A Long Short-Term Memory Network for Sparse Spatiotemporal EEG Source Imaging.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biomedical Engineering

Background:

• The electroencephalography (EEG) inverse problem is inherently underdetermined, challenging noninvasive neuroimaging.
• Existing EEG source imaging methods often rely on direct inverse operations, making them susceptible to noise and solution strategy biases.
• Understanding cortical directional networks is crucial for exploring neural processes noninvasively.

Purpose of the Study:

• To develop a robust source imaging technique for sparse spatiotemporal EEG source estimation.
• To introduce Deep Brain Neural Network (DeepBraiNNet) as a novel approach to overcome limitations of traditional EEG inverse solutions.
• To validate DeepBraiNNet's performance in recovering spatiotemporal sources and analyzing brain networks.

Main Methods:

• Developed Deep Brain Neural Network (DeepBraiNNet), a novel source imaging technique utilizing Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM).
• DeepBraiNNet approximates the inverse operation for the lead field matrix, avoiding direct inverse calculations prone to noise.
• Simulations were performed on various source patterns and noise conditions to assess performance.

Main Results:

• DeepBraiNNet accurately recovered spatiotemporal sources across diverse simulations, outperforming state-of-the-art methods.
• The technique successfully estimated sparse Motor Imagery (MI)-related activation patterns from a real dataset.
• Constructed MI-related cortical neural networks revealed strong contralateral patterns for MI tasks.

Conclusions:

• DeepBraiNNet offers a robust and effective alternative for spatiotemporal EEG source imaging.
• The method provides accurate source estimation, outperforming conventional techniques, especially under noisy conditions.
• DeepBraiNNet facilitates the construction of functional brain networks, aiding in the understanding of neural processes like motor imagery.