TF-MCA: A Dual-Branch Time-Frequency Network with Multi-Scale Cross-Axis Attention for Generalizable EEG Decoding

Electroencephalography (EEG) decoding remains challenging due to low signal-to-noise ratios, non-stationary dynamics, and high inter-subject variability. Existing methods operate predominantly in either the temporal or spectral domain, failing to exploit complementary cross-domain information. We propose the Time-Frequency Multi-Scale Cross-Axis Attention Network (TF-MCA), a dual-branch framework for generalizable EEG decoding. The temporal branch employs a novel Multi-Scale Cross-Axis Attention (MCA) module capturing spatiotemporal dependencies via bidirectional cross-attention between time-axis and channel-axis features, augmented by multi-kernel axial convolutions with kernel sizes {7, 11, 21}. The frequency branch leverages HiLoAttention to separately encode high-frequency local spectral details and low-frequency global spectral structures. A lightweight DilateFormer fusion module integrates cross-domain features through multi-scale dilated attention at$\mathbf{O}(n)$complexity. Experiments on three benchmarks demonstrate consistent state-of-the-art performance: BCIC-IV2a (69.95% accuracy), MAMEM-SSVEP-II (56.36% accuracy), and BCI-ERN (76.82% AUC). Ablation studies confirm the complementary contribution of each module.