Objectives: Although many studies in epilepsy have examined the synaptic mechanisms constituting the basis for most of the current principles of brain activity, relatively less studies have tried to characterize changes in the cellular environment that might predispose a network to pathologic synchronization.

Methods: In this study, near-infrared optical imaging was used with ECoG and EEG to investigate variations in the optical properties of cortical tissue directly associated with neuronal activity in 15 rats and 3 human epileptic patients. Time-frequency analysis was also used to track variations of (de)synchronization concomitantly with changes in optical signals during IES.

Results: Changes in Fast optical signals (FOS) occurred 320 msec before to 370 msec after the IES peak. These changes started before any changes in ECoG signal. In addition, time-frequency domain ECoG revealed an alternating decrease-increase-decrease in the ECoG spectral power (pointing to desynchronization-synchronization-desynchronization), which occurred concomitantly with an increase-decrease-increase in relative optical signal (pointing to shrinking-swelling-shrinking of the neuronal assembly) during the IES.

Discussion: These relationships between electrical and optical changes highlights the complexity of the interplay between the neuronal network activity and its environment around an IES.

Conclusions: These changes in the neuronal environment around IESs raise new questions about the mechanisms that provides the suitable conditions for the neuronal synchronization during IESs.

Significance: The multimodal-multiscale FOS-ECoG approach opens new avenues to better analyze the mechanisms of neuronal synchronization in the pathologic epileptic brain, which is applicable in clinic.

The interpretation of EEG is improved using source reconstruction. Accurate reconstructions necessitate anatomical models and electrode coregistration. The golden standard consists of anatomical MRIs and an electromagnetic digitizer for the electrodes. These are costly, require considerable time, and are not always feasible. In this study we investigated an 800 euro optical 3-D scanner as an alternative.

We scanned 49 subjects with MRI, Polhemus digitizer and a 3-D scanner. We created volume conduction models from this and from template models. We used these to compute the EEG scalp distribution for sources distributed over the cortical sheet. We compared these to evaluate the difference between golden standard individual models, individualized template models and a common template model.

With on average 2 minutes lab-time, the 3D scan procedure is considerably faster. The quality of the electrode model is significantly better than a common electrode model, although the magnetic digitizer remains more accurate. The quality of the head model is not significantly different than a common template model. The quality of individual MRI-based head models is not reached.

Our model comparisons show a strong improvement if individual electrode positions are considered. Optical 3D scanners are a cost and time efficient alternative for recording these. The difference between individualized template head models and a common template is not significant. The golden standard is not to be replaced where applicable, but optical 3-D scanner based electrode models are better than template electrodes. Hence we recommend using an optical 3-D scanner to improve source reconstruction.

High-density-EEG is a powerful tool to estimate brain connectivity, but the choice of parameters, i.e., data pre-processing steps, model order p of the multivariate-autoregressive (MVAR) model, number of samples and level of signal-to-noise-ratio (SNR), in the connectivity pipeline is fundamental to avoid spurious connectivity. We provide a simulation framework to test the influence of the choice of the model order p in the estimation of connectivity based on information partial directed coherence (iPDC). The time-course of four sources J(t) was reconstructed imposing the connectivity structure by manipulating the autoregressive matrices A_p of the following equation: J(t)=∑_(p=1:5) A_p J(t-p)+w(t) where w(t) is zero-mean gaussian noise and SNR is equal to 5. Then, iPDC was estimated from 100 data sets varying p in the range [2 20]. We also simulated 100 data sets of gaussian noise to compute iPDC thresholds. We evaluated the performance in connectivity estimation by computing sensitivity, specificity and accuracy from iPDC results. The true connections were detected in a narrow range order of p=[4 9] obtaining an average accuracy in the interval [0.995 0.999]. False positive connections appeared both for orders minor to 4 and major to 9 with a quickly deterioration of accuracy results [up to 20%]. In this work, we proved how much the choice of p order matters in the connectivity estimation. Future developments of this work will be simulations involving interacting sources at different frequency range to test both the different connectivity methods and the effects of the other parameters in the connectivity pipeline.

Model individualisation is a key factor to increase the accuracy in the EEG forward problem (EEG-FP), and consequently in source localisation. To this end, there exist methods accepted in the community for acquiring the necessary data to make these computational models as personalised as possible. However, there is one factor that is generally neglected in the model individualisation process: the electrical conductivity. Although few methodologies exist to deal with this issue, they usually rely on iterative solutions in which the complete lead-field matrix is computed (i.e. refined) in each step. This makes such methods computationally expensive, for which the convergence could take several hours or days even in simplistic scenarios.

In this work, we present a solution to this problem by applying a reduced order methodology to the dual version of the EEG-FP. This technique exploits the affine dependence of the stiffness matrix and load vector with respect to the electrical conductivities to speed-up the calculation of lead-field matrices. This is done by computing a set of suitable, problem-dependent basis functions to express any solution of the EEG-FP, considering arbitrary electrical conductivities, up to a certain error. The EEG-FP is then solved in the reduced space. Using a five-layered model, we found that approximations with a relative error of 10^-5 with respect to the high-fidelity solutions are obtained considering less than 20 basis functions, allowing to compute accurate lead-field matrices in less than a second. The convergence of the method and other theoretical and practical aspects are also discussed.

