A parallel multi-scale time-frequency block convolutional neural network based on channel attention module for motor imagery classification. (January 2023)

• Record Type: Journal Article
• Authors: Li, Hongli; Chen, Hongyu; Jia, Ziyu; Zhang, Ronghua; Yin, Feichao
• Is Part Of: Biomedical signal processing and control. Volume 79(2023)Part 1
• Publication Date: 2023-01
• Subjects: Brain-computer interface -- Motor imagery -- Deep learning -- Convolutional Neural Networks -- Attention mechanism; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104066
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: The proposed model adaptively extracts the time, frequency, and time-frequency features with a multi-branch structure. The proposed ResneXt and channel attention module fuse and filter the features, and further improve the accuracy. The proposed model has been applied to multiple datasets with perfect generalization performance and robustness. Compared with the state-of-the-art baseline models, the proposed model achieves the highest average accuracy. Abstract: The motor imagery brain- computer interface (MI-BCI) based on electroencephalography (EEG) enables direct communication between the human brain and external devices. In this paper, the MTFB-CNN, a parallel multi-scale time-frequency block convolutional neural network based on the channel attention module, is proposed for EEG signals decoding, which can adaptively extract the time, frequency, and time-frequency domain features through parallel multi-scale time-frequency blocks, and then fuses and filters the features through attention mechanism and residual module. Experimental results based on the BCI Competition IV 2a and 2b datasets and the high gamma dataset show that the model achieves the highest average accuracy and kappa compared with existing baseline models. The MTFB-CNN is a novel and effective end-to-end model for decoding EEG signals without complex signals pre-processing operations, which has multi-scale feature extraction capability, making it successful in MI-BCI applications.