculated voxel-
procedure as described in Grova et al. (2006) and Chowdhury wise between the true source vector j and the predicted source
et al. (2013) with minor changes to the procedure. The vector ĵ:
dipole clusters in the target source vector j have a Gaussian
distribution of dipole moments in our simulations. Therefore,
1X
p
we had to normalize all members of all dipole clusters (i.e. the MSE = (ji − ĵi )2 (6)
target source vector) to unit amplitude (i.e. values between 0 p
i=1
and 1). The estimated source vector ĵ was normalized between
0 and 1 by division of the maximum. The AUC metric requires 4) Normalized mean squared error (nMSE): The nMSE is
equal as many positives as there are negatives in the data. calculated by first normalizing both ĵ and j to values
However, in our simulations only few dipoles were active between 0 and 1. Then, the voxel-wise MSE is calculated as
(positives) compared to those not being active (negatives). described above.
Therefore, we adapted the procedure as described by Grova
et al. (2006) and Chowdhury et al. (2013) by calculating two
types of AUC: Both AUCs included all positives and only 2.6. Statistics
differ in the selection of negatives. AUCclose contained negative Statistical comparison between the outcomes of different inverse
examples that closely surrounded the positive voxels by solutions was done using an unpaired permutation test with
sampling randomly from the closest 20% voxels. This metric 106 permutations. The rationale behind this choice is that most

Frontiers in Neuroscience | www.frontiersin.org 6 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

TABLE 1 | Comparison of inverse algorithm performance for samples containing a single source cluster.

Inverse solution AUC [%] (SD) MSE (SD) nMSE (SD) MLE [mm] (SD)

ConvDip 98.31 (3.41) 3.9 · 10−19 (2.9 ∗ 10−19 ) 0.0033 (0.0015) 11.05 (4.96)
cMEM 90.99 (7.88) 4.3 · 10−13 (3.2 · 10−12 ) 0.0087 (0.0045) 18.80 (7.37)
eLORETA 85.06 (6.22) 6.4 · 10−08 (1.1 · 10−07 ) 0.0436 (0.0122) 13.26 (6.22)
LCMV 94.96 (4.04) 0.4 (0.1) 0.1274 (0.0356) 20.52 (6.83)

Source clusters were of varying spatial extents from two to seven neighborhood orders. AUC: Area under the receiver operator curve. MSE: Mean squared error, nMSE: normalized
mean squared error, MLE: mean localization error. Best performances are highlighted in bold font.

distributions did not meet the criteria for parametric and rank- et al. (2013) suggested that an AUC of 80% and above can be
based tests. Additionally, Cohens d is given for each statistical considered acceptable.
test (Cohen, 1992). This criterion is met by all inverse algorithms in the present
evaluation (Table 1). ConvDip achieves the best scores in this
2.7. Evaluation With Real Data comparison with an overall AUC of 98%, closely followed by
To evaluate the performance of ConvDip and the other inverse LCMV (95%) and cMEM (91%). Furthermore, ConvDip inverse
algorithms in a more realistic set-up, we used data of a real solutions show the highest global similarity with nMSE at
EEG recording. The data were recorded while a single participant 0.0033, which is less than half of the error yielded by cMEM.
(first author of this article) viewed fast presentations of faces Moreover, the other two methods are far behind (Table 1). MLE
and scrambled faces. The participant had to indicate whether a of ConvDip is lowest with 11.05 mm, which is remarkably small
face was presented by button press. The duration of a stimulus considering the variance introduced by the AGM. eLORETA
presentation was 600 ms, followed by 200 ms of black screen. yields significantly larger MLE than ConvDip (diffmedian =
Note that 300 trials of both faces and scrambled faces were 2.21mm, p = 10−6 , d = 0.34).
presented, which corresponds to a total of 4 min.
The EEG was recorded with 31 electrodes of the 10–20 system
using the ActiCap electrode cap and the Brain Vision amplifier 3.1.1.1. AUC
ActiCHamp. Data was sampled at 1, 000Hz and filtered online In Figure 2 (left graph), the AUC for each inverse algorithm is
using a band-pass of 0.01–120 Hz. The EEG data were imported depicted per simulated source extent. ConvDip is able to reach
to MNE Python and filtered using a band-pass filter between best performance for extents of s3 and s4 (AUCconvdip,s3 ,s4 = 99%)
0.1 and 30 Hz. Data were re-referenced to common average. with attenuated AUC especially for larger sizes (AUCconvdip,s7 =
The trials in which faces were presented were then selected and 94%). The AUC drops by ≈ 5.2% from s3 to s7 . cMEM, an inverse
baseline corrected by subtracting the average amplitude in the algorithm aiming at estimating both location and size of source
interval –0.06 to 0.04 s relative to stimulus onset. clusters, shows similar properties. cMEM shows it is best AUC at
s2 (AUCcMEM,s2 = 93%). The AUC decreases by ≈ 5.5% from
3. RESULTS the smallest to the largest source extent, which is comparable to
ConvDip. We conclude that while ConvDip yields significantly
In this section, we evaluate the performance of ConvDip higher AUC overall (p ≤ 10−6 , d > 0.40), its behavior to
and compare it to state-of-the-art inverse algorithms, namely varying source sizes is similar to that of cMEM. The other inverse
eLORETA, LCMV beamformer, and cMEM. Note that the algorithms perform significantly worse (Figure 2).
evaluation set was not part of the training set of ConvDip, hence
it is unknown to the model. Furthermore, all samples in the
evaluation set were created using the AGM in order to avoid the 3.1.1.2. MLE
inverse crime. Exemplary and representative samples are shown MLE of ConvDip is affected by source extent, as MLEconvdip,s2
in Appendices A, B. is 39% higher compared to MLEconvdip,s7 . This observation does
not apply to the MLE yielded by inverse solutions of cMEM,
3.1. Evaluation With Single Source Set which is fairly stable (7% variation). Overall, MLE was smaller
3.1.1. Evaluation of Source Extents for ConvDip inverse solutions compared to the other three
We will now evaluate the ability of ConvDip to estimate the inverse algorithms. One particular exception is the case of small
correct size of sources and to correctly localize sources with source clusters (s2 ), for which eLORETA yields the most accurate
varying depth. For this purpose, we used the data samples from estimates (MLEeLORETA,s2 = 10.34mm, see Figure 2).
the single-source set. An exemplary and representative sample is
shown in Appendix–A. 3.1.2. Evaluation of Source Eccentricity
One of the advantages of inverse algorithms such as cMEM Next, we evaluated the ability of ConvDip to correctly localize
over minimum norm solutions is their capability to estimate source clusters of different eccentricity (i.e., distance from all
not only the position of a source cluster but also its spatial electrodes). The further away a source is from the electrodes, the
extent. This can be tested using the AUC metric. Chowdhury harder it is to localize due to volume conduction.

