A novel multi-scale convolutional neural network for motor imagery classification. (July 2021)

• Record Type: Journal Article
• Authors: Riyad, Mouad; Khalil, Mohammed; Adib, Abdellah
• Is Part Of: Biomedical signal processing and control. Volume 68(2021)
• Publication Date: 2021-07
• Subjects: Deep learning -- Electroencephalography -- Brain–computer interfaces -- Motor imagery -- Convolutional neural network; 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.2021.102747
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We propose AMSI-EEGNet as an novel end-to-end Convolutional Neural Network for MI. The proposed method uses multi-scale inputs to extract the robust features. AMSI-EEGNet achieves higher accuracies than the state-of-the-art. Ablation study and relevance analysis prove the benefit of the proposed method. Abstract: Brain-Computer Interfaces (BCI) allows us to use brain activity as a pathway to interact with machines for varied purposes that may enhance humans life. Most of those systems rely on motor imagery that generates different patterns of the electroencephalography signals. However, those phenomenon produce signals with low quality which makes their recognition a troublesome job. Thus, we come up with a novel Convolutional Neural Network (CNN) that is constructed specially for this application. A multi-branch system is used to preserve and deal with each band separably, which will grant a better classification. By using the dataset BCI IV-2a, we demonstrate that our method achieves better training time compared with the existing CNNs, and better classification compared with the state-of-the-art baselines. We conduct an ablation study to evaluate the impact of the parallel pipelines on the performances by adding them gradually, and we show that adding features extracted from different scales has indeed a positive influence on diverse metrics. Further, we compared the training time of the network with and without Separable and Depthwise convolutional layers, we conclude that they permit the reduction of the training time thanks to their low computational requirement along with the positive impact on the accuracy. We note that there is some benefit in the tasks that provides the weakest patterns. To further clarify the results, we carry out an extensive analysis of the network based on the relevance of the input feature, which explained the failed classification.