The characterisation of electrical sources of brain activity by means of EEG is fundamental for understanding brain processes. The accuracy with which we perform such analysis is limited by the models used to represent the sources of electrical activity (among other factors). In this regard, the dipolar model is generally adopted. Although useful, it was shown to be limited to represent sources with non-negligible spatial extent. To increase the reliability, Jerbi et al. (2004) proposed to use multipolar source components, which they show to increase the accuracy of the source estimation process using MEG recordings. Even though this framework showed great improvements with respect to the standard dipolar models, it was presented for MEG only and considering spherical head models. This limits the applicability of the technique to individualised models, for which numerical methodologies need development.

In this work, we present a full subtraction version of the finite element method for solving the EEG forward problem considering multipolar source models. This framework allows to perform computational simulations of electrical brain activity utilising multipolar sources in anisotropic and personalised head models. In particular, we analysed the cases of dipolar and quadrupolar source components. Numerical solutions are compared with analytical formulas in a multi-layered spherical model with anisotropic electrical conductivity. These formulas are available in the case of dipolar sources, and here derived for quadrupolar components. Results in idealised and realistic head models show the reliability of the method for further multipolar characterisation of electrical brain sources.

We present a subtraction approach for solving the EEG forward problem (EEG-FP) considering the complete electrode model (CEM) and multipolar source models. Differently from other approaches, we deal with the singularity in the source term by splitting the electric potential into a singularity potential, for which analytical expressions are available, and a singularity-free correction potential that is approximated using the finite element method (FEM). This approach allows the use of standard finite elements for solving the EEG-FP in personalised head models with anisotropic electrical conductivity field. Moreover, the subtraction method is used to show the existence and uniqueness of the solution, extending the results previously obtained for dipolar sources and the point electrode model (Wolters et al., 2007). The methodology here presented consists in an alternative to the approach proposed by Pursiainen et al. (2017) based on Whitney basis functions, for which the simulation of sources in arbitrary locations would require extra efforts. Numerical experiments are shown considering the full and projected FEM versions.

Several methods have been proposed for the inversion of the magnetoencephalography (MEG) and the Electroencephalography (EEG) problems, i.e. the localization of the active brain regions from the measured M-EEG signals, and all of them require two main ingredients: the forward M-EEG model and the source space. The forward model relates the electric potential/magnetic field at sensors' positions produced by a known neuroelectric current distribution while the source space reflects our a priori knowledge on the current flowing inside the brain. An accurate source space, approximating the cortical surface, consists of thousand of points and this large number of possible sources is responsible for the high computational cost for the computation of the forward model first and then of the solution of the inverse problem.

Here, we propose the RAndoM Sampling invErSion method (RAMSES) which uses a sampling procedure to significantly reduce the dimension of the source space. The accuracy of the method in localizing brain activity is investigated using both synthetic and real MEG data. We employ three different model for the forward solution -a BEM model, a spherical model and a less sophisticated method based on the Biot-Savart operator- and compare the results of RAMSES when employing different methods to solve the inverse problem -a simple least square (LSQR) Matlab routine, dynamic statistical parametric map (dSPM), weighted Minimum Norm Estimates (wMNE).

The tests show that the random sampling procedure does not compromise the capability of localizing brain activity.

Introduction

Accurate segmentation of an individual human head in conjunction with the resulting volume conductor model has been shown to improve the results of EEG source analysis. However, comparison of some state-of-the-art segmentation techniques and their effectiveness in source analysis, namely Multi-Atlas and Convolutional Neural Networks (CNN) segmentation, is lacking. We present a comparison between these techniques to segment five tissues using ground-truth data from BrainWeb.

Method

For the Multi-Atlas method, a 2-step registration process with affine and non-rigid registration along with a classification step involving label-fusion from registered atlases was performed, using weighted local similarity. For CNN, a 3D multi-path pipeline with an 8-layer architecture, a kernel of size 3x3x3, and receptive-field of size 25x25x25, was designed. An extra pathway for sub-sampled images was used to exploit spatial cues. We calculated the lead-field matrix for each segmentation result using 2000 sources and 97 scalp electrodes.

Results

Dice scores for cerebral spinal fluid (CSF), grey matter (GM), white matter (WM), muscle/skin and skull were 0.85, 0.94, 0.94, 0.96, 0.91 and 0.78, 0.82, 0.83, 0.94, 0.88 for CNN and multi-atlas respectively. Mean values of the relative and magnitude difference measure for computing lead-field matrices were 0.1071, 0.9808 and 0.2089, 0.9544 for CNN and multi-atlas respectively.

Conclusion

CNN outperforms multi-atlas method in segmenting CSF, GM, and WM, primarily due to the tissue’s structural variability across subjects which affects consensus-based algorithms. The CNN lead-field matrix values were closest to those of the ground-truth; the effect on source analysis will be investigated further.

Localizing the seizure onset zone (SOZ) from ictal EEGs is a difficult process since different physiological and non-physiological noises occurring during the seizure. Therefore, ictal EEG source imaging (ESI) is currently not used in the presurgical evaluation for epilepsy patients. To localize the SOZ, we propose a methodology coupling decomposition with ESI.

