FBMSNet

FBMSNet decodes motor imagery (MI) from electroencephalogram (EEG) data to improve classification accuracy for brain-computer interface (BCI) applications.

Key Features:

• Multiview Spectral Representation: Employs a filter bank to generate a multiview spectral representation of EEG signals, capturing diverse frequency components for MI classification.
• Mixed Depthwise Convolution: Uses mixed depthwise convolution to extract multiscale temporal features from EEG signals.
• Spatial Filtering: Incorporates spatial filtering to mitigate volume conduction and reduce interference from overlapping cortical sources.
• Joint Supervision Mechanism: Trains the network with joint supervision combining cross-entropy and center loss to increase interclass dispersion and intraclass compactness.

Scientific Applications:

• Motor Imagery Classification: Validated on the BCI Competition IV 2a and OpenBMI datasets, achieving four-class hold-out classification accuracy of 79.17% and two-class hold-out classification accuracy of 70.05%.
• Brain-Computer Interfaces (BCIs): Enhances EEG decoding performance for research and clinical BCI applications by improving robustness of MI classification.

Methodology:

End-to-end deep learning integrating a filter bank for spectral analysis, mixed depthwise convolution for temporal multiscale feature extraction, spatial filtering for noise reduction, and joint supervision combining cross-entropy and center loss.