Deep CNN model based on serial-parallel structure optimization for four-class motor imagery EEG classification. (February 2022)

• Record Type: Journal Article
• Authors: Zhao, Xuefei; Liu, Dong; Ma, Li; Liu, Quan; Chen, Kun; Xie, Shane; Ai, Qingsong
• Is Part Of: Biomedical signal processing and control. Volume 72(2022)Part A
• Publication Date: 2022-02
• Subjects: Motor imagery electroencephalogram (MI-EEG) signal -- Deep learning -- Convolutional neural network -- Serial-parallel (SP) structure -- Multi-dimensional feature extraction; 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.2021.103338
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Motor imagery electroencephalogram (MI-EEG) is one of the most important brain-computer interface (BCI) signal. It is vital to analyze the MI-EEG for the manipulation of external BCI actuator. However, traditional methods usually undertake EEG feature extraction and classification separately, which may lose efficient feature information. It behaves beyond our satisfaction for multi-class MI activity evoked by space-close and cannot eliminate the influence of individual differences. To solve these problems, we propose a convolutional neural network (CNN) with an end-to-end serial-parallel (SP) structure followed by tranfer learning. In detail, we use the serial module to extract the rough features in time–frequency-space domain, and the parallel module for fine feature learning in different scales. Meanwhile, a freeze-and-retrain fune tuning transfer learning strategy is proposed to improve the cross-subject accuracy. When our model is compared with the other three typical networks, results show that the proposed model performs best with the average testing accuracy of 72.13% and the average loss of 0.47, among which one subject only takes 0.7 s to reach 89.17% as the highest one. Through transfer learning, we reduce the training parameters by 53%. The average cross-subject classification accuracy increases by approximate 15%, and the individual highest accuracy reaches 76.98%. In conclusion, the integrity and separability of SPCNN determine that we require no additional EEG signal feature analysis, which is conducive to the realization of an efficient online BCI. It can also get rid of the dependence on training time and subject data to rapidly advance BCI in the future.