We decomposed 68 artifact-free ictal epochs recorded in 18 patients having Engel Class I surgical outcome. The decomposition techniques were independent component analysis (ICA), canonical polyadic decomposition (CPD) and block term decomposition (BTD). Based on the clinical reports of ictal discharges, ictal component was manually identified. Then, ictal source was generated from the maximum of sLORETA reconstruction. To investigate the decomposition effects on localization, ESI was also applied on the aforementioned epochs without decomposition. For both approaches, the distance between the estimated SOZ and the border of resected zone (RZ) was calculated.

By considering the estimated SOZ inside the RZ (or within 20mm from the RZ) ESI alone was correct in 13%(34%) of the seizures, while it increased to 18%(43%), 22%(38%) and 26%(46%) by combining ESI with BTD, CPD, and ICA, respectively. Without decomposition, 39% of patients had more than 2/3 of their seizures localized within 20mm from the RZ. By including decomposition, this increased to 56%, 61% and 78% for BTD,CPD and ICA, respectively.

We showed that decomposing the ictal EEG before applying ESI is beneficial to localize the SOZ. The technique is promising, but currently the accuracy still needs to be improved for clinical application.

Our objective was to determine if there are any distinguishable phase clustering patterns present before, during and after the onset of epileptic spikes. The phase clustering activity was derived from 256-channel high density (dEEG) data of an adult patient who had epileptic activity in the left central and parietal areas as determined from invasive subdural recordings. The analysis was performed in the ripple band ( 80-150 Hz) and in the low gamma band (30-50 Hz). The dEEG data was filtered in the appropriate band. Hilbert transform was applied to compute the analytic phase and it was unwrapped. Spatiotemporal contour plots of the unwrapped analytic phase with 1.0 ms intervals were constructed using a montage layout of 256 electrode positions. These plots exhibited dynamic formation of phase cones which are similar to bubbles in boiling water. Several criterions were applied to select stable phase cone patterns. These included: phase frequency was within the temporal band, sign of spatial gradient did not change for at least 3 time samples and the frame velocity should be within the range of propagation velocities of cortical axons. Analysis was performed during ±5 seconds from the location of spike with a resolution of one sec. Stable phase cluster patterns were higher in the seizure area as compared with the nearby surrounding brain areas. Spatiotemporal oscillatory patterns were also visible during ±5 sec period. These results show the feasibility to localize epileptic spikes and also to study the dynamics of cortical neurons.

Being given pointwise measurements of the electric potential taken by electrodes on part of the scalp, the EEG (electroencephalography) inverse problem consists in estimating current sources within the brain that account for this activity.

A model for the behaviour of the potential rests on Maxwell equation in the quasi-static case, under the form of a Poisson-Laplace equation.

We will describe our approach for solving the inverse problem in spherical geometry, for piecewise constant electric conductivity values, and pointwise dipolar source terms.

It relies on consecutive steps, consisting of (see [CLMP]):

(i) singular value decomposition, in order to separate the time independant activities;

(ii) spherical harmonics expansion, for data transmission from scalp to cortex ("cortical mapping") using best constrained approximation;

(iii) best rational approximation on 2D slices in order to compute singularities in circular sections;

(iv) clustering of these singularities in order to localize the sources, dipole fitting, moment computation.

The algorithm has been encoded in the software FindSources3D (see http://www-sop.inria.fr/apics/FindSources3D/). Numerical simulations will be presented.

[CLMP] M. Clerc, J. Leblond, J.-P. Marmorat, T. Papadopoulo, Source localization in EEG using rational approximation on plane sections, Inverse Problems, 28, 055018, 2012.

Magnetoencephalography (MEG) and Electroencephalography (EEG) are widely used systems to study the electrophysiological dynamics of the human brain. MEG/EEG are able to provide data streams with millisecond-temporal resolution. This makes them ideal candidates for real-time monitoring and processing of neuronal activity. In conjunction with neurofeedback sce-narios, real-time MEG/EEG data processing allows the adaption of the experiment to the subject’s reaction, creating a whole set of new options and possible experiments. By further advancing the open source MNE-CPP project we want to provide a state of the art framework, which offers tools to develop novel real-time processing methods and to build standalone applications for electrophysiological data processing.

The MNE-CPP project is structured into libraries, which guarantee a modular and easily extendable architecture. MNE-CPP hosts libraries to support the Fiff and FreeSurfer data format as well as source estimation and 2D/3D displaying routines. We have kept the external dependencies to a minimum, namely Qt5 and Eigen. We were able to build several MNE-CPP based soft-ware applications for real-time (MNE Scan) as well as offline (MNE Browse) data processing. Next to usage in research envi-ronments, MNE-CPP applications can also function in clinical environments with regulatory requirements (BabyMEG). We recently added new EEG device support (BrainAmp, EGI) to MNE Scan and further improved overall 3D visualization, i.e. by including real-time smoothing of cortical activity. Furthermore, we added tools to track head motion in MEG scenarios via HPI coils. The new tracking tools provide 3D visualization of the subject’s head relative to the sensors in real-time.

The surgical treatment of patients with complex epilepsies is changing from open, invasive surgery towards minimally invasive, image guided treatment. Brain imaging is becoming more and more important for preoperatively identifying the region of the brain which is responsible for the epilepsy of the patient. The ultimate aim is to provide the neurosurgeon with a clear, intuitive image of the targeted epileptogenic region, to enable a resection which renders the patient seizure free while avoiding damage to the cortex.