Frontiers in Neuroscience | www.frontiersin.org 7 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

FIGURE 2 | Area under the curve and mean localization error for each inverse solution algorithm grouped by the source extent. AUC reflects capability of finding
source at correct location, estimating its size and reducing false positives. MLE reflects the capability to correctly localize the center of the source cluster. Simulations
used for this analysis contained only a single source cluster per sample with varying size from two to seven neighborhood orders.

3.1.2.1. AUC systematically more robust behavior of ConvDip compared to

In Figure 3 left, the AUC metric is shown for each inverse all other inverse algorithms. With depth, however, the ability to
algorithm grouped by bins of eccentricity. Bins were defined estimate the extent of a source cluster is reduced.
from 20 mm (deep source clusters) to 80 mm (superficial source
clusters). Since deeper sources are harder to localize accurately, 3.1.3. Evaluation of Dipole Moments
AUC is expected to be correlated with eccentricity. However, Finally, we utilized the single-source set to evaluate how well
this could be verified only for ConvDip (r = 0.27, p < 10−16 ) the moments of the dipoles was reconstructed by ConvDip and
and cMEM (r = 0.31, p < 10−23 ). For LCMV beamformer, the other inverse algorithms. We calculated the MSE between
depth is barely associated with AUC (r = 0.07, p < 0.05) and each true and predicted source vector (as described in section
for eLORETA no such association could be found (r = 0, p > 2.5). We find that the scaled predictions of ConvDip yielded the
0.9). Despite these interesting dynamics, ConvDip shows for all lowest MSE (3.93 · 10−19 ± 2.91 · 10−19 ), followed by cMEM
eccentricities higher AUC compared to almost all other inverse (4.32·10−13 ±3.24·10−12 ) and eLORETA (6.40·10−8 ±1.05·10−7 ).
algorithms. Only the LCMV beamformer shows a non-significant We excluded LCMV beamformer since the approach does not
tendency (p < 0.1, d = 0.2) for a higher AUC for low eccentricity yield dipole moments that can be interpreted in this way. This
of 20mm (median = 93.39% ± 3.19) than ConvDip (median = result is less surprising in the light of the aforementioned results
93.44% ± 5.35). since ConvDip is able to estimate the extent of sources well, which
facilitates the estimation of correct dipole moments through the
3.1.2.2. MLE scaling process (refer to section 2.3.3 for an explanation of our
MLE is also affected by source depth as can be seen in scaling approach).
Figure 3 right. We calculated the ratio between the MLE of the
deepest source clusters (eccentricity = 20 mm) to the MLE of
superficial source clusters (eccentricity = 80 mm) in order to 3.2. Evaluation With Multiple Sources
calculate depth sensitivity of each inverse algorithm, henceforth In a realistic situation, the EEG may pick up signals originating
called drop-off. cMEM shows a drop-off of 64%, eLORETA has from many regions in the brain simultaneously, aggravating
a drop-off of 23%, and LCMV beamformer has a drop-off of the inversion process. Due to the low spatial resolution
18%. ConvDip shows the lowest MLEs across eccentricities, of the EEG, not all sources can be identified reliably.
compared to the other inverse methods. Moreover, ConvDip To evaluate the performance of the inverse algorithms in
shows an inverted U-shaped behavior, with lowest MLE for the reconstructing more challenging samples of the multi-source set,
deepest and most superficial source clusters. MLE of ConvDip we calculated the aforementioned metrics that are summarized
decreases slightly with depth of the source cluster and are not in Table 2.
correlated with eccentricity (r = −0.02, p > 0.5), whereas
this relation was found for all other inverse algorithms: cMEM 3.2.1. AUC
(r = −0.30, p < 10−21 ), eLORETA (r = −0.15, p < 10−5 ), and Overall, the AUC for the multi-source set is low since the
LCMV beamformer (r = −0.09, p < 0.01). desirable 80% AUC is undercut by each inverse algorithm.
Concluding, we find no drop-off in localization accuracy with ConvDip achieved the highest AUC of 78%, thereby surpassing
depth of source clusters when using ConvDip, highlighting a the next best competitor LCMV by ≈ 7.2%.

Frontiers in Neuroscience | www.frontiersin.org 8 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

FIGURE 3 | Area under the curve and mean localization error for each inverse algorithm grouped by the source eccentricity. Simulations used for this analysis
contained only a single source cluster per sample of varying size from two to seven neighborhood orders.

TABLE 2 | Comparison of inverse algorithm performance for samples containing multiple source clusters.

Inverse solution AUC [%] (SD) MSE (SD) nMSE (SD) MLE (SD) % of sources found (SD)

ConvDip 78.29 (9.39) 1.6 · 10−18 (4.4 · 10−19 ) 0.0087 (0.0042) 28.18 (12.10) 62.71 (22.38)
cMEM 70.76 (8.95) 1.2 · 10−12 (3.9 · 10−12 ) 0.0139 (0.0057) 38.61 (16.11) 50.12 (24.02)
eLORETA 69.94 (7.54) 1.9 · 10−07 (1.6 · 10−07 ) 0.0559 (0.0152) 29.51 (8.98) 60.71 (22.60)
LCMV 71.13 (10.30) 0.4 (0.1) 0.1592 (0.0415) 28.09 (5.50) 62.36 (23.67)

Samples contained between one and ten source clusters of varying spatial extent. Each cell contains the median and mean absolute deviation of the medians over all samples in the
multi-source set. For the “percentage of sources found” metric, the mean was calculated instead of median. AUC: Area under the receiver operator curve; nMSE: normalized mean
squared error, MLE: mean localization error, Percentage of sources found: The percentage of sources whose maxima was correctly localized. Best performances are highlighted in
bold font.

