MSCANet: a cross-attention-based multi-scale convolutional fusion neural network for EEG motor imagery classification

Brain-computer interfaces (BCIs) utilize physiological neural signals from the brain to control external devices, offering significant potential to restore normal life in patients with brain injuries or motor impairments. However, electroencephalogram (EEG) signals are inherently non-stationary, possess low signal-to-noise ratios, and exhibit inter-subject variability, posing substantial decoding challenges. To effectively integrate multi-scale spatiotemporal features, this study proposes a cross-attention-based multi-scale convolutional fusion neural network (MSCANet) that integrates local and global features while capturing temporal dependencies across multiple scales. Specifically, MSCANet first employs a multi-scale spatio-temporal convolutional module to extract localized spatio-temporal information from variable-sized windows within individual frequency bands. Subsequently, channel and spatial attention mechanisms are incorporated to enhance discriminative feature representation by prioritizing salient information. A temporal convolution module with multi-level residual connections then preliminarily captures both short- and long-term dependencies. Finally, cross-attention mechanisms further capture temporal correlations and fuse features across frequency bands before classification via fully connected layers. In subject-dependent experiments, MSCANet achieved classification accuracies of 82.06% and 87.45%, with kappa values of 0.76 and 0.76 on the BCI IV-2a and BCI IV-2b Datasets, respectively. The proposed method outperforms several comparative models and demonstrates promising potential for BCI applications.

The online version contains supplementary material available at 10.1007/s11571-026-10485-5.