EEG source localization using spatio-temporal neural network

Background: Electroencephalography (EEG) may be used as an objective diagnosis tool for diagnosing various disorders. Recently, source localization from EEG is being used in the analysis of real-time brain monitoring applications. However, inverse problem reduces the accuracy in EEG signal processing systems. Objectives: This paper presents a new method of EEG source localization using variational mode decomposition (VMD) and standardized the low resolution brain electromagnetic tomography (sLORETA) inverse model. The focus is to compare the effectiveness of the proposed approach for EEG signals of depression patients. Method: As the first stage, real EEG recordings corresponding to depression patients are decomposed into various mode functions by applying VMD. Then, closely related functions are analyzed using the inverse modelling-based source localization procedures such as sLORETA. Simulations have been carried out on real EEG databases for depression to demonstrate the effectiveness of the proposed techniques. Results: The performance of the algorithm has been assessed using localization error (LE), mean square error and signal to noise ratio output corresponding to simulated EEG dipole sources and real EEG signals for depression. In order to study the spatial resolution for cortical potential distribution, the main focus has been on studying the effects of noise sources and estimating LE of inverse solutions. More accurate and robust localization results show that this methodology is very promising for EEG source localization of depression signals. Conclusion: It can be said that proposed algorithm efficiently suppresses the influence of noise in the EEG inverse problem using simulated EEG activity and EEG database for depression. Such a system may offer an effective solution for clinicians as a crucial stage of EEG pre-processing in automated depression detection systems and may prevent delay in diagnosis.

Evaluation of different cortical source localization methods using simulated and experimental EEG data

Performance of source imaging techniques of spatially extended generators of uterine activity

Chemical exchange saturation transfer (CEST) magnetic resonance imaging has shown promise for classifying tumors based on their aggressiveness, but CEST contrast is complicated by multiple signal sources and thus prolonged acquisition times are often required to extract the signal of interest. We investigated whether deep learning could help identify pertinent Z-spectral features for distinguishing tumor aggressiveness as well as the possibility of acquiring only the pertinent spectral regions for more efficient CEST acquisition. Human breast cancer cells, MDA-MB-231 and MCF-7, were used to establish bi-lateral tumor xenografts in mice to represent higher and lower aggressive tumors, respectively. A convolutional neural network (CNN)-based classification model, trained on simulated data, utilized Z-spectral features as input to predict labels of different tissue types, including MDA-MB-231, MCF-7, and muscle tissue. Saliency maps reported the influence of Z-spectral regions on classifying tissue types. The model was robust to noise with an accuracy of more than 91.5% for low and moderate noise levels in simulated testing data (SD of noise less than 2.0%). For in vivo CEST data acquired with a saturation pulse amplitude of 2.0 μT, the model had a superior ability to delineate tissue types compared with Lorentzian difference (LD) and magnetization transfer ratio asymmetry (MTRasym ) analysis, classifying tissues to the correct types with a mean accuracy of 85.7%, sensitivity of 81.1%, and specificity of 94.0%. The model's performance did not improve substantially when using data acquired at multiple saturation pulse amplitudes or when adding LD or MTRasym spectral features, and did not change when using saliency map-based partial or downsampled Z-spectra. This study demonstrates the potential of CNN-based classification to distinguish between different tumor types and muscle tissue, and speed up CEST acquisition protocols.

Normally, detection, enumeration, and localization of sources are treated as separate problems, and different processing techniques are employed to solve them. Source enumeration or prior knowledge of the number of sources is a pre-requisite for the existing methods of source localization. In this paper, we present a method called Residual Data Vector Processing (RDVP) for joint detection and 3-dimensional (3-D) localization of an unknown number of underwater acoustic sources using measurements by a 3-D array of sensors. Theoretically, maximum likelihood estimation (MLE) is the best method of source localization since MLE provides asymptotically efficient estimates. But the computational complexity of MLE is prohibitively high. The localization accuracy of the proposed method compares very favorably with that of other methods of 3-D source localization such as Capon and MUSIC processors, while its computational complexity is much lower than that of the existing methods.