3.2.2. nMSE cMEM (r = 0.60, p < 10−97 ), eLORETA (r = 0.73, p < 10−163 ),
nMSE is substantially smaller for ConvDip inverse solutions and LCMV (r = 0.73, p < 10−167 ).
compared to all other inverse algorithms with 60% lower nMSE This relationship was also found to exists between MLE and
compared to the next best performing algorithm cMEM. the number of source clusters. For large numbers of source
clusters (≥ 5), LCMV shows the lowest MLE (29.41mm) or
3.2.3. MLE 4.24mm less than ConvDip (p = 10−6 , d = 0.60, Figure 5). This
LCMV beamforming is the only inverse solution (of the tested may be a result of the inherent noise-suppressing properties of
ones) with lower MLE than ConvDip (Table 2), albeit the the LCMV beamformer.
absolute difference of the median MLEs is small (diff_MLE ≈ A core result of this analysis is that although ConvDip was
0.1mm, p = 10−6 , d = 0.22). trained with samples containing between 1 and 5 source clusters,
it is capable of inverting more challenging samples as well
3.2.4. Percentage of Found Sources (Figure 4).
The percentage of sources found was fairly similar for ConvDip, Furthermore, we showed that discrepancies between the
eLORETA, and LCMV and 10% lower for cMEM (Table 2). assumed GM and the true bio-physical properties underlying
Figure 4 shows the results for the metrics AUC and nMSE an EEG measurement (e.g., tissue conductivity, precise electrode
grouped by the number of sources present in the samples. It locations, spacing of the discrete dipole model) can be handled
is apparent that the accuracy of all inverse solutions depends by ConvDip.
strongly on the total number of sources that contribute to the ConvDip provides inverse solutions that are valid for samples
EEG signal on the scalp since for each inverse algorithm, the generated with the AGM and different configurations (namely,
results worsen with increasing numbers of source clusters present different numbers of source clusters). This shows that ConvDip
in the sample. Pearson correlation between AUC and the number is able to generalize well beyond samples within the training
of present source clusters was strong and statistically significant set. This in turn confirms the validity of this purely data-driven
for each inverse algorithm: ConvDip (r = 0.70, p < 10−147 ), approach to solve the EEG inverse problem.

Frontiers in Neuroscience | www.frontiersin.org 9 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

FIGURE 4 | Area under the curve and normalized mean squared error for each inverse algorithm grouped by the number of present source clusters. The right-most
columns in each of the graphs display the results for samples containing between 5 and 10 source clusters.

4. DISCUSSION
4.1. Overview
We have presented ConvDip, a CNN that aims to find extended
sources of EEG signals. The approach of using a neural network
to solve the inverse problem is distinct from classical methods,
mainly because it works purely data driven. All constraints on
the solution of ConvDip are given implicitly. The performance
of ConvDip was evaluated and tested against commonly used
inverse algorithms cMEM, eLORETA, and LCMV beamforming.
In the following, we discuss the significance of these results.

4.2. Using ANNs to Solve the Inverse

Problem FIGURE 5 | Mean localization error for each inverse algorithm grouped by the
One major result of this work is that the CNN ConvDip was number of present source clusters. The rightmost column displays the results
capable to reliably find the locations of neural sources that give for samples containing between 5 and 10 source clusters.
rise to the EEG. ConvDip transforms a 7 × 11 image to a
large vector of 5,124 voxels. In image classification, CNNs are
utilized to do the opposite: To extract hidden information from deviations from the GM source model grid), and provides
a large input matrix and output, a single value that indicates another remarkable advantage of these data-driven approach.
some property, e.g., the presence of a cat or a dog. Our work We evaluated the performance of ConvDip and the alternative
verifies that CNNs are capable to solve an underdetermined inverse algorithms concerning multiple objectives, the results of
problem. This issue was already addressed by Lucas et al. (2018), which we will discuss in the following.
who demonstrated the advantages of ANNs over analytical
approaches to solve the inverse problem. 4.3. Estimating Source Extent
We trained ConvDip on data generated by GM and The ability to estimate not only the location of a neural source
subsequently evaluated ConvDip on simulated samples that were but also its extent was tested for all inverse algorithms on the
generated by the AGM in order to avoid committing the inverse single-source set using the AUC and nMSE performance. We
crime. Overall, the performance of ConvDip was sometimes find that ConvDip reaches a nearly perfect AUC of ≈ 98%,
comparable but in most cases better than the classical inverse which is not met by any other inverse algorithm. nMSE, a
solutions cMEM, eLORETA, and LCMV beamformer. This measure of global similarity between the true and estimated
provides evidence that ConvDip learned to generalize beyond source, was smallest using ConvDip, with less than half the
samples in the training data and remains robust to deviations error compared to the next best inverse solutions yielded
from the training data. This holds for small deviations in the from cMEM (Table 1). ConvDip furthermore estimated different
position of the electrodes (±3.7 mm), tissue conductivities, but source sizes appropriately (Figure 2) with higher AUC than all
also in the configuration of sources (number and size of clusters, other inverse algorithms. Although these findings are remarkable

