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

Abstract

1AbstractMagneto- and electroencephalography (M/EEG) are widespread techniques to measure neural activityin-vivoat a high temporal resolution but relatively low spatial resolution. Locating the sources underlying the M/EEG poses an inverse problem, which is itself ill-posed. In recent years, a new class of source imaging methods was developed based on artificial neural networks. We present a long-short term memory (LSTM) network to solve the M/EEG inverse problem. It integrates low computational cost, exploitation of both the coarse spatial but also the excellent temporal information from the EEG, input flexibility and robustness to noise. We compared the LSTM network with classical inverse solutions using both simulation data and real EEG data, recorded in epileptic patients during intracranial stimulation. The LSTM network shows higher accuracy on multiple metrics and for varying numbers of neural sources, compared to classical inverse solutions but also compared to our alternative architecture without integration of temporal information. The performance of the LSTM network regarding its robustness to noise and low localization errors renders it a promising inverse solution to be considered in future source localization studies and for clinical applications.