In the construction industry, construction machinery are an important factor in the overall productivity and efficiency of a worksite. Thus, emphasis is put on the monitoring of actions conducted by construction machinery. This was traditionally done manually by humans, which is a timeconsuming and laborious task. Automatic action recognition of construction machinery is therefore needed. The field of action recognition is predominantly occupied by Deep Learning approaches and several previous works focused on adapting such approaches for construction machinery. However, the issue of obtaining training data is particularly troublesome for construction machinery. Our previous work proposed a Deep Learning method for learning an action recognition model from training data generated in a simulator using video filters but the precise contributions of the introduced video filter were unclear. The purpose of this study is therefore to clarify the effects of video filters for learning an action recognition model for construction machinery from simulated training data.

Brain oscillations from EEG and MEG shed light on neurophysiological mechanisms of human behavior. However, to extract information on cortical processing, researchers have to rely on source localization methods that can be very broadly classified into current density estimates such as exact low-resolution brain electromagnetic tomography (eLORETA), minimum norm estimates (MNE), and beamformers such as dynamic imaging of coherent sources (DICS) and linearly constrained minimum variance (LCMV). These algorithms produce a distributed map of brain activity underlying sustained and transient responses during neuroimaging studies of behavior. On the other hand, there are very few comparative analyses that evaluates the “ground truth detection” capabilities of these methods. The current article evaluates the reliability in estimation of sources of spectral event generators in the cortex using a two-pronged approach. First, simulated EEG data with point dipoles and distributed dipoles are used to validate the accuracy and sensitivity of each one of these methods of source localization. The abilities of the techniques were tested by comparing the localization error, focal width, false positive (FP) ratios while detecting already known location of neural activity generators under varying signal-to-noise ratios (SNRs). Second, empirical EEG data during auditory steady state responses (ASSRs) in human participants were used to compare the distributed nature of source localization. All methods were successful in recovery of point sources in favorable signal to noise scenarios and could achieve high hit rates if FPs are ignored. Interestingly, focal activation map is generated by LCMV and DICS when compared to eLORETA while control of FPs is much superior in eLORETA. Subsequently drawbacks and strengths of each method are highlighted with a detailed discussion on how to choose a technique based on empirical requirements.

Underwater source localization using receiver arrays is often achieved with a purely model-based approached, i.e., matched-field processing using simulated replica-field. However, these approaches only yield reasonable predictions if the complicated ocean environment can be precisely described, which is often a daunting task. Alternatively, by turning to a data-driven approach, source localization can be achieved without the strong dependence on model-parameters. Recently, source localization in the Santa Barbara Channel (∼580 m depth, downward refracting profile) has been achieved via measurements of ships of opportunity, and machine learning classifiers. In this presentation, relative channel impulse responses, estimated via the ray-based blind deconvolution, are used to train a neural network to predict the latitude and longitude of a surface ship within the Santa Barbara Channel. Localization results for simulated and experiment data are presented and implications for future work are discussed.Underwater source localization using receiver arrays is often achieved with a purely model-based approached, i.e., matched-field processing using simulated replica-field. However, these approaches only yield reasonable predictions if the complicated ocean environment can be precisely described, which is often a daunting task. Alternatively, by turning to a data-driven approach, source localization can be achieved without the strong dependence on model-parameters. Recently, source localization in the Santa Barbara Channel (∼580 m depth, downward refracting profile) has been achieved via measurements of ships of opportunity, and machine learning classifiers. In this presentation, relative channel impulse responses, estimated via the ray-based blind deconvolution, are used to train a neural network to predict the latitude and longitude of a surface ship within the Santa Barbara Channel. Localization results for simulated and experiment data are presented and implications for future work are discussed.