Frontiers in Neuroscience | www.frontiersin.org 10 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

we have to regard that the shape and size of the sources was 360) to calculate the iterative cMEM inverse solution using
rather predictable for ConvDip. All sources in the training brainstorm. Additionally, one has to provide the noise covariance
and evaluation set had a more or less circular shape due the matrix, which takes on average ≈ 45 ms to compute using the
region growing approach we chose. It may be presumably empirical estimation of MNE. In practical terms, ConvDip is
a large challenge for ConvDip to estimate ellipitical shapes capable to compute 31.25 inverse solutions per second (ips) on
or shapes that are even more deviant to a circle. Despite a GPU, whereas eLORETA reaches 7.75 ips LCMV beamformer
this caveat, it is remarkable how well ConvDip performs 10.20 ips on a CPU. Possibly, implementations of eLORETA and
in this comparison considering the variability introduced by LCMV beamforming that can be run on a GPU achieve lower
the AGM. computation times compared to CPU. The short computation
times of ConvDip are at a distinct advantage over other inverse
4.4. Localizing Deep Sources solutions, e.g., in real-time applications as in neurofeedback
A further surprising result for us during the evaluation was experiments or in the development of brain–computer interfaces.
that for ConvDip the MLE did not depend on the depth of When comparing our approach to the existing literature on
the source cluster, i.e., ConvDip is capable to correctly localize ANN-based inverse solutions, we can now confirm that an
even deep sources with nearly no drop-off (see Figure 3). This ANN not only provides competitively accurate single dipole
finding was a strong contrast to the findings from other inverse locations but also a distributed dipole model of the brain-electric
algorithms, which all tend to mislocate deeper sources more than activity, taking into account more than one dipole. This can
superficial ones. Although source depth did not influence the be attributed mostly to the rapid developments in the machine
MLE of ConvDip, we find that the AUC is indeed compromised learning domain such as the introduction of convolutional
by depth. This holds also true for cMEM but not for eLORETA layers (LeCun and Bengio, 1995) and the improvements of
and barely for LCMV beamformer. Overall, ConvDip achieved GPUs that render the training of large ANNs possible in
comparable AUCs and MLEs for various depth of source clusters, acceptable time.
compared to other inverse solutions, rendering it a viable option
for localizing deep sources. 4.7. Realistic Simulations
The simulation of EEG data in this study was based on
4.5. Performance With Multiple Active different assumptions on the number, distribution, and shape
Source Clusters of electromagnetic fields, which is inspired by Nunez and
In section 3.2, we evaluated how ConvDip performs when Srinivasan (2006). The validity of predictions of any inverse
many source clusters are simultaneously active. We observed solution is only granted if the assumptions on the origin of
a drop-off in AUC with increasing numbers of active source the EEG signal are correct. This is indeed one of the critical
clusters from ≈ 99% with single sources to ≈ 73% when 5 challenges toward realistic inverse solutions in general, both
or more sources were active simultaneously. Similar behavior for data-driven inverse solutions (such as ConvDip) as well as
was observed for all inverse algorithms tested (see Figure 4). classical inverse solutions with physiological constraints (e.g.,
But, ConvDip achieved the highest AUC compared to all eLORETA). Therefore, when interpreting the applicability of
alternatives tested, surpassing the best competitor LCMV by ConvDip to real data, it is important to specify and justify these
≈ 7.2%. Furthermore, ConvDip also outperformed all other assumptions. We show that the knowledge about brain-electric
inverse algorithms concerning nMSE (Table 2). Only when five activity underlying EEG signals is emulated by ConvDip after
or more source clusters are present, we find slightly lower MLE it was trained. However, it is evident from the present work,
using LCMV beamformer compared to ConvDip. This increased how critical the particular choice of parameters is with respect
localization ability of LCMV beamformer could be due to the to the performance, as ConvDip produces solutions that closely
active suppression of noise signals in the beamforming process. resemble the training samples (see Appendices A, B). A more
Another reason could be that ConvDip was never trained with sophisticated approach to this problem would be to diversify the
samples that contained more than five active source clusters, training data, e.g., by changing the region-growing procedure
therefore struggling to accurately localize at least a subset of the to produce sources with higher variety of shapes beyond simple
source clusters. spheric source clusters. Nonetheless, ConvDip finds reasonable
source clusters when presented with real data (see Appendix–C
4.6. Computation Speed for an example).
As already pointed out by Sclabassi et al. (2001), ANNs naturally An important observation in real EEG data is the inter-
yield faster inverse solutions than classical methods. On our individual variability, which is in part a direct result of individual
workstation (CPU: Intel i5 6400, GPU: Nvidia RTX 2070, 16 Gb brain anatomies. For ConvDip application to real EEG data,
RAM), one forward pass of ConvDip took on average ≈ 31 ms particularly with group data, it is thus advised to collect
on the GPU. When all necessary preprocessing steps (such as anatomical MRI brain scans of each individual subject. Individual
re-referencing, scaling, and interpolation to 2-D scalp maps) are neural networks then need to be trained based on each subjects’
taken into account, we reached ≈ 32 ms of computation time on individual anatomy provided by the MRI data. This may help to
average. In comparison, it takes ≈ 129 ms (≈factor 4) to calculate achieve accurate group-study source estimations. This, however,
the eLORETA inverse solution using MNE, ≈ 98 ms ms (≈factor raises the problem of computation time, since training individual
3) for LCMV beamforming and astonishing 11.55 s (≈factor CNNs is time consuming. A solution to this is an inter-individual

Frontiers in Neuroscience | www.frontiersin.org 11 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