A software product is developed for the visualization of multi-modal brain images and analysis results of non-invasive and invasive epilepsy recordings. The software is designed for three main tasks. At the preparation step the data is collected, pre-processed and saved together with the patient info in the application database. During the exploration step, different aspects of the data can be investigated and at the final step of visualization, individual images can be combined in multi-modal 2D- and 3D-MRI viewports. The software contains several pre-programmed sequences for creating multi-modal visualizations used to identify epileptic tissue versus functional areas, like visualizations of inverse solutions of high density EEG and MEG, EEG informed functional MRI visualizations and functional Near-infrared Spectroscopy projections. The end result is a software tool that supports the decision process involving the preoperative planning of surgical resections of epileptic tissue, which is less time consuming and yields a more precise delineation of epileptic tissue with a higher success rate in case of surgery.

OpenMEEG implements boundary element solutions for simulating electromagnetic fields in the quasistatic regime. Originally designed for the forward EEG and MEG problems (MEEG collectively), it has also been applied to compute forward solutions for ECoG, for implanted EEG, for cochlear implant stimulation, for tDCS and other electrostimulation settings. In this presentation we detail the features of the latest release of OpenMEEG.

Geometrical models have now been extended to handle non-nested geometries: the various domains must still have a constant isotropic conductivity but they need no longer be nested inside one another. OpenMEEG supports CGAL meshing tools (allowing to remesh or decimate existing meshes, or to mesh a levelset). Linear algebra packages MKL and OpenBLAS are now supported on all platforms (Linux, MacOS and Windows). Interface with Python is improved with wrappers that allow to pass data to Python without memory copies. Gifti and VTK mesh formats are supported, and some visualization tools are provided (VTK or mayavi). More tests have been included, and last but not least, some bugs fixed.

The effect of the choice of the inverse method on the cortico-cortical coupling analysis has been largely overlooked in the literature. Here, we set out to investigate the impact of three inverse methods on source coherence detection using simulated MEG data. To this end, we created thousands of randomly located pairs of sources and varied their inter- and intra-source correlation strength, source size and spatial configuration. Then, we used the simulated pairs of sources to generate sensor-level MEG measurements at varying signal-to-noise ratios (SNR). Next, we reconstructed the sources using L2-Minimum-Norm Estimate (MNE), Linearly Constrained Minimum Variance (LCMV) beamforming, and Dynamic Imaging of Coherent Sources (DICS) beamforming; and calculated source level power and coherence maps. We evaluated the performance of the methods using the Receiver Operating Characteristic (ROC) curves. The results indicate that beamformers perform better than MNE for coherence reconstructions of interacting point-like sources; but MNE provides better connectivity estimation than beamformers of interacting extended cortical patches, if each patch consists of dipoles with identical time series (high intra-patch coherence). However, the performance of the beamformers for interacting patches improves substantially if each cortical patch is simulated with partly coherent time series (partial intra-patch coherence). These results demonstrate that the choice of the inverse method impacts the results of MEG source-space coherence analysis, and that the optimal choice of the inverse solution depends on the spatial and synchronization profile of the interacting cortical sources. Our conclusions can guide method selection and help improve data interpretation regarding MEG connectivity estimation.

The bioelectromagnetic inverse problem cannot be solved based on EEG/MEG data alone and requires additional assumptions. In linear reconstruction methods, spatial smoothness is often used as an additional constraint. This is equivalent to the prior assumption of a particular source covariance structure. Recent publications (Knösche et al., NeuroImage 2013) have suggested altering this spatial correlation structure such that it reflects available knowledge on the functio-anatomical organization of the brain. In particular, it is possible to derive borders between different brain areas from various types of brain images. This allows assuming that sources located within the same area exhibit similar activity and sources in different areas are mutually uncorrelated. We present a technique based on the well-known LORETA method (Pascual-Marqui et al., Int. J. Psychophysiol. 1994), which is capable of incorporating such function-anatomical priors. We show that our method embodies the intended prior knowledge in the prior source covariance in an unbiased way. We present Monte-Carlo simulations, which provide a systematic evaluation of how the prior knowledge influences the estimate of different linear inverse procedures. The study answers questions like “What happens if the course of boundaries is uncertain?”, “What if our knowledge on functional areas is limited to certain cortical regions?” and “Can prior knowledge improve source localization?”. Besides presenting answers to these questions we demonstrate our method to localize auditory N100 activity from experimental EEG/MEG data. The results clearly suggest that spatially informed linear inverse methods provide very plausible reconstruction results.

Redundancy in high-resolution electroencephalography (EEG) may cause numerical instability and inaccuracies in source reconstruction. Dimensionality reduction via spatial projection, which is based on singular value decomposition (SVD), suppresses this redundancy while largely preserving the benefits of improved head coverage and higher spatial resolution for surface-EEG. The authors have successfully used spatial projection in conjunction with the spatiotemporal Kalman filter (STKF) to alleviate this problem. The choice of the optimal threshold value for spatial projection, however, has not yet been investigated. This proof-of-principle work uses different threshold values for spatial projection and studies the effect thereof on the accuracy and spatial resolution of source reconstruction via STKF and its generalized variant, the regional spatiaotemporal Kalman filter (RSTKF). RSTKF allows for region-specific dynamics in the state-space model to approximate the brain’s modularity. In this work we use 256-electrode EEG recordings from a patient with bilateral temporal lobe epilepsy caused by hippocampal sclerosis. The patient was operated in the right temporal lobe and is now seizure-free. The reconstructed source will be compared to the resected volume from the post-operative magnetic resonance image (MRI). First results show a reduction in spatial blurring for the source in the temporal lobe with decreasing threshold values for STKF until the point when redundancy dominates. Compared to the STKF, we expect the RSTKF to be more robust to redundancy and produce results with a higher accuracy and a better spatial resolution for the same threshold value, since its dynamical model is more sophisticated than that of the STKF.

