AUTHOR=Jiao Meng , Wan Guihong , Guo Yaxin , Wang Dongqing , Liu Hang , Xiang Jing , Liu Feng 

TITLE=A Graph Fourier Transform Based Bidirectional Long Short-Term Memory Neural Network for Electrophysiological Source Imaging

JOURNAL=Frontiers in Neuroscience

VOLUME=Volume 16 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2022.867466

DOI=10.3389/fnins.2022.867466

ISSN=1662-453X

ABSTRACT=The electroencephalogram (EEG) source localization has been one of the most critical brain science research issues. Over the past few decades, the most preeminent ways for solving the EEG inverse problem are to employ prior information or regularization, but this approach is intractable to localize the deep active brain source. Here, the graph Fourier transform (GFT) and the bi-directional long-short term memory (BiLSTM) neural network are introduced to solve the EEG inverse problem in a more efficient and robust way. The presented GFT-BiLSTM in this paper not only fully utilizes spatial information of the brain source signal, but also makes full use of the powerful self-learning ability of the BiLSTM. In the GFT-BiLSTM, the GFT is used for the signal decomposition of the source signal to reduce its dimension. The BiLSTM is adopted to learn the mapping relationship between the brain sources and the recorded EEG. The results show that GFT-BiLSTM outperforms other state-of-the-art inverse models in synthetic data. Regardless of whether the activated region is single or multiple, the area under the curve (AUC) corresponding to GFT-BiLSTM can be reliably above 0.96. When the signal-to-noise ratio (SNR) varies, the GFT-BiLSTM exhibits strong robustness with the highest AUC while the lowest localization error (LE). This fully demonstrates the superiority of GFT-BiLSTM when applied to solve the EEG inverse problem.