A Novel Two-Branch Convolutional Neural Network for EEG Decoding of Motor Imagery

In recent years, deep learning has been applied to the classification of motor imagery EEG, which can decode and classify EEG signals end-to-end. However, compared with picture datasets, EEG datasets are relative small and difficult to adapt to deeper networks. To solve this problem, we propose a shallow two-branch convolutional neural network (TBCNNet) to decode EEG information. More EEG information can be obtained by fusing the spatio-temporal features extracted from each single branch. Thus, the accuracy of EEG decoding can be improved without increasing the depth of the network. We tested this method on the BCI Competition IV-2a dataset, and the average decoding accuracy reached 86.59%, the average kappa value reached 0.82, and the average Fl-score reached 0.86. The results show that the proposed method can effectively improve the performance of motor imagery decoding.