transfer learning approach, where ConvDip is trained on one network for 2D-3D projections (Lin et al., 2018) or by using
subjects’ anatomy and fine-tuned for each additional subject with spatiotemporal information with a LSTM network (Cui et al.,
new training data of the individual anatomies. Fine tuning could 2019). One obvious further way to improve the performance of
be achieved by replacing the output-layer, lowering the learning ConvDip may be to increase the capacity of the neural network,
rate and retraining for only few epochs. Training time per se is e.g., by adding more layers. In our initial testing phase, we
another important topic, when considering CNNs to handle the tried different neural architectures without observing significant
inverse problem. improvements by adding more layers that justify the increase in
training time. Neural architecture search should be used to find a
4.8. Training Time compact, well-performing architecture to solve the EEG inverse
Although the availability of high-performing hardware resources problem (Liu et al., 2018).
(e.g., high performing graphic cards) is growing rapidly, the Finally, the development of a both meaningful and
training of the presented neural network architecture requires a computationally inexpensive loss function, which makes
considerable amount of processing power and processing time. training faster and allows for faster convergence, is key for
The version of ConvDip shown in this work required ≈ 10.5 h ANN-based approaches to the inverse problem that can run on
of training for 500 epochs and 100,000 samples of data using a ordinary PC within a reasonable amount of time.
NVIDIA RTX 2070 GPU. Improving the architecture of ConvDip
to shorten training time is an important task for future research,
especially when individualized models are required. Limiting the 4.10. Outlook
solution space to fewer voxels may be one way to decrease the We showed that ConvDip, a shallow CNN, is capable of solving
computation time since most trainable parameters are located the inverse problem using a distributed dipole solution. The
at the final fully connected layers. The easiest way would be association between single time points of EEG measurements
to reduce the number of parameters of the neural network, and underlying source distributions can be learned and used
e.g., by targeting a CNN with volumetric output of the ANN to predict plausible inverse solutions. These inverse solutions
corresponding to a volumetric source model of the brain. This were, furthermore, shown to globally reflect the topology of the
could spare the computationally expensive fully connected layers brain-electric activity given by the training data.
and reduce the number of weights in the network. Another Predictions yielded from ConvDip are in rare cases
possibility is to segment the cortical surface into parcels, i.e., comparable to but in the majority of cases better than the
clusters of neighboring dipole positions. This would reduce the here tested existing inverse algorithms in the measures we
size of the output layer by one to two orders of magnitude and focused on. Furthermore, the application of ConvDip to real
thereby reduce the number of trainable parameters dramatically. EEG data yields reasonable results. The fully trained ConvDip
The feasibility of such a dimension reduction is to be investigated requires only little computational costs of 32 ms, which makes
in future developments. it a promising tool for real-time applications where only single
Another computationally expensive processing step is the time points of EEG data are available.
generation of realistic artificial training data. In this study, we We want to emphasize the importance of realistic simulations
projected neural activity using a three-layer BEM head model. of real neural activity that can be measured by EEG. Although
Generating 100,000 simulations of neural activity took 12 min a lot of knowledge about the generation of scalp potentials
using a PC with CPU @ 4 × 4.3 GHz and 16 GB of RAM. Software was implemented in our EEG simulations, including more
that enables EEG researchers to perform physiologically realistic physiological constraints can further reduce the complexity
simulations of neural activity for their own ConvDip application of the inverse problem. One example is the estimation of
is available from different resources, e.g., from the MNE library cortical column orientation (Bonaiuto et al., 2019) or the
in Python (Gramfort et al., 2014). incorporation of empirical knowledge about region-specific
activity from in-vivo (e.g., electrocorticogram) studies. Taking
4.9. Further Perspectives for Improvement into account these additional aspects may further improve
The present version of ConvDip is trained on single time ConvDip’s performance.
instances of EEG data. From a single EEG data point, it The current ConvDip version was based on single time points
is, however, not possible to estimate a noise baseline. As a of artificial EEG data. Exploiting temporal aspects of brain
consequence, such a baseline has not yet been taken into account dynamics (neighboring time points) may provide additional
in ConvDip, which leaves room for improvements toward more valuable information and increase the predictive power that may
regularized and adaptive solutions that are computed flexibly further refine future ConvDip versions.
depending on noise conditions. In contrast, LCMV, eLORETA,
and cMEM explicitly make use of a noise covariance matrix,
which is used to regularize the inverse solution. ConvDip could
be improved by extending the data input to an EEG time series, DATA AVAILABILITY STATEMENT
allowing the neural network to learn spatiotemporal patterns.
Possible further advances could be made by preserving the The simulations and the ConvDip model can be
3D structure of the output space by, instead of a flattened reproduced using the python package located at:
output layer. This can be realized using a deconvolutional neural https://www.github.com/LukeTheHecker/ESINet.

Frontiers in Neuroscience | www.frontiersin.org 12 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

ETHICS STATEMENT procedure, and discussion. All authors participated in discussing

the methodology and evaluating the results and revising
Ethical review and approval was not required for the study on the manuscript.
human participants in accordance with the local legislation and
institutional requirements. The patients/participants provided FUNDING
their written informed consent to participate in this study.
The project was funded by the European Campus (Eucor)
AUTHOR CONTRIBUTIONS seeding money.

LH conceptualisation, initial idea, implementation, testing, ACKNOWLEDGMENTS

evaluation, and writing the manuscript. RR technical supervision
for mathematics, implementation, and artificial neural networks. The authors want to thank NVIDIA Corporation for supplying
LT conceptualisation, remarks on results, manuscript, and us with a Titan V GPU as part of the academic seeding program.
remarks on future perspectives. JK conceptualization, writing We thank Roman Tankelevich for his remarks in the early stage
and revising the manuscript, development of evaluation of the project.

REFERENCES for the cortical surface. Hum. Brain Mapp. 8, 272–284.