Objective. There is longstanding interest in using EEG measurements to inform transcranial Electrical Stimulation (tES) but adoption is lacking because users need a simple and adaptable recipe. The conventional approach is to use anatomical head-models for both source localization (the EEG inverse problem) and current flow modeling (the tES forward model), but this approach is computationally demanding, requires an anatomical MRI, and strict assumptions about the target brain regions. We evaluate techniques whereby tES dose is derived from EEG without the need for an anatomical head model, target assumptions, difficult case-by-case conjecture, or many stimulation electrodes. Approach. We developed a simple two-step approach to EEG-guided tES that based on the topography of the EEG: (1) selects locations to be used for stimulation; (2) determines current applied to each electrode. Each step is performed based solely on the EEG with no need for head models or source localization. Cortical dipoles represent idealized brain targets. EEG-guided tES strategies are verified using a finite element method simulation of the EEG generated by a dipole, oriented either tangential or radial to the scalp surface, and then simulating the tES-generated electric field produced by each model-free technique. These model-free approaches are compared to a ‘gold standard’ numerically optimized dose of tES that assumes perfect understanding of the dipole location and head anatomy. We vary the number of electrodes from a few to over three hundred, with focality or intensity as optimization criterion. Main results. Model-free approaches evaluated include (1) voltage-to-voltage, (2) voltage-to-current; (3) Laplacian; and two Ad-Hoc techniques (4) dipole sink-to-sink; and (5) sink to concentric. Our results demonstrate that simple ad hoc approaches can achieve reasonable targeting for the case of a cortical dipole, remarkably with only 2–8 electrodes and no need for a model of the head. Significance. Our approach is verified directly only for a theoretically localized source, but may be potentially applied to an arbitrary EEG topography. For its simplicity and linearity, our recipe for model-free EEG guided tES lends itself to broad adoption and can be applied to static (tDCS), time-variant (e.g., tACS, tRNS, tPCS), or closed-loop tES.

Images of brain activation typically comprise of small patches of activation with large regions of non-activation. Such images can be considered as minimum support images, where non-zero values occupy only a small area. In this paper, we propose a new minimum support imaging method that solves the bioelectromagnetic inverse problem. We formulate the problem using Lead-fields under Tikhonov theory which replaces the original ill-posed inverse problem by a well-posed minimization of a Tikhonov parametric functional consisting of a sum of the misfit and stabilizer functions. The stabilizer function embeds apriori assumptions about brain activation, which we assume is minimum support. We illustrate the proposed method in simulations. Using realistic head and sensor geometry, we generate forward model MEG data for multiple-dipole sources. This simulated forward model data with additive noise was used to perform data inversion. Monte Carlo simulations indicate that the accuracy in localization of single dipoles is 2.1 mm for low noise conditions. While this performance is consistent with performance of conventional parametric dipole inversion, the minimum support method is capable of reconstructing multiple distributed sources.

Shallow water is a complex sound propagation medium, which is affected by the varying spatial–temporal ocean environment. Taking this complexity into account, the classical processing techniques of source localization and environmental inversion may be improved. In this work, a joint tracking approach for the moving source and environmental parameters of the range-dependent and time-evolving environment in shallow water is presented. The tracking scheme treats both the source parameters (e.g., source depth, range, and speed) and the environmental parameters (e.g., water column sound speed profile (SSP) and sediment parameters) at the source location as unknown variables that evolve as the source moves. To counter sample impoverishment and robustly characterize the evolution of the parameters, an improved particle filter (PF), which is an extension of the standard PF, is proposed. Two examples with simulated data in a slowly changing environment and experimental data collected during the ASIAEX experiment are utilized to demonstrate the effectiveness of the joint approach. The results show that we were able to track the source and environmental parameters simultaneously, and the uncertainties were evaluated in the form of time-evolving posterior probability densities (PPDs). The performance comparison confirms that the improved PF is superior to the standard PF, as it can reduce the parameter uncertainties. The tracking capabilities of the improved PF were verified with high accuracy in real-time source localization and well-estimated rapidly varying parameters. Moreover, the influence of different particle numbers on the improved PF tracking performance is also illustrated.

