IFNet: An Interactive Frequency Convolutional Neural Network for Enhancing Motor Imagery Decoding From EEG.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) decoding for electroencephalogram (EEG)-based Brain-Computer Interfaces (BCIs) requires efficient feature extraction from spectral, spatial, and temporal domains.
• Limited, noisy, and non-stationary EEG data pose challenges for advanced decoding algorithms.

Purpose of the Study:

• To propose a lightweight Interactive Frequency Convolutional Neural Network (IFNet) for MI decoding.
• To explore cross-frequency interactions for enhanced representation of MI characteristics.
• To improve the accuracy and efficiency of EEG-based BCI systems.

Main Methods:

• IFNet extracts spectro-spatial features from low and high-frequency bands separately.
• Cross-frequency interactions are learned via element-wise addition and temporal average pooling.
• Repeated trial augmentation is used as a regularizer to achieve spectro-spatio-temporally robust features.

Main Results:

• IFNet significantly outperforms state-of-the-art MI decoding algorithms on benchmark datasets (BCIC-IV-2a and OpenBMI).
• Achieved an 11% improvement over the previous best result on the BCIC-IV-2a dataset.
• Demonstrated an optimal trade-off between decoding speed and accuracy, capturing cross-frequency coupling and MI signatures.

Conclusions:

• The proposed IFNet effectively decodes motor imagery from EEG signals.
• IFNet demonstrates superior performance and robustness compared to existing MI decoding methods.
• IFNet shows significant promise for applications requiring rapid and accurate control in MI-BCI systems.