Multi-Domain Dynamic Weighting Network for Motor Imagery Decoding

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) are standard for decoding electroencephalogram (EEG) signals in motor imagery (MI)-based brain-computer interfaces (BCIs).
• Existing CNNs face limitations in fully capturing intricate time-frequency features of EEG signals due to fixed kernel sizes and uniform feature attention.
• This necessitates advanced methods to improve the accuracy and robustness of MI-BCI decoding.

Purpose of the Study:

• To propose the Multi-Domain Dynamic Weighted Network (MD-DWNet) for enhanced MI-BCI decoding performance.
• To address the limitations of traditional CNNs in capturing complex EEG signal characteristics.
• To improve the adaptive modeling and generalization capabilities of BCI systems.

Main Methods:

• MD-DWNet integrates multimodal features across time, frequency, and spatial domains using a branch structure.
• It employs multi-band filtering, spatial convolution, and temporal variance for spatial-spectral feature extraction.
• A dual-scale CNN, dynamic global filter, attention mechanism, and dual-branch joint loss function are utilized for comprehensive feature processing and optimization.

Main Results:

• MD-DWNet achieved high classification accuracies on multiple datasets: 83.86% (BCI Competition IV 2a), 88.67% (IV 2b), 75.25% (OpenBMI), and 84.85% (laboratory dataset).
• The proposed network outperformed several advanced methods in MI signal decoding tasks.
• Experimental results validate the superior performance and effectiveness of MD-DWNet.

Conclusions:

• MD-DWNet significantly enhances the decoding performance of MI-based BCIs by effectively capturing complex EEG signal features.
• The network's multi-domain feature integration and adaptive mechanisms contribute to improved accuracy and generalization.
• The findings suggest MD-DWNet as a promising advancement for practical BCI applications.