Deep brain stimulation (DBS) is considered by some authors as the second most important therapeutic advance in Parkinson's disease (PD) after the introduction of L-dopa and dopamine agonists. We acquired scalp 256-channel EEG data from 10 PD patients with DBS of subthalamic nucleus (STN) during DBS ON and OFF state while performing 3-stimulus visual oddball task.

The preprocessing steps include DBS artefact filtering, bandpass filtering 1-40Hz, ICA for cardiac and eye-blinking artefact suppression, interpolation of bad channels and manual detection of bad segments.

The usage of event related potentials (ERP) analysis as an exploratory technique can be demanding for such high-density EEG. Thus, we decided to apply Parallel Factor Analysis (PARAFAC) on 3-way data array composed by averaged trials from all patients, both states (DBS ON/OFF) and all stimulus types (frequent, target, distractor). The resulting estimated PARAFAC components have 3 signatures - topography, time series and subject/state/stimulus type loadings for particular averaged trials. Finally, we compared loadings between trial types during both states by Wilcoxon test.

PARAFAC decomposition revealed evoked activity which showed significant difference between loadings of frequent and target stimuli in the DBS ON state and no difference in DBS OFF state. Thereafter we transformed the topography of the component into the source space, which points to areas of the fronto-parietal attention network.

We conclude that our results support a hypothesis that the DBS improves not only motor control but also affects cognitive networks.

Sleep induce changes in human brain connectivity/excitability [1]. According to [2], these modifications can be attributed to changes in the dynamics of neuronal responsiveness. In the epileptic brain, these activities and networks are also affected by these changes of state between sleep and wakefulness. The aim of this work is to estimate the sleep-induced changes in connectivity maps based on Cortico Cortical Evoqued Potentials (CCEP) for structures close to the epileptogenic zone in temporal lobe epilepsy. Intra-cerebral electrical stimulations are used to produce the CCEP, the resulting electrophysiological responses and connectivities are analyzed using SEEG recordings. Seven drug resistant epileptic patients were stimulated during 30 seconds in different sites during both sleep/wake states. CCEP are immediate causal response and can then be contaminated by the stimulation artefact. The first pre-processing step consists in removing this artefact while preserving the underlying CCEP. Two methods are evaluated, based on subspace decompositions: the Generalised Eigen Value Decomposition (GEVD) and the Common Spatial Subspace Decomposition (CSSD). The best separation results between the CCEP and the stimulation artefact is achieved using CSSD. Temporal connectivity Graph based on parametric model (DTF, PDC...) are then estimated in the sensor as well as in the reconstructed source space, for both sleep and wake states. The identified networks are validated by experts.

[1] Pigorini et. al., Bistability breaks-off deterministic responses to intracortical stimulation during non-REM sleep, 2015.

[2] Massimini et al., Cortical mechanisms of loss of consciousness: insight from TMS/EEG studies. Archives italiennes de biologie, 2012.

Introduction: Electrocorticography (ECoG) is an invasive technique commonly used in patients or animals. ECoG measures the electrical potential using strips of electrodes placed directly onto the cortex. The high SNR allows to infer about brain activity with millimetric spatial resolution under the strips. This resolution can be improved using source estimation techniques (Cho, 2011) though it strongly decreases away from the electrodes (Zhang, 2008). Our goal is to characterize the resolution properties of eLORETA and MNE at increasing distance from the electrodes. This is particularly interesting in studies where the strip doesn’t completely cover the areas of interest.

Methods: A realistic setup from a monkey ECoG study with 128 channels (Nagasaka, 2011) was considered. The source-to-sensor mapping was implemented using a FEM approach (Simbio, 2014) for grids with different electrode number. We characterized the resolution properties with metrics quantifying the localization error, the activity spread and the relative sensitivity of source estimates (Hauk, 2011) for Point-Spread-Functions (PSF) and Cross-Talk-Functions (CTF) in the whole source space.

Results: For a single active source (PSF), as its distance from the electrodes increases the eLORETA spread increases while the localization error is always zero, whereas for MNE all metrics increase. When all sources are simultaneously active (CTF), for both inverse algorithms the resolution metrics depend on the distance from the grid, slightly less for MNE.

Discussion: Source estimate from ECoG is reliable only near the electrodes and must be carefully interpreted accordingly to the resolution properties of the inverse algorithm.

