SSA-DCNet: a cross-session MI-EEG classification network based on deformable convolution and spatial-shift attention

Brain-computer interfaces (BCIs) based on motor imagery (MI) electroencephalogram (EEG) signals have shown tremendous potential in neurorehabilitation due to their non-invasive acquisition and ease of use. However, the cross-session nature of EEG signals-where recordings from the same subject at different sessions may vary due to fluctuations in physiological state and environmental conditions-presents a significant challenge. Efficient extraction and preservation of temporal and spatial features from EEG signals can capture invariant neural activation patterns while suppressing session-dependent noise and variability, thereby greatly enhancing the robustness of cross‑session motor imagery classification. To address the suboptimal performance of existing models in cross-session MI-EEG classification, this paper proposes Spatial-Shift Attention Deformable Convolution Network-SSA-DCNet, a compact convolutional neural network in which temporal filtering is implemented via a two-dimensional deformable convolution of size 1 × 64, so that the sampling grid dynamically adapts to the non-uniform distributions of informative EEG segments while operating on a 1 × 64 kernel along the temporal axis. Thereafter, a spatial-shift attention architecture expands each intermediate feature map from C to 3 C channels, evenly splits them into three subsets, applies distinct spatial-shift operations to each subset, and finally merges them via a split-attention that recalibrates channel weights to emphasize spatial patterns stable across sessions. On the public BCI Competition IV-2a and 2b datasets, SSA-DCNet achieved classification accuracies of 84.72% and 90.45%, respectively. Moreover, t-SNE visualizations provide intuitive evidence, underscoring its superior discriminative power and robust cross-session generalization.