doi: 10.1002/(SICI)1097-0193(1999)8:4&lt;272::AID-HBM10&gt;3.0.CO;2-4
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, Friston, K., Harrison, L., Daunizeau, J., Kiebel, S., Phillips, C., Trujillo-Barreto,
G. S., et al. (2016). Tensorflow: large-scale machine learning on heterogeneous N., et al. (2008). Multiple sparse priors for the M/EEG inverse problem.
distributed systems. arXiv preprint arXiv:1603.04467. NeuroImage 39, 1104–1120. doi: 10.1016/j.neuroimage.2007.09.048
Abeyratne, U. R., Zhang, G., and Saratchandran, P. (2001). EEG source Fuchs, M., Kastner, J., Wagner, M., Hawes, S., and Ebersole, J. S. (2002).
localization: a comparative study of classical and neural network A standardized boundary element method volume conductor model. Clin.
methods. Int. J. Neural Syst. 11, 349–359. doi: 10.1142/S01290657010 Neurophysiol. 113, 702–712. doi: 10.1016/S1388-2457(02)00030-5
00813 Glorot, X., Bordes, A., and Bengio, Y. (2011). “Deep sparse rectifier neural
Amblard, C., Lapalme, E., and Lina, J.-M. (2004). Biomagnetic source detection networks,” in Proceedings of the Fourteenth International Conference on
by maximum entropy and graphical models. IEEE Trans. Biomed. Eng. 51, Artificial Intelligence and Statistics (Fort Lauderdale, FL), 315–323.
427–442. doi: 10.1109/TBME.2003.820999 Gramfort, A., Luessi, M., Larson, E., Engemann, D. A., Strohmeier, D., Brodbeck,
Awan, F. G., Saleem, O., and Kiran, A. (2019). Recent trends and advances in C., et al. (2013). MEG and EEG data analysis with MNE-Python. Front.
solving the inverse problem for EEG source localization. Inverse Prob. Sci. Eng. Neurosci. 7:267. doi: 10.3389/fnins.2013.00267
27, 1521–1536. doi: 10.1080/17415977.2018.1490279 Gramfort, A., Luessi, M., Larson, E., Engemann, D. A., Strohmeier, D.,
Bonaiuto, J. J., Afdideh, F., Ferez, M., Wagstyl, K., Mattout, J., Bonnefond, M., Brodbeck, C., et al. (2014). MNE software for processing MEG and
et al. (2019). Estimates of cortical column orientation improve MEG source EEG data. Neuroimage 86, 446–460. doi: 10.1016/j.neuroimage.2013.
inversion. bioRxiv 810267. doi: 10.1101/810267 10.027
Brent, R. P. (1971). An algorithm with guaranteed convergence for finding Grech, R., Cassar, T., Muscat, J., Camilleri, K. P., Fabri, S. G., Zervakis, M., et al.
a zero of a function. Comput. J. 14, 422–425. doi: 10.1093/comjnl/14. (2008). Review on solving the inverse problem in EEG source analysis. J.
4.422 Neuroeng. Rehabil. 5:25. doi: 10.1186/1743-0003-5-25
Chollet, F. (2015). Keras. Available online at: https://keras.io Grova, C., Daunizeau, J., Lina, J.-M., Bénar, C. G., Benali, H., and
Chowdhury, R. A., Lina, J. M., Kobayashi, E., and Grova, C. (2013). MEG Gotman, J. (2006). Evaluation of EEG localization methods using
source localization of spatially extended generators of epileptic activity: realistic simulations of interictal spikes. Neuroimage 29, 734–753.
comparing entropic and hierarchical bayesian approaches. PLoS ONE 8:e55969. doi: 10.1016/j.neuroimage.2005.08.053
doi: 10.1371/journal.pone.0055969 Hämäläinen, M. S., and Ilmoniemi, R. J. (1994). Interpreting magnetic fields
Cohen, J. (1992). Statistical power analysis. Curr. Direct. Psychol. Sci. 1, 98–101. of the brain: minimum norm estimates. Med. Biol. Eng. Comput. 32, 35–42.
doi: 10.1111/1467-8721.ep10768783 doi: 10.1007/BF02512476
Cui, S., Duan, L., Gong, B., Qiao, Y., Xu, F., Chen, J., et al. (2019). EEG source He, B., Sohrabpour, A., Brown, E., and Liu, Z. (2018). Electrophysiological source
localization using spatio-temporal neural network. China Commun. 16, 131– imaging: a noninvasive window to brain dynamics. Ann. Rev. Biomed. Eng. 20,
143. doi: 10.23919/JCC.2019.07.011 171–196. doi: 10.1146/annurev-bioeng-062117-120853
Delorme, A., Palmer, J., Onton, J., Oostenveld, R., and Makeig, S. Hochreiter, S., and Schmidhuber, J. (1997). Long short-term memory. Neural
(2012). Independent EEG sources are dipolar. PLoS ONE 7:e30135. Comput. 9, 1735–1780. doi: 10.1162/neco.1997.9.8.1735
doi: 10.1371/journal.pone.0030135 Hoey, G. V., Clercq, J. D., Vanrumste, B., de Walle, R. V., Lemahieu, I.,
Deru, M., and Ndiaye, A. (2019). Deep learning with tensorFlow, keras and D’Havé, M., et al. (2000). EEG dipole source localization using artificial
tensorFlow. js. Rheinwerk Comput. 1, 15–55. neural networks. Phys. Med. Biol. 45, 997–1011. doi: 10.1088/0031-9155/4
Ebersole, J. S. (1994). Non-invasive localization of the epileptogenic 5/4/314
focus by EEG dipole modeling. Acta Neurol. Scand. 89, 20–28. Ioannides, A. A., Bolton, J. P. R., and Clarke, C. J. S. (1990). Continuous
doi: 10.1111/j.1600-0404.1994.tb05179.x probabilistic solutions to the biomagnetic inverse problem. Inverse Probl. 6:523.
Fedorov, M., Koshev, N., and Dylov, D. V. (2020). “Deep learning for Non-invasive doi: 10.1088/0266-5611/6/4/005
cortical potential imaging,” in Machine Learning in Clinical Neuroimaging Ioffe, S., and Szegedy, C. (2015). Batch normalization: accelerating deep network
and Radiogenomics in Neuro-Oncology: Third International Workshop, MLCN training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167.
2020, and Second International Workshop, RNO-AI 2020, Held in Conjunction Jin, K. H., McCann, M. T., Froustey, E., and Unser, M. (2017). Deep convolutional
with MICCAI 2020, Lima, Peru, October 4–8, 2020, Proceedings, Vol. 12449, neural network for inverse problems in imaging. IEEE Trans. Image Proc. 26,
(Springer Nature), 45. 4509–4522. doi: 10.1109/TIP.2017.2713099
Fischl, B., Sereno, M. I., Tootell, R. B., and Dale, A. M. (1999). Joos, E., Giersch, A., Hecker, L., Schipp, J., Heinrich, S. P., Tebartz van
High-resolution intersubject averaging and a coordinate system Elst, L., et al. (2020). Large EEG amplitude effects are highly similar

