Evaluation of Long-Short Term Memory Networks for M/EEG Source Imaging with Simulated and Real EEG Data

Hecker, Lukas; Rupprecht, Rebekka; Elst, Ludger Tebartz van; Kornmeier, Jürgen

doi:10.1101/2022.04.13.488148

Cited by 5 publications

(5 citation statements)

References 83 publications

(127 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As we have shown in this work, combining FLEX-MUSIC and a Champagne algorithm may be useful, although Champagne may overestimate the size of neural sources in some cases. An interesting family of recently developed inverse solution are based on artificial neural networks (ANNs, Hecker et al, 2020Hecker et al, , 2022. In a next step we plan to compare a potentially further optimized version of FLEX-MUSIC with this family of ANN-based solvers.…”

Section: Discussionmentioning

confidence: 99%

“…A novel class of inverse solvers arose in the past decade that utilize the recent advances in machine learning, predominantly artificial neural networks (ANNs) to solve M/EEG inverse problems. These approaches require training an ANN to produce a source estimate based on simulated pairs of source and M/EEG activity and achieve high accuracy compared to many conventional methods (Cui et al, 2019; Hecker, Rupprecht, van Elst, & Kornmeier, 2020, 2022; Pantazis & Adler, 2021).…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Source Localization Using Recursively Applied and Projected MUSIC with Flexible Extent Estimation

Hecker

Elst

Kornmeier

2023

Preprint

Self Cite

View full text Add to dashboard Cite

Magneto- and electroencephalography (M/EEG) are widespread techniques to measure neural activity in-vivo at a high temporal resolution but low spatial resolution. Locating the neural sources underlying the M/EEG poses an inverse problem, which is ill-posed. We developed a new method based on Recursive Application of Multiple Signal Classification (MUSIC). Our proposed method is able to recover not only the locations but, in contrast to other inverse solutions, also the extent of active brain regions flexibly (FLEX-MUSIC). This is achieved by allowing it to search not only for single dipoles but also dipole clusters of increasing extent to find the best fit during each recursion. FLEX-MUSIC achieved the highest accuracy for both single dipole and extended sources compared to all other methods tested. Remarkably, FLEX-MUSIC was capable to accurately estimate the level of sparsity in the source space (r = 0.82), whereas all other approaches tested failed to do so (r ≤ 0.18). The average computation time of FLEX-MUSIC was considerably lower compared to a popular Bayesian approach and comparable to that of another recursive MUSIC approach and eLORETA. FLEX-MUSIC produces only few errors and was capable to estimate the extent of sources reliably. The accuracy and low computation time of FLEX-MUSIC renders it an improved technique to solve M/EEG inverse problems both in neuroscience research and potentially in pre-surgery diagnostic in epilepsy.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Source Localization Using Recursively Applied and Projected MUSIC with Flexible Extent Estimation

Hecker

Elst

Kornmeier

2023

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…Neural sources underlying the ERPs were calculated using a fully-connected ANN, which has shown to yield accurate inverse solutions with respect to source location and estimation of source extent [31,13]. We used simulated neural sources and scalp EEG data to train the fully-connected (FC) ANN model.…”

Section: Artificial Neural Network (Ann)mentioning

confidence: 99%

“…We found that the SNR of our participants ranged from 4.2 -19.9, which was again used as the parameter range for the simulation. For further details on the simulation procedure please refer to [13].…”

Section: Artificial Neural Network (Ann)mentioning

confidence: 99%

Source Reconstruction of the ERP Uncertainty Effects reveals common Neural Mechanisms for different Stimulus Categories

Hecker

Joos

Feige

et al. 2023

Preprint

View full text Add to dashboard Cite

During the observation of an ambiguous figure, our perception becomes unstable and alternates repeatedly between mutual exclusive interpretations. Tiny changes of the stimulus features can disambiguate the figure and stabilize percepts. Recent EEG studies found much smaller amplitudes of two event-related potentials (ERPs), an anterior P200, 200 ms after stimulus onset and a posterior P400, 200 ms later, when participants observed an ambiguous stimulus compared to disambiguated stimulus variants. Interestingly, this pattern of results was found across stimuli differing in ambiguity (geometry, motion and gestalt) and in visibility. We postulate a meta-perceptual / meta-cognitive evaluation instance that rates the reliability of perceptual constructs at a high processing level generalized beyond sensory details. We further postulate that the above described ERP effects reflect the outcome of this evaluation process. According to these hypotheses, the distributions of these ERP effects on the scalp across three different stimulus categories, should originate from the same neural structures in the sensor and source space. This was tested by calculating EEG inverse solutions using a novel artificial neural network-based approach.We found very similar sources across stimulus categories, both on the level of individual participants and on the group level. Regions involved in the earlier processing steps (P200) encompass lateral occipital cortex, inferior parietal cortex and medial cingulate cortex. Later processes (P400) originated mostly from the right inferior temporal cortex. Our findings were consistent with comparable studies using functional magnetic resonance imaging and EEG source imaging. In summary, we found highly coherent neural sources of the ERP Uncertainty Effects. The underlying processes may be related to a common uncertainty resolution at a higher processing level beyond sensory details.

show abstract

“…A novel class of inverse solvers arose in the past decade that utilize the recent advances in machine learning, predominantly artificial neural networks (ANNs) to solve M/EEG inverse problems. These approaches require training an ANN to produce a source estimate based on simulated pairs of source and M/EEG activity and achieve high accuracy compared to many conventional methods (Cui et al, 2019 ; Hecker et al, 2020 , 2022 ; Pantazis and Adler, 2021 ). ANNs are prone to biases in the training data, wherefore their application is yet limited.…”

Section: Introductionmentioning

confidence: 99%

Source localization using recursively applied and projected MUSIC with flexible extent estimation

2023

Self Cite

View full text Add to dashboard Cite

Magneto- and electroencephalography (M/EEG) are widespread techniques to measure neural activity in-vivo at a high temporal resolution but low spatial resolution. Locating the neural sources underlying the M/EEG poses an inverse problem, which is ill-posed. We developed a new method based on Recursive Application of Multiple Signal Classification (MUSIC). Our proposed method is able to recover not only the locations but, in contrast to other inverse solutions, also the extent of active brain regions flexibly (FLEX-MUSIC). This is achieved by allowing it to search not only for single dipoles but also dipole clusters of increasing extent to find the best fit during each recursion. FLEX-MUSIC achieved the highest accuracy for both single dipole and extended sources compared to all other methods tested. Remarkably, FLEX-MUSIC was capable to accurately estimate the level of sparsity in the source space (r = 0.82), whereas all other approaches tested failed to do so (r ≤ 0.18). The average computation time of FLEX-MUSIC was considerably lower compared to a popular Bayesian approach and comparable to that of another recursive MUSIC approach and eLORETA. FLEX-MUSIC produces only few errors and was capable to reliably estimate the extent of sources. The accuracy and low computation time of FLEX-MUSIC renders it an improved technique to solve M/EEG inverse problems both in neuroscience research and potentially in pre-surgery diagnostic in epilepsy.

show abstract

Evaluation of Long-Short Term Memory Networks for M/EEG Source Imaging with Simulated and Real EEG Data

Cited by 5 publications

References 83 publications

Source Localization Using Recursively Applied and Projected MUSIC with Flexible Extent Estimation

Source Localization Using Recursively Applied and Projected MUSIC with Flexible Extent Estimation

Source Reconstruction of the ERP Uncertainty Effects reveals common Neural Mechanisms for different Stimulus Categories

Source localization using recursively applied and projected MUSIC with flexible extent estimation

Contact Info

Product

Resources

About