Classification of motor imagery based on multi-scale feature extraction and fusion-residual temporal convolutional network.

Brain-computer interface (BIC) decodes electroencephalogram (EEG) signals to realize the interaction between brain and external devices. However, traditional methods show limited performance in motor imagery electroencephalogram (MI-EEG) classification. In this paper, we introduce a multi-scale temporal convolutional network (MS-TCNet) that employs parallel multi-scale convolutions for spatiotemporal feature extraction, efficient channel attention (ECA) for channel weights optimization, and fusion-residual temporal convolution (FR-TCN) for high-level temporal feature capture. Experimental results show that MS-TCNet achieved remarkable decoding accuracies of 87.85% and 92.85% on the BCI IV-2a and BCI IV-2b datasets, respectively. The proposed MS-TCNet surpasses existing baseline models across various performance metrics, demonstrating its effectiveness in advancing MI-EEG decoding.