. Electroencephalography (EEG) signals can reflect motor intention signals in the brain. In recent years, motor imagery (MI) based brain-computer interfaces (BCIs) have attracted the attention of neuroinformatics researchers. Numerous deep learning models have been developed to decode EEG signals. Although deep learning models, particularly those based on convolutional neural networks, have shown promise in decoding EEG signals, most existing methods focus on attention mechanisms while neglecting high-order statistical dependencies that are critical for accurately capturing the complex structure of EEG data.. To address this limitation, we propose a neural network integrating a transpose-attention mechanism and second-order pooling (SecTNet). The proposed model tackles two fundamental challenges in EEG decoding. It metrics the covariance structure of EEG signals using Riemannian geometry on symmetric positive definite (SPD) matrices, and it enhances the discriminability of these SPD features by introducing attention mechanisms that adaptively model inter-channel dependencies. Specifically, SecTNet is composed of three key components. First, a multi-scale spatial-temporal convolution module extracts detailed local features. Second, a transpose-attention mechanism captures dependency information from the internal interactions between channels. Lastly, a second-order pooling layer captures high-order statistical correlations in the EEG feature space.. SecTNet is evaluated on two publicly available EEG datasets, namely BCI competition IV 2a dataset and OpenBMI dataset. It achieves an average accuracy of 86.88% on the BCI competition IV dataset 2a and 74.99% on the OpenBMI dataset. Moreover, results show that SecTNet maintains competitive performance even when trained on only 50% of the data, demonstrating strong generalization under limited data conditions.. These results demonstrate the broad applicability and effectiveness of SecTNet in enhancing MI-BCI performance. SecTNet provides a robust and generalizable framework for EEG decoding, supporting the development of BCI applications across diverse real-world scenarios. Objective Approach Main results Significance
