A Multi-Scale Adaptive Fusion Network: End-to-End Interpretable Small-Sample Classifier for Motor Imagery EEG
Qiulei Han, Yan Sun, Ze Song, Hua Ye, Ting-Wei Chen, Jian Ming Zhao · IEEE Access · 2025
Brain-computer interface (BCI) technology based on motor imagery (MI) represents an advanced human-computer interaction approach that enables us to communicate and exchange directly with the external environment through neural activities. However, the non-stationarity and individual variability of EEG signals present significant challenges to improving decoding accuracy. Existing studies often struggle with feature extraction, dynamic feature selection, and temporal modeling, failing to capture critical EEG patterns effectively. In this study, we propose a multi-scale adaptive fusion network (MSAFNet) to improve the accuracy of MI-EEG classification. The MSAFNet integrates two primary modules: scale-aware channel attention (SCA) and adaptive dynamic feature fusion (ADF). The SCA module strengthens the model’s ability to perceive features of different frequencies and spatio-temporal information through multi-scale feature interactions to extract richer primary feature representations. The ADF module captures short-term dynamic variations in EEG signals while performing key feature selection and cross-layer feature fusion. Additionally, long-term temporal dependence is modeled by a temporal convolution network (TCN) to enhance the extraction capability of temporal features of EEG signals. Ablation experiments and t-SNE visualization analysis verified the effectiveness of the modules. Gradient-weighted Class Activation Mapping generated EEG topography maps visualized the brain regions of interest of the model, and the experimental results show that the decision-making process of the model in the MI-EEG decoding task is in accordance with neurophysiological principles. The within-subject average classification accuracies on the BCI Competition IV-2a and IV-2b public datasets are 88.08% and 90.47%, respectively, and the subject-independent average classification accuracies reach 68.79% and 80.86%, achieving state-of-the-art decoding accuracies and providing new possibilities for practical applications of BCI.