A cross-regional multi-scale network for motor imagery EEG decoding with discriminative representation learning

Motor imagery (MI)-based brain–computer interfaces (BCIs) have shown great potential for neurorehabilitation and human–machine interaction. However, robust decoding of MI electroencephalography (EEG) remains challenging due to its low signal-to-noise ratio, non-stationarity, and inter-subject variability. This work presents CrossDMNet, an end-to-end MI-EEG decoder that jointly improves representation richness, cross-regional modeling, and feature discriminability within a unified architecture. CrossDMNet builds multi-resolution temporal descriptors and refines them through a lightweight global-context modulation, allowing scale-specific responses to be adaptively emphasized according to trial-level context. To preserve region-dependent neurophysiological patterns, the model learns two parallel streams from central and parietal electrodes and employs a learnable cross-stitch exchange to share complementary information while retaining region-specific characteristics. In addition, CrossDMNet introduces region-wise supervision with two dedicated classifier heads and aggregates their outputs via a joint decision module, reducing the reliance on a single global boundary. Finally, a dual discriminative objective that combines center loss and triplet loss is adopted to produce compact and well-separated embeddings. Experiments on BCI Competition IV-2a, HGD, OpenBMI, and GigaDB, together with systematic ablations, verify the effectiveness of each component and show that CrossDMNet achieves consistently superior performance compared with competitive deep MI-EEG baselines.