MS-TSEFNet: Multi-Scale Spatiotemporal Efficient Feature Fusion Network.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Motor imagery (MI) decoding using electroencephalogram (EEG) is crucial for brain-computer interfaces (BCIs).
• Current deep learning models struggle to effectively fuse multi-level features in complex EEG signals, limiting classification performance.
• There is a need for advanced models that can capture spatiotemporal dynamics and integrate information across different feature levels.

Purpose of the Study:

• To propose a novel deep learning network, the Multi-scale Spatiotemporal Efficient Feature Fusion Network (MS-TSEFNet), for enhanced motor imagery decoding from EEG signals.
• To improve the fusion of multi-level features extracted from EEG data.
• To enhance the accuracy and robustness of EEG-based BCIs.

Main Methods:

• Developed MS-TSEFNet incorporating multi-scale convolution modules to capture temporal dynamics at various time scales.
• Integrated a spatial attention mechanism to effectively identify spatial correlations between EEG electrodes.
• Employed an efficient feature fusion strategy to deeply integrate features from different levels, enhancing model expressiveness.

Main Results:

• MS-TSEFNet achieved high classification accuracies on public datasets: 80.31% (BCIC-IV2a), 86.69% (BCIC-IV2b), and 71.14% (ECUST).
• The proposed network demonstrated superior performance compared to current state-of-the-art algorithms.
• Ablation studies confirmed the significant contribution of each module, particularly the multi-scale convolution and feature fusion modules, to overall performance.

Conclusions:

• MS-TSEFNet effectively decodes motor imagery signals by leveraging multi-scale spatiotemporal feature extraction and fusion.
• The network offers improved accuracy and robustness for EEG-based brain-computer interfaces.
• The findings highlight the importance of advanced feature fusion techniques for complex EEG signal processing.