A Novel 3D Approach with a CNN and Swin Transformer for Decoding EEG-Based Motor Imagery Classification

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Motor imagery (MI) is vital for brain-computer interfaces (BCIs), enabling device control for individuals with motor impairments.
• Decoding electroencephalography (EEG) signals for MI-BCI is challenging due to high dimensionality and low signal-to-noise ratio (SNR).
• Existing deep learning methods struggle with capturing spatio-temporal dependencies in EEG data.

Purpose of the Study:

• To develop a novel end-to-end deep learning network for enhanced MI-BCI classification accuracy.
• To address limitations in processing high-dimensional, low-SNR EEG signals.
• To improve the capture of intrinsic correlations between electrode channels and long-distance temporal dependencies.

Main Methods:

• A novel network integrating Convolutional Neural Networks (CNNs) and a Swin Transformer was proposed.
• EEG signals were transformed into a 3D data structure for spatio-temporal feature extraction.
• 1D CNNs processed temporal data, 2D CNNs processed electrode distribution, and a 3D Swin Transformer module extracted deep features.

Main Results:

• The proposed method achieved 83.99% classification accuracy on the BCI Competition IV-2a dataset.
• This accuracy significantly outperformed existing deep learning methods.
• The combination of CNNs and Swin Transformer in a 3D space proved effective for EEG signal processing.

Conclusions:

• The developed CNN-Swin Transformer network offers a promising approach for improving MI-BCI performance.
• This method provides a new perspective for processing complex EEG signals in BCI applications.
• Further research could explore broader BCI tasks and clinical applications.