Independent components analysis (ICA) has previously been used to denoise EEG/magnetoencephalography (MEG) signals before performing neural source localization. Source localization is then performed using a method such as beamforming or dipole fitting. Here we show how ICA can also be used as a source localization method, negating the need for beamforming and dipole fitting. This type of approach is valid whenever an estimate of the forward (mixing) model for all putative source locations is available, which includes EEG and MEG applications. The proposed method consists of estimating the forward model using the laws of physics, estimating a second forward model using ICA, and then correlating the columns of the matrices that represent the two forward models. We show that, when synthetic data are used, the proposed localization method produces a smaller localization error than several alternatives. We also show localization results for real auditory-evoked MEG data.

Owing to the multipath effect, the source localization in shallow water has been an area of active interest. However, most methods for source localization in shallow water are sensitive to the assumed model of the underwater environment and have poor robustness against the underwater channel uncertainty, which limit their further application in practical engineering. In this paper, a new method of source localization in shallow water, based on vector optimization concept, is described, which is highly robust against environmental factors affecting the localization, such as the channel depth, the bottom reflection coefficients, and so on. Through constructing the uncertainty set of the source vector errors and extracting the multi-path sound rays from the sea surface and bottom, the proposed method can accurately localize one or more sources in shallow water dominated by multipath propagation. It turns out that the natural formulation of our approach involves minimization of two quadratic functions subject to infinitely many nonconvex quadratic constraints. It shows that this problem (originally intractable) can be reformulated in a convex form as the so-called second-order cone program (SOCP) and solved efficiently by using the well-established interior point method, such as the software tool, SeDuMi. Computer simulations show better performance of the proposed method as compared with existing algorithms and establish a theoretical foundation for the practical engineering application.

Neural synchronization is a key mechanism to a wide variety of brain functions, such as cognition, perception, or memory. High temporal resolution achieved by EEG recordings allows the study of the dynamical properties of synchronous patterns of activity at a very fine temporal scale but with very low spatial resolution. Spatial resolution can be improved by retrieving the neural sources of EEG signal, thus solving the so-called inverse problem. Although many methods have been proposed to solve the inverse problem and localize brain activity, few of them target the synchronous brain regions. In this paper, we propose a novel algorithm aimed at localizing specifically synchronous brain regions and reconstructing the time course of their activity. Using multivariate wavelet ridge analysis, we extract signals capturing the synchronous events buried in the EEG and then solve the inverse problem on these signals. Using simulated data, we compare results of source reconstruction accuracy achieved by our method to a standard source reconstruction approach. We show that the proposed method performs better across a wide range of noise levels and source configurations. In addition, we applied our method on real dataset and identified successfully cortical areas involved in the functional network underlying visual face perception. We conclude that the proposed approach allows an accurate localization of synchronous brain regions and a robust estimation of their activity.

We present a source localization method based upon a sparse representation of sensor measurements with an overcomplete basis composed of samples from the array manifold. We enforce sparsity by imposing an /spl lscr//sub 1/-norm penalty; this can also be viewed as an estimation problem with a Laplacian prior. Explicitly enforcing the sparsity of the representation is motivated by a desire to obtain a sharp estimate of the spatial spectrum which exhibits superresolution. To summarize multiple time samples we use the singular value decomposition (SVD) of the data matrix. Our formulation leads to an optimization problem, which we solve efficiently in a second-order cone (SOC) programming framework by an interior point implementation. We demonstrate the effectiveness of the method on simulated data by plots of spatial spectra and by comparing the estimator variance to the Cramer-Rao bound (CRB). We observe that our approach has advantages over other source localization techniques including increased resolution; improved robustness to noise, limitations in data quantity, and correlation of the sources; as well as not requiring an accurate initialization.