Dynamic Granger-causality methods aim to quantify directed interaction strengths between brain areas with high temporal resolution, using simultaneously recorded electrophysiological signals. These methods are often based on time-varying multivariate autoregressive (tvMVAR) modeling, and while several such approaches have been proposed there currently is a lack of unbiased analyses and comparisons of their performance. Our aim was to compare the performance of commonly used tvMVAR methods using numerical simulations and real benchmark EEG data along fixed criteria. We compared classical Kalman filter (MVAAR), Dual Extended Kalman Filter (DEKF), Recursive Least Squares (RLS) and General Linear Kalman Filter (GLKF), and two ways of exploiting repeated observations: 1) single-trial modeling followed by averaging, and 2) multi-trial modeling where one tvMVAR model is fitted across trials. Our results show that while most approaches can adequately model simulated and real data, the best performance was often achieved with GLKF and a multi-trial approach. This approach’s accuracy, however, more strongly depended on model order choice and sampling rate. In fact, this algorithm produced highly variable estimates at high sampling rate and when large model order was required. In this scenario downsampling successfully reduced the estimates’ variability. Single-trial approaches using GLKF and MVAAR were more robust against setting model order too high and showed good performance at high sampling rates. For these algorithms downsampling degraded performance, because of slower adaptation speeds. Our findings help understand the strengths of various tvMVAR approaches and provide practical recommendations for their use in modeling dynamic directed interactions from electrophysiological signals.

Simulated MRI data have great value for segmentation algorithm development since they provide a large set of priors. However, such simulated MRI are not readily available for children’s heads. This along with the prevalence of adult simulated data frequently used as priors (BrainWeb), was our motivation to build an MR simulator for children.

The simulator works similarly to the adult version but with a child head phantom. Although some ground-truth datasets for children’s brains are available, data covering all tissue types of children’s heads for use as a phantom is not readily available.

Our head phantom was created in two steps. In the first step, the skull and scalp from an MRI-CT pair obtained from the Retrospective Image Registration Evaluation (RIRE) database was extracted. Skull extraction used thresholding applied to the CT image at 700 Hounsfield units (HU), followed by a morphological step to remove any anomalies.

In the second step, an MRI ground-truth pair was obtained from the UNC Infant Atlas. Non-rigid registration between MRIs from UNC and RIRE databases was performed. A phantom was created by fusing the skull data with the ground-truth of the brain, with visual validation. MR simulation of this phantom was performed using a hybrid of Bloch equation and tissue template simulation, enabling simulation of image contrast, partial volume, and noise.

The ground-truth and MRI obtained from the simulation can be used as priors for segmentation algorithms of complete children’s heads, with the aim of creating realistic head models for EEG/MEG source analysis.

Intro

Cortico-cortical Evoked Potentials (CCEPs), through the use of comparable bi-directional stimulations in pairs of ROIs, allow studying the directionality and reciprocity of functional connections. This paradigm is compatible with the concept of effective connectivity which evaluates the influence of one neural population onto another.

Brain activity changes according to the vigilance state (wakefulness/non-REM sleep). In intracerebral and scalp EEG recordings, it was demonstrated that irrative zones are more active (spikes) and large during sleep than wakefulness (R.Rocamora et al. 2015). In this study, we aim to investigate the influence of vigilance state on cortical reactivity.

Material and methods

We included one drug-resistant epileptic patient from a cohort of 5 patients.

Intracerebral electrical stimulations were performed using biphasic pulses (1050µs, 1mA,1Hz) in 120 intracerebral contact pairs during non-REM sleep and wakefulness conditions.

5 anatomical ROI were selected: amygdala, anterior and posterior hippocampus, entorhinal cortex and temporal pole. To characterize and compare the cortical reactivity during non-REM sleep and wakefulness, we compared the occurences, the averaged CCEPs amplitudes and latencies.

All CCEPs were analyzed using EEGLab and homemade algorithms. Data analysis comprised: 1.Filtering, 2.Stimulation peak detection, 3.Time boxes creation around the peak, 4.Artefact rejection, 5.Time boxes averaging around the peak and 6.CCEPs detection using a permutation and t-test.

Preliminary results

We observe a generally larger amplitude and spatial distribution of the CCEPs during non-REM sleep than wakefulness. These results are in accordance with the irrative zone behavior observed in epilepsy. No significant delay was observed between sleep and wakefulness latencies.

A previously introduced method for source connectivity analysis allows the projection of multivariate autoregressive (MVAR) model coefficients from MEG sensor space to source space using a lead-field matrix and a lead-field based inverse operator (weight matrix). However, the method shows deficits when source positions are not a priori known. This could be due to the fact that the product of the weight and lead-field matrices is not an identity matrix and to the crosstalk between sources. We improved the method to mitigate these drawbacks and examined the improved method using simulations and a real MEG dataset. For the estimation of MVAR model coefficients in source space, we used an inverse of the weight matrix instead of the lead-field matrix to reduce errors caused by the assumption that multiplying weight and lead-field matrices results in an identity matrix, and we applied a nulling beamformer for crosstalk suppression between sources. The partial directed coherence (PDC) was used as a connectivity measure calculated from the estimated MVAR model coefficients. In simulations, applying the inverse of the weight matrix reduced the errors in in/out-degree of the PDC, and spurious connections were reduced by using the nulling beamformer. In a case study, we applied our method to the interictal MEG recordings and could identify information flows from left to right regions nearby the focal cortical dysplasias found in MRI. These results suggest that the proposed method has considerable potential as a non­invasive approach for source connectivity analysis without a priori knowledge about source locations.

Introduction: Measurements on animals are important in clinical practice. The aim of our study was to develop a software module for statistical brain-mapping. This study compares brain activity during application of psilocin.

Methods: In this study we measured electrical activity of 9 rat’s brain at 4 times and computed absolute spectrum of each signal. Splines mapping was used for imaging electrical activity of the brain. Statistical differences were calculated using one way ANOVA. Subsequently color was assigned to individual points of the 3D map at 3 different levels of significance: α = 0.05; α = 0.01 and α = 0.001.

