Deep-Learning-Based Motor Imagery EEG Classification by Exploiting the Functional Connectivity of Cortical Source Imaging

Neurology Information Science Electronic Information Artificial Intelligence Automation Medicine Medical Imaging
deep learning motor imagery EEG cortical source imaging brain functional connectivity spatiotemporal features graph convolutional network

Deep-learning-based Motor Imagery EEG Classification by Exploiting the Functional Connectivity of Cortical Source Imaging

Research Background and Motivation

A brain-computer interface (BCI) is a system that directly decodes and outputs brain activity information without relying on related neural pathways and muscles, thereby achieving communication or control of external devices. Common signals used in BCI systems include electroencephalogram (EEG), magnetoencephalography (MEG), and functional magnetic resonance imaging (fMRI). Among them, EEG is the most commonly used signal because it is non-invasive, easy to implement, cost-effective, and devoid of ethical challenges.

Motor Imagery (MI) is an important paradigm in BCIs. Under non-stimulus conditions, motor imagery tasks spontaneously generate motor imagery EEG signals (MI-EEG) during the period. The MI-EEG signal may embed patterns of neural activity in the motor cortex during motor intention. Hence, decoding MI-EEG signals has become a hot research topic for achieving mental control of external devices through BCI systems.

Current MI-EEG classification methods involve various feature extraction and machine learning schemes. However, these methods still need improvements in classification accuracy and model adaptability among individuals. This paper proposes a new source domain MI-EEG classification algorithm to address these issues.

Research Origin

This paper, titled “Deep-learning-based motor imagery EEG classification by exploiting the functional connectivity of cortical source imaging,” was authored by Bian Doudou, Ma Yue, Huang Jiayin, Xu Dongyang, Wang Zhi, Cai Shengsheng, Wang Jiajun, and Hu Nan. The paper was published online on February 10, 2024, in the journal “Signal, Image and Video Processing.” The research was exclusively licensed by Springer-Verlag London Ltd and is part of Springer Nature.

Research Details

Research Workflow

The workflow of this research includes the following steps:

  1. High-spatial-resolution De-noised Electrophysiological Source Imaging (ESI): Using the Champagne algorithm with noise self-learning to generate high-spatial-resolution cortical source imaging.
  2. Functional Connectivity Measurement: Calculating imaginary coherence (ICoh) in the motor cortex to form a graph structure of the motor cortex source space during MI.
  3. Graph Convolutional Network (GCN): Constructing a GCN to extract spatial features using the graph structure built by ICoh.
  4. Temporal Convolutional Network (TCN): Using TCN and multi-head attention mechanisms to extract multi-scale temporal features, based on the GCN’s spatial attention mechanism for interaction between spatial and temporal features.
  5. Feature Combination and Classification: Combining all extracted features to obtain the final classification result.

Subjects and Samples

The study used the PhysioNet EEG Motor Movement/Imagery dataset, which contains 109 subjects. Each subject performed 14 rounds of MI tasks, with multiple MI tasks conducted in each round.