MSHANet: a multi-scale residual network with hybrid attention for motor imagery EEG decoding.

EEG decoding plays a crucial role in the development of motor imagery brain-computer interface. Deep learning has great potential to automatically extract EEG features for end-to-end decoding. Currently, the deep learning is faced with the chanllenge of decoding from a large amount of time-variant EEG to retain a stable peroformance with different sessions. This study proposes a multi-scale residual network with hybrid attention (MSHANet) to decode four motor imagery classes. The MSHANet combines a multi-head attention and squeeze-and-excitation attention to hybridly focus on important information of the EEG features; and applies a multi-scale residual block to extracts rich EEG features, sharing part of the block parameters to extract common features. Compared with seven state-of-the-art methods, the MSHANet exhits the best accuracy on BCI Competition IV 2a with an accuracy of 83.18% for session- specific task and 80.09% for cross-session task. Thus, the proposed MSHANet decodes the time-varying EEG robustly and can save the time cost of MI-BCI, which is beneficial for long-term use.