Frontiers in Neuroscience | www.frontiersin.org 13 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

across Necker cube, smiley, and abstract stimuli. PLoS ONE 15:e0232928. Robert, C., Gaudy, J.-F., and Limoge, A. (2002). Electroencephalogram
doi: 10.1371/journal.pone.0232928 processing using neural networks. Clin. Neurophysiol. 113, 694–701.
Kaipio, J., and Somersalo, E. (2006). Statistical and Computational Inverse doi: 10.1016/S1388-2457(02)00033-0
Problems. Luxemburg: Springer Science & Business Media. Scherg, M. (1990). Fundamentals of dipole source potential analysis. Auditory
Kanwisher, N., and Yovel, G. (2006). The fusiform face area: a cortical region Evoked Magn. Fields Electr. Potent. Adv. Audiol. 6, 40–69.
specialized for the perception of faces. Philos. Trans. R. Soc. B Biol. Sci. 361, Schirrmeister, R. T., Springenberg, J. T., Fiederer, L. D. J., Glasstetter, M.,
2109–2128. doi: 10.1098/rstb.2006.1934 Eggensperger, K., Tangermann, M., et al. (2017). Deep learning with
Kavanagk, R. N., Darcey, T. M., Lehmann, D., and Fender, D. H. (1978). convolutional neural networks for EEG decoding and visualization. Hum. Brain
Evaluation of methods for three-dimensional localization of electrical Mapp. 38, 5391–5420. doi: 10.1002/hbm.23730
sources in the human brain. IEEE Trans. Biomed. Eng. 25, 421–429. Sclabassi, R. J., Sonmez, M., and Sun, M. (2001). EEG source localization: a neural
doi: 10.1109/TBME.1978.326339 network approach. Neurol. Res. 23, 457–464. doi: 10.1179/0161641011011
Kim, Y. (2014). Convolutional neural networks for sentence classification. arXiv 98848
preprint arXiv:1408.5882. doi: 10.3115/v1/D14-1181 Sharma, P., Scherg, M., Pinborg, L. H., Fabricius, M., Rubboli, G., Pedersen,
Kingma, D. P., and Ba, J. (2014). Adam: a method for stochastic optimization. arXiv B., et al. (2018). Ictal and interictal electric source imaging in pre-
preprint arXiv:1412.6980. surgical evaluation: a prospective study. Eur. J. Neurol. 25, 1154–1160.
Koles, Z. J. (1998). Trends in EEG source localization. Electroencephalogr. Clin. doi: 10.1111/ene.13676
Neurophysiol. 106, 127–137. doi: 10.1016/S0013-4694(97)00115-6 Simonyan, K., and Zisserman, A. (2014). Very deep convolutional networks for
Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). “Imagenet classification large-scale image recognition. arXiv preprint arXiv:1409.1556.
with deep convolutional neural networks,” in Advances in Neural Information Tadel, F., Baillet, S., Mosher, J. C., Pantazis, D., and Leahy, R. M. (2011).
Processing Systems (Lake Tahoe, NV), 1097–1105. Brainstorm: a user-friendly application for MEG/EEG analysis. Comput. Intell.
Lantz, G., Holub, M., Ryding, E., and Rosén, I. (1996). Simultaneous intracranial Neurosci. 2011:879716. doi: 10.1155/2011/879716
and extracranial recording of interictal epileptiform activity in patients with Tankelevich, R. (2019). Inverse problem’s solution using deep learning: An
drug resistant partial epilepsy: Patterns of conduction and results from EEG-based study of brain activity. Part 1. Available online at: https://www.
dipole reconstructions. Electroencephalogr. Clin. Neurophysiol. 99, 69–78. researchgate.net/publication/331153114_Inverse_Problem’s_Solution_Using_
doi: 10.1016/0921-884X(96)95686-6 Deep_Learning_An_EEG-based_Study_of_Brain_Activity_Part_1_-_rel_10
LeCun, Y., and Bengio, Y. (1995). Convolutional networks for images, speech, and Van Veen, B. D., van Drongelen, W., Yuchtman, M., and Suzuki, A.
time series. Handbook Brain Theory Neural Netw. 3361:1995. (1997). Localization of brain electrical activity via linearly constrained
Lin, C.-H., Kong, C., and Lucey, S. (2018). “Learning efficient point cloud minimum variance spatial filtering. IEEE Trans. Biomed. Eng. 44, 867–880.
generation for dense 3d object reconstruction,” in Thirty-Second AAAI doi: 10.1109/10.623056
Conference on Artificial Intelligence (New Orleans, LA). Willemse, R. B., Hillebrand, A., Ronner, H. E., Peter Vandertop, W., and
Liu, C., Zoph, B., Neumann, M., Shlens, J., Hua, W., Li, L.-J., et al. (2018). Stam, C. J. (2016). Magnetoencephalographic study of hand and foot
“Progressive neural architecture search,” in Proceedings of the European sensorimotor organization in 325 consecutive patients evaluated for tumor
Conference on Computer Vision (ECCV) (Munich), 19–34. or epilepsy surgery. Neuroimage Clin. 10, 46–53. doi: 10.1016/j.nicl.2015.
Lucas, A., Iliadis, M., Molina, R., and Katsaggelos, A. K. (2018). Using 11.002
deep neural networks for inverse problems in imaging: beyond analytical Wipf, D. and Nagarajan, S. (2009). A unified Bayesian framework
methods. IEEE Signal Proc. Mag. 35, 20–36. doi: 10.1109/MSP.2017.27 for MEG/EEG source imaging. Neuroimage 44, 947–966.
60358 doi: 10.1016/j.neuroimage.2008.02.059
Luck, S. J. (2014). An Introduction to the Event-Related Potential Technique. Wirgin, A. (2004). The inverse crime. arXiv:math-ph/0401050.
Cambridge: MIT press. Wolters, C. H., Lew, S., MacLeod, R. S., and Hämäläinen, M. (2010). Combined
Michel, C. M., and Brunet, D. (2019). EEG source imaging: a practical review of EEG/MEG source analysis using calibrated finite element head models. Biomed.
the analysis steps. Front. Neurol. 10:325. doi: 10.3389/fneur.2019.00325 Technik Biomed. Eng. Rostock Germany 55(Suppl. 1), 64–68.
Nair, V., and Hinton, G. E. (2010). “Rectified linear units improve restricted Yuasa, M., Zhang, Q., Nagashino, H., and Kinouchi, Y. (1998). “EEG source
boltzmann machines,” in Proceedings of the 27th International Conference on localization for two dipoles by neural networks,” in Proceedings of the
Machine Learning (ICML-10) (Madison, WI), 807–814. 20th Annual International Conference of the IEEE Engineering in Medicine
Nunez, P. L., and Srinivasan, R. (2006). Electric Fields of the Brain: The and Biology Society. Vol. 20 Biomedical Engineering Towards the Year
Neurophysics of EEG. Oxford: Oxford University Press, USA. 2000 and Beyond (Cat. No. 98CH36286), Vol. 4 (Hong Kong: IEEE), 21
Pascual-Marqui, R., Michel, C. M., and Lehmann, D. (1994). Low-resolution 90–2192.
electromagnetic tomography–a new method for localizing electrical activity Zhang, Q., Yuasa, M., Nagashino, H., and Kinouchi, Y. (1998). “Single dipole
in the brain. Int. J. Psychophysiol. 18, 49–65. doi: 10.1016/0167-8760(84) source localization from conventional EEG using BP neural networks,” in
90014-X Proceedings of the 20th Annual International Conference of the IEEE Engineering
Pascual-Marqui, R. D. (1999). Review of methods for solving the EEG inverse in Medicine and Biology Society. Vol. 20 Biomedical Engineering Towards the
problem. Int. J. Bioelectromagn. 1, 75–86. Year 2000 and Beyond (Cat. No. 98CH36286), Vol. 4 (IEEE), 2163–2166.
Pascual-Marqui, R. D. (2007). Discrete, 3D distributed, linear imaging methods of
electric neuronal activity. Part 1: Exact, zero error localization. arXiv preprint Conflict of Interest: The authors declare that the research was conducted in the
arXiv:0710.3341. absence of any commercial or financial relationships that could be construed as a
Razorenova, A., Yavich, N., Malovichko, M., Fedorov, M., Koshev, N., and Dylov, potential conflict of interest.
D. V. (2020). “Deep Learning for Non-Invasive Cortical Potential Imaging,”
in Machine Learning in Clinical Neuroimaging and Radiogenomics in Neuro- Copyright © 2021 Hecker, Rupprecht, Tebartz Van Elst and Kornmeier. This is an
Oncology (Lima: Springer), 45–55. open-access article distributed under the terms of the Creative Commons Attribution
Ribera, J., Guera, D., Chen, Y., and Delp, E. J. (2019). “Locating objects without License (CC BY). The use, distribution or reproduction in other forums is permitted,
bounding boxes,” in Proceedings of the IEEE Conference on Computer Vision provided the original author(s) and the copyright owner(s) are credited and that the
and Pattern Recognition (San Juan), 6479–6489. original publication in this journal is cited, in accordance with accepted academic
Ritter, P., and Villringer, A. (2006). Simultaneous EEG–fMRI. Neurosci. Biobehav. practice. No use, distribution or reproduction is permitted which does not comply
Rev. 30, 823–838. doi: 10.1016/j.neubiorev.2006.06.008 with these terms.

