MSVTNet: Multi-Scale Vision Transformer Neural Network for EEG-Based Motor Imagery Decoding.

Area of Science:

• Neuroscience and Artificial Intelligence
• Brain-Computer Interfaces (BCI)

Background:

• Transformer networks are used for electroencephalography (EEG) decoding in motor imagery (MI).
• Existing methods often overlook cross-frequency coupling and effective integration of diverse neural network architectures.
• Advanced decoding algorithms require improved feature extraction and network integration.

Purpose of the Study:

• To propose a novel end-to-end Multi-Scale Vision Transformer Neural Network (MSVTNet) for MI-EEG classification.
• To address limitations in capturing cross-frequency coupling and integrating CNNs with Transformers.
• To enhance the discriminative power of feