Section: New Results

MEG, EEG and cochlear modeling

MEM-diffusion MRI framework to solve MEEG inverse problem

Participants : Brahim Belaoucha, Jean-Marc Lina, Maureen Clerc, Théodore Papadopoulo.

In this work, we present a framework to fuse information coming from diffusion magnetic resonance imaging (dMRI) with Magnetoencephalography (MEG)/ Electroencephalography (EEG) measurements to reconstruct the activation on the cortical surface. The MEG/EEG inverse-problem is solved by the Maximum Entropy on the Mean (MEM) principle and by assuming that the sources inside each cortical region follow Normal distribution. These regions are obtained using dMRI and assumed to be functionally independent. The source reconstruction framework presented in this work is tested using synthetic and real data. The activated regions for the real data is consistent with the literature about the face recognition and processing network.

This work was published in the proceedings of the conference EUSIPCO 2015 [22] .

MEG/EEG reconstruction in the reduced source space

Participants : Brahim Belaoucha, Théodore Papadopoulo.

Obtaining the brain activity with the distributed source model from MEG or EEG measurements is ill-posed problem due to the high number of unknowns compared to the number of measurements. The idea of ths work is to reduce the solution space size from the number of sources to a smaller space. Assuming that sources inside each functional region have equal activation allows us to reduce the number of columns in the leadfield matrix from the number of nodes $S$ required to model the cortex to a number of regions $K$, which is much smaller. These regions are obtained from a dMRI parcellation-based region growing algorithm. A region is assumed to contain sources that have similar fibers distribution. To obtain a sparse solution, we assume that only a few regions are active simultaneously. BIC1 is used to obtain the optimal number of regions (Kp) that explains the MEG/EEG data.

We compared the results of the proposed method to the ones from Minimum Norm Estimate (MNE) and LASSO. The first gives a smooth solution and the second gives a sparse solution. To test the accuracy of the reconstruction, we activated simultaneously from two to five regions in both hemispheres with synthetic low SNR signals (10 dB). Our approach could detect the right number of activated regions and provided more accurate reconstructions compared to MNE and LASSO.

Our approach assumes that few regions are active simultaneously which allows us to reduce the space to a few unkowns. It can be seen as an approximation to the l0 norm. Even though assuming a constant activation in each functional region is a hard constraint, it allows us to reduce the space size from $S$ to $K$. The obtained solution can be used to detect extended sources (e.g epileptic activity) or as an initialization step to other approaches to obtain more detailed solutions in the active regions.

This work was presented at the conference BaCI 2015 [24] .

Realistic simulation of electric potential distributions of different stimulation modes in an implanted cochlea

Participants : Kai Dang, Maureen Clerc, Clair Vandersteen [Institut Universitaire de la Face et du Cou, Nice] , Nicolas Guevara [Institut Universitaire de la Face et du Cou, Nice] , Dan Gnansia [Oticon Medical/Neurelec] .

Simulation of the intracochlear potentials is an important approach to study the activation of auditory nerve fibers under electrical stimulations. However, it is still unclear to which extent the simulation results are affected by precision in reproducing the exact cochlear geometry. In this study, we address to this question by comparing the actual electric potential measured from implanted human specimen with the simulation outputs from two different parametric 3D cochlear models. One of the models is created from the default values while the other is adapted to the micro-CT scan data of the implanted cochlea.

This work was presented at the Association of Research in Otolaryngology 38th MidWinter Meeting, Feb 2015, Baltimore, United States [38] .

We also made an in situ validation of electrical models: Cochlear implants have been proved to be an effective treatment for patients with sensorineural hearing loss. Among all the approaches that have been developed to design better cochlear implants, 3D model-based simulation stands out due to its detailed description of the electric field which helps reveal the electrophysiological phenomena inside the cochlea. With the advances in the cochlear implant manufacturing technology, the requirement on simulation accuracy increases. Improving the simulation accuracy relies on two aspects: 1) a better geometrical description of the cochlea that is able to distinguish the subtle differences across patients; 2) a comprehensive and reliable validation of the created 3D model. In this paper, targeting at high precision simulation, we propose a parametric cochlea model which uses micro-CT images to adapt to different cochlea geometries, then demonstrate its validation process with multi-channel stimulation data measured from a implanted cochlea. Comparisons between the simulation and validation data show a good match under a variety of stimulation configurations. The results suggest that the electric field distribution is affected by the geometric characteristics of each individual cochlea. These differences can be correctly reflected by simulations based on a 3D model tuned with personalized data.

This work was presented at the 7th International IEEE EMBS Conference on Neural Engineering, Apr 2015, Montpellier, France [25] .

Influence of skull modelling on conductivity estimation for EEG source analysis

Participants : Christos Papageorgakis, Maureen Clerc, Benjamin Lanfer [BESA GmbH] .

The skull conductivity strongly influences the accuracy of EEG source localization methods. As the conductivity of the skull has strong inter-individual variability, conductivity estimation techniques are required. Typically, conductivity estimation is performed on data from a single event-related stimulation paradigm, which can be explained by one dipole source. A conductivity value for the skull can be estimated as the value for which the single dipole source provides the best goodness of fit to the data. This conductivity value is then used to analyse the actual data of interest. It is known that the optimal local skull conductivity when modelling the skull as one compartment depends on the amount of spongiosa present locally. The research question arising is: Is conductivity estimation based on data from a single paradigm meaningful without accounting for the internal skull structure ?

This work was presented at the conference BaCI 2015 [33] , and is submitted for journal publication.

Dictionary learning for M/EEG multidimensional data

Participants : Christos Papageorgakis, Sebastian Hitziger, Théodore Papadopoulo.

Signals obtained from magneto- or electroencephalography (M/EEG) are very noisy and inherently multi-dimensional, i.e. provide a vector of measurements at each single time instant. To cope with noise, researchers traditionally acquire measurements over multiple repetitions (trials) and average them to classify various patterns of activity. This is not optimal because of trial-to-trial variability (waveform variation, jitters). The jitter-adaptive dictionary learning method (JADL) has been developed to better handle for this variability (with a particular emphasis on jitters). JADL is a data-driven method that learns a dictionary (prototype pieces) from a set of signals, but is currently limited to a single channel, which restricts its capacity to work with very noisy data such as M/EEG. We propose an extension to the jitter-adaptive dictionary learning method, that is able to handle multidimensional measurements such as M/EEG.

This work was presented at the conference BaCI 2015 [32] .