DESAnet: dual-branch spatial-temporal CNN with efficient channel and sparse attention networks for motor imagery classification

Motor imagery brain-computer interfaces (MI-BCIs) play a vital role in neural interaction and clinical rehabilitation. However, their real-world deployment is often hindered by two major limitations: the relatively low decoding accuracy of MI tasks and the high computational cost of deep learning models. To address these challenges, we propose DESAnet—a lightweight end-to-end framework that incorporates a dual-branch spatial–temporal convolution module (DBSTCV), an efficient channel attention (ECA) module, and a sparse attention (SA) module. The dual branches capture multi-scale spatial and temporal structure in EEG, ECA adaptively emphasizes informative channels, and SA efficiently models long-range temporal dependencies with reduced overhead. We evaluate DESAnet on BCI Competition IV-2a (four-class) and IV-2b (two-class) under a subject-dependent setting. The model attains an average accuracy of 85.69% on 2a and 89.40% on 2b, outperforming representative deep-learning-based methods reported under comparable conditions. In addition, DESAnet maintains a low parameter count, compact model size, and few multiply-accumulate operations per forward pass, demonstrating favorable computational efficiency for resource-constrained or real-time MI-BCI use. Visualization with t-distributed stochastic neighbor embedding (t-SNE), attention weight maps, and class activation mapping (CAM) indicates clearer class separation and temporally focused responses, supporting the model’s interpretability and design choices. DESAnet can effectively improve the classification performance of MI-BCI signals while ensuring computational efficiency. This study provides a potentially promising technical solution for advancing the practical application of MI-BCI, and may hold valuable reference significance for the development of medical rehabilitation and intelligent interaction systems in relevant scenarios.