Frontiers in Neuroscience | www.frontiersin.org 14 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

A. APPENDIX

FIGURE A1 | Inverse solution of a simulation containing a single source cluster. An exemplary, representative sample of inverse solutions for a single source cluster.
(A) The ERP at each of the 31 channels containing both signal (central peak) and realistic noise from real recordings. (B) The scalp map at the central ERP peak (as
indicated by the vertical red line in A). (C) The dipole moments plotted on the white matter surface of the template brain in lateral view of the left hemisphere. On the
left, the ground truth source pattern is depicted with a source cluster in the frontal cortex of the left hemisphere. Various inverse solutions that aim to recover this
pattern are depicted next to it. Voxels below 25% of the respective maximum are omitted for a clearer representation of the current distribution.

Frontiers in Neuroscience | www.frontiersin.org 15 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

B. APPENDIX

FIGURE A2 | Inverse solution of a simulation containing four source clusters. An exemplary, representative sample of inverse solutions for four source clusters.
(A) The ERP at each of the 31 channels containing both signal (central peak) and realistic noise from real recordings. (B) The scalp map at the central ERP peak
(vertical red line in A). (C) The dipole moments plotted on the white matter surface of the template brain in lateral view of the left hemisphere. On the left, the ground
truth source pattern is depicted with a source cluster in the motor cortex, supplementary motor area, insula and the middle temporal lobe of the left hemisphere.
Various inverse solutions that aim to recover this pattern are depicted next to it. Voxels below 25% of the respective maximum are omitted for a clearer representation
of the current distribution.

Frontiers in Neuroscience | www.frontiersin.org 16 June 2021 | Volume 15 | Article 569918

Hecker et al. ConvDip: EEG Source Imaging

C. APPENDIX

FIGURE A3 | Inverse solution of real face-evoked data. All inverse algorithms were subjected to real EEG data recorded from a single participant containing 150 trials
of face-evoked potentials. (A) The ERP at each of the 31 channels. (B) The scalp map at a N170-like component (vertical red line in A, 211 ms after stimulus onset).
(C) The dipole moments plotted on the white matter surface of the template brain in ventral view of the right hemisphere to reveal the inferior temporal cortex (ITC).
Various inverse solutions show activity in the ITC. Notably, ConvDip and cMEM recover a focal source cluster close or within the fusiform face area, a region known to
be involved in face processing (e.g., Kanwisher and Yovel, 2006; Joos et al., 2020). Voxels below 25% of the respective maximum are omitted for a clearer
representation of the current distribution.

Frontiers in Neuroscience | www.frontiersin.org 17 June 2021 | Volume 15 | Article 569918