A Multi-Branch 3D Convolutional Neural Network for EEG-Based Motor Imagery Classification

• Published in: IEEE transactions on neural systems and rehabilitation engineering Vol. 27; no. 10; pp. 2164 - 2177
• Main Authors: Zhao, Xinqiao, Zhang, Hongmiao, Zhu, Guilin, You, Fengxiang, Kuang, Shaolong, Sun, Lining
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.10.2019
• Subjects: 3D convolutional neural network (3D CNN); Deep learning; Electrodes; Electroencephalogram (EEG); Electroencephalography; Feature extraction; motor imagery (MI); multi-branch structure; Task analysis; Three-dimensional displays; Two dimensional displays
• ISSN: 1534-4320; 1558-0210
• DOI: 10.1109/TNSRE.2019.2938295

One of the challenges in motor imagery (MI) classification tasks is finding an easy-handled electroencephalogram (EEG) representation method which can preserve not only temporal features but also spatial ones. To fully utilize the features on various dimensions of EEG, a novel MI classification framework is first introduced in this paper, including a new 3D representation of EEG, a multi-branch 3D convolutional neural network (3D CNN) and the corresponding classification strategy. The 3D representation is generated by transforming EEG signals into a sequence of 2D array which preserves spatial distribution of sampling electrodes. The multi-branch 3D CNN and classification strategy are designed accordingly for the 3D representation. Experimental evaluation reveals that the proposed framework reaches state-of-the-art classification kappa value level and significantly outperforms other algorithms by 50% decrease in standard deviation of different subjects, which shows good performance and excellent robustness on different subjects. The framework also shows great performance with only nine sampling electrodes, which can significantly enhance its practicality. Moreover, the multi-branch structure exhibits its low latency and a strong ability in mitigating overfitting issues which often occur in MI classification because of the small training dataset.