Results: The module for significant probability mapping (SPM) was used to find differences between repeated measures EEG on rats. We proofed that there is a significant difference after application of psilocin to rats.

Conclusion: The module for significant probability mapping (SPM) was successfully implemented. MATLAB was used as a programming language. Validity of brain model was confirmed.

This study was supported by projects LO1611/NPUI, MICR VI20172020056; Progress Q35; European Regional Development Fund and by Czech Technical University research program SGS (SGS15/229/OHK4/3T/17).

Visual inspection of the EEG to determine the seizure onset zone (SOZ) in the context of the presurgical evaluation in epilepsy is time-consuming and often challenging or impossible. We offer an approach that uses EEG source imaging (ESI) in combination with functional connectivity analysis (FC) to localize the SOZ from ictal EEG.

Ictal low-density scalp EEG from 111 seizures in 27 patients who were rendered-seizure free after surgery was analyzed. For every seizure, ESI (LORETA) was applied on an artifact-free epoch (1-5s) selected around the seizure onset. Additionally, FC (swADTF) was applied on the reconstructed sources. We estimated the SOZ in two ways: the source with (i)highest power after ESI and (ii)the most outgoing connections after ESI and FC. For both approaches, the distance between the estimated SOZ and the resected zone (RZ) of the patient were calculated.

Using ESI alone, the SOZ was estimated inside the RZ in 31% of the seizures and within 10mm from the border of the RZ in 42%. For 18.5% of the patients, all seizures were estimated within 10mm of the RZ. Using ESI and FC, 72% of the seizures were estimated inside the RZ, and 94% within 10mm. For 85% of the patients, all seizures were estimated within 10mm of the RZ. FC provided a significant added value to ESI alone (p<0.001).

ESI combined with subsequent FC is able to localize the SOZ in a non-invasive way with high accuracy. Therefore it could be a valuable tool in the presurgical evaluation of epilepsy.

EEG-fMRI maps the generators of interictal epileptiform discharges (IEDs) in epilepsy patients to aid pre-surgical evaluation. This requires 1) EEG-defined IED timings and 2) a model of the haemodynamic response (HRF). Research suggests the standard HRF used in cognitive neuroscience may be inappropriate for modelling IEDs. We aimed to derive the HRF to IEDs in a group of children with focal epilepsy. We then tested if this derived HRF could improve EEG-fMRI maps for pre-surgical evaluation.

We collected simultaneous EEG-fMRI data at 1.5T using a 64-channel EEG system. Epileptic events were visually coded following artefact correction. Twenty seven drug resistant focal epilepsy patients in whom the epileptogenic region was confirmed were recruited (post-surgical Engel outcome=1 or visible MR lesion).

Sixteen of these patients had concordant EEG-fMRI maps when using the canonical HRF. This group was used to generate a new HRF using a smooth FIR deconvolution (time window of -22 to 22sec surrounding the IED onset). Subsequently a principal component analysis was used to determine the IED-HRF response across this group. The remaining 11 patients with discordant EEG-fMRI maps were used to assess the improvement this new HRF had on localisation.

Haemodynamic changes up to ~20sec prior to IEDs onset were observed. This early response may represent a metabolic change in state that is predictive of the epileptiform activity. In subjects where the standard canonical basis set failed the IED-HRF was able to localise in 64% of patients. This could potentially increase the clinical yield of EEG-fMRI.

The neural overlap between spectrotemporal sound feature representations in the human cortex during listening to speech and music still remains unclear. To assess this we recorded electrocorticographic data from 8 epileptic patients. Participants listened to natural speech and a music stimulus. For both conditions we built encoding (predicts high gamma neural activity [70-150 Hz] using the spectrogram representation of the sound) and decoding models (predicts the sound spectrogramn by using the high gamma neural activity). Further we used a cross-condition analysis by applying the decoding model built on the speech condition on the music condition and vice versa. We found robust overlaps between the speech and the music condition in terms of anatomical location and frequency tuning in the auditory areas.

This work aims at building a pipeline to analyze the recorded magnetic field from the human heart. We constructed the pipeline using simplified meshes of the heart and torso and simple simulated signals at first and we will continue with sophisticated simulations and validated recordings. The pipeline contains two main parts, namely the forward and the inverse solutions.

The forward solution starts with segmenting an individual Magnetic resonance image (MRI), which we manually segmented into two triangular surface meshes for heart and torso, as a preliminary model, and then we constructed cubic meshes for heart and torso. We also simulated one dipole at a single time point in the middle of the heart, we considered a vertical direction (from head to feet) to represent the potential from the bundle branches, and then we calculated the forward solution using Finite Element Method (FEM) for a magnetometer sensors model. We used the Simbio software for the forward calculation. In order to solve the inverse problem, we also calculated the Lead Field Matrix (LFM) by defining a volumetric grid in the heart.

We performed the inverse solution using two methods, namely low-resolution electromagnetic tomography (LORETA) and Spatiotemporal Kalman Filtering (STKF), which is based on linear state space modeling. We applied the pipeline on two magnetocardiographic (MCG) datasets and we will present our first localization results.

In the future, we intend to use a nonlinear state space model for the STKF, so that it can better describe the dynamics of the heart signal.

