FBDA-TSRNet: A Filter Bank Dual-Path Attention and Temporal–Spatial Residual Network for Motor Imagery Decoding

Abstract

Electroencephalography (EEG)-based motor imagery (MI) decoding is an important task in noninvasive brain-computer interfaces (BCIs). However, reliable MI-EEG decoding remains challenging because of low signal-to-noise ratios, inter-session and inter-subject variability, and multiscale spatiotemporal patterns. This study proposes the Filter Bank Dual-Path Attention and Temporal–Spatial Residual Network (FBDA-TSRNet), a compact framework for MI decoding. EEG trials are decomposed into narrow frequency bands using a Chebyshev Type II filter bank, and a sample-level frequency-band attention mechanism adaptively reweights informative sub-bands. The weighted multiband representation is then processed by a multiscale temporal–spatial residual block and a dual-path feature-channel-temporal attention module to capture cross-scale patterns and emphasize discriminative feature channels and task-relevant temporal responses. An attention-aware discriminative objective integrates center loss, band-attention entropy regularization, and temporal-attention smoothness regularization to improve feature compactness and stabilize attention learning. Experiments were conducted on three public datasets using subject-dependent 10-fold cross-validation and hold-out validation, with subject-independent leave-one-subject-out cross-validation additionally performed on BCI Competition IV 2a. On this dataset, FBDA-TSRNet achieved 79.95% accuracy under session-to-session hold-out validation, 69.8% under leave-one-subject-out cross-validation, and 85.84% under pooled 10-fold cross-validation. The model contains 14.2K trainable parameters, with measured network and filter-bank latencies of 7.2 and 0.8 ms/trial, respectively, on the reported hardware. Across the three datasets, FBDA-TSRNet showed consistently competitive mean performance under both cross-validation and hold-out validation. These results indicate that FBDA-TSRNet provides a compact solution for offline MI-EEG decoding across subject-dependent and subject-independent settings.