Transformer-Based Multiscale 3-D Convolutional Network for Motor Imagery Classification

Due to the variability and nonstationarity of electroencephalography (EEG) signals across different recording scenarios and subjects, it is crucial to have methods with strong generalization capabilities that can effectively capture both temporal and spatial features while maintaining high accuracy. In this article, we introduce a Transformer-based multiscale 3-D-convolution network (TMCNet), a novel end-to-end deep learning model that integrates multiscale 3-D convolutional neural networks with the Transformer for enhanced feature extraction. First, the temporal convolution is applied to EEG signals to extract detailed temporal features. Then, we utilize multiscale 3-D convolutional branches to perform spatial convolution, capturing spatial information across diverse receptive fields. In the second stage, we utilize the multihead attention mechanism in the Transformer to extract more refined global features, with fully connected layers used for classification. We evaluate the proposed method on three publicly available datasets. In cases where the datasets contain information for two sessions, we conduct evaluations for both within-session and cross-session scenarios. The experimental results demonstrate that the proposed TMCNet exhibits advanced performance, showcasing strong decoding ability and robustness.