Background: Brain source mapping from distant measurements such as M/EEG brings more insight into brain normal and pathological functioning. This inverse problem is commonly carried out based on non-invasive electrode setup, however these surfacic data do not well capture the activities of deep structures. To get a wider picture of the brain activation map, we propose to use the Stereo-EEG (SEEG) setup, consisting in shaft electrodes implanted within the structures of interest.

Methods: Following preliminary works [1] dealing with the necessary conditions for successfull dipole localization from SEEG, we adress the distributed source imaging problem [2]. In particular, we explicitely take into account the forward model uncertainties. Using a variational Bayesian framework, the source time-course and the dipole projection gains are simultaneously optimized. The gain posterior distributions are constrained to remain close to a confident physical model through the introduction of multivariate Gaussian priors, preventing from non-physiological estimates.

Results: We demonstrate under simulations that the method enhance the accuracy of the source time-course estimates as well as the sparsity of the resulting source map. The approach is validated on data of intra-cranial stimulations (for which the position of the source is known), as well as on SEEG data of epileptic spikes, validated by the surgery outcomes.

Références:

[1] V. Caune et. al., Evaluating dipolar source localization feasibility from intracerebral SEEG recordings, NeuroImage, 2014.

[2] S. Le Cam et. al., SEEG dipole source localization based on an empirical Bayesian approach taking into account forward model uncertainties, NeuroImage, 2017.

The uncertainty in MEG-to-MRI co-registrations propagates through the forward model to uncertainty in source reconstruction results of MEG data. However, the common tools for source reconstruction in MEG or EEG analysis do not account for that source of uncertainty and usually only the variance of the noise is considered in the assessment of the standard error or covariance of source parameters. For realistic head models, the computational costs of forward modeling are unfeasible for straightforward Monte Carlo simulations. To overcome this problem, a polynomial expansion of the forward model is constructed as a function of the co-registration parameters.

The six-dimensional co-registration space of three rotation and three translation parameters is sampled using a Metropolis algorithm. The eigen-decomposition of the Metropolis sample covariance matrix provides a map to a six-dimensional uncorrelated parameter space. Based on the uncorrelated parameters, the forward model matrix is expanded in terms of Hermite polynomials. The number of polynomial terms is limited using an adaptive expansion with an error tolerance of 1% resulting in 61 terms for our demonstration model.

For this expansion, the forward model is evaluated at 97 different co-registration parameterizations. The polynomial expansion is used as a computationally cheap surrogate of the forward model construction.

We demonstrate the benefit of the expansion by drawing 10000 independent samples of the linearly constrained minimum variance beamformer for 42 target sources.

Our methods provide a computationally feasible assessment of the distribution of source estimates, e.g. the beamformer activation maxima, based on the uncertainty in MEG-to-MRI co-registrations.

Transcranial electric stimulation (TES) is a non-invasive technique for cortical stimulation. Depending on the electrode positions and polarity, the current density distribution in the head changes its amplitude and orientation. For treatment of depression, several stimulation montages for targeting the dorsolateral prefrontal cortex (DLPFC), were introduced.

With the present study, we aimed to evaluate magnitude and orientation differences in the DLPFC originating from different stimulation montages.

We generated an individual five compartment finite element model from structural magnetic resonance images of a volunteer (age 23 years). For TES simulations with 1 mA current strength, we placed 5x7 cm patch electrodes as anode at position F3 and used a cathode at positions Fp2, F4, F8 and P2. We analysed the amplitude and orientation of the resulting current density distributions in the DLPFC.

The mean current density in the DLPFC was 0.11±0.03 mA/m² for F3/Fp2, 0.06±0.01 mA/m² for F3/F4, 0.10±0.02 mA/m² for F3/F8 and 0.08±0.01 mA/m² for F3/P2. The current density orientation difference in the DLPFC between the montages F3/F4 and F3/Fp2 was 32.9±6.4 degrees, between F3/F4 and F3/F8 27.6±5.6 degrees, and between F3/F4 and F3/P2 41.6±6.6 degrees.

Our simulation results demonstrate considerable effects of the stimulation montage on the amplitude and the orientation on the current density in the DLPFC. Bai et al. (2014) compared similar stimulation montages and also found highest stimulation intensities in the DLPFC for the F3/F4 montage. With our results, we underline the importance of detailed models in TES simulations and the consideration of the stimulation montage.

The N1 component of the visual ERP is considered to be sensitive to print, as it becomes larger for words than for symbols during learning to read in children. In adults, however, the N1 in response to visual words typically corresponds to a lengthy GFP segment, which opens the question whether different neural processes occur in this time range. Here, we summarize 4 different ERP studies in adults with various designs and in different writing systems. The studies analyze different parts of the N1 component in order to test differential sensitivity to various aspects of visual word processing. Study 1 shows that lexicality effects occur in the late N1, but not in the early N1 in German readers. Study 2 shows that repetition effects in Chinese readers occur in the late N1, but not in the early N1. Similarly, study 3 shows masked priming effects in Chinese readers in the late N1, but not in the early N1. Finally, study 4 shows sensitivity for print irrespective of task in Chinese readers in the N1 onset, but task modulation in the N1 offset. Taken together, the results suggest that the N1 component is not as functionally homogeneous as previously thought and that during the lengthy N1 duration, different visual-orthographic processes are unfolding.