Deep-learning-based motor imagery EEG classification by exploiting the functional connectivity of cortical source imaging
Doudou Bian,Yue Ma,Jiayin Huang,Dongyang Xu,Zhi Wang,Shengsheng Cai,Jiajun Wang,Nan Hu
DOI: https://doi.org/10.1007/s11760-023-02965-6
IF: 1.583
2024-02-11
Signal Image and Video Processing
Abstract:Motor imagery (MI) is a commonly used brain–computer interface paradigm, and decoding the MI-EEG signals has been an active research area in recent years. The existing methods involved various feature extraction and machine learning schemes, while classification accuracy and inter-individual model adaptation still need to be improved. To address these issues, a novel source-domain MI-EEG classification algorithm is proposed in this paper. First, the Champagne algorithm with noise self-learning, is adopted to achieve high-spatial-resolution denoised electrophysiological source imaging (ESI) on the cortex. Second, a kind of brain functional connectivity metric, imaginary coherence (iCOH), is used to exploit the source spatial features in the motor cortex. The iCOH in the motor cortex is calculated to form the graph structure of motor cortical source space during MI, by which graph convolutional networks (GCNs) are constructed to extract the spatial features. Multi-scale temporal features are also derived by temporal convolutional network (TCN) along with multi-head attention mechanism, and spatial attention based on GCN is also used for the interaction of spatio-temporal features. Finally, all the extracted features are combined to give the ultimate classification result. The MI-EEG classification performance of the proposed algorithm is evaluated on the PhysioNet EEG Motor Movement/Imagery Datasetis superior, both for the results of intra-subject fivefold cross validation experiments and subject-specific model training experiments.
engineering, electrical & electronic,imaging science & photographic technology