A Multi-Domain Convolutional Neural Network for EEG-Based Motor Imagery Decoding.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) decoding is vital for electroencephalography (EEG)-based brain-computer interface (BCI) advancement.
• Complex deep learning models often lead to overfitting and inaccurate MI decoding due to redundant information.
• Existing methods struggle to fully leverage multi-domain EEG features effectively.

Purpose of the Study:

• To propose an efficient Multi-Domain Temporal-Spatial-Frequency Convolutional Neural Network (TSFCNet) for enhanced MI decoding.
• To address limitations of complex deep learning structures in MI decoding by utilizing multi-domain EEG features.
• To develop a network capable of powerful feature extraction without intricate architectures.

Main Methods:

• The TSFCNet utilizes MixConv-Residual blocks for multiscale temporal feature extraction from multi-band filtered EEG data.
• A temporal-spatial-frequency convolution block extracts discriminative representations from spatial, frequency, and time-frequency domains.
• Features are aggregated using average pooling and variance layers, trained with cross-entropy and center loss.

Main Results:

• The TSFCNet achieved superior classification accuracy and kappa values compared to state-of-the-art models.
• Results on BCI competition IV 2a dataset: 82.72% accuracy and 0.7695 kappa.
• Results on BCI competition IV 2b dataset: 86.39% accuracy and 0.7324 kappa.

Conclusions:

• The proposed TSFCNet demonstrates significant potential for improving MI decoding performance in BCIs.
• The network's ability to integrate multi-domain EEG features offers a promising alternative to complex deep learning models.
• TSFCNet provides a powerful yet efficient approach for enhancing BCI applications through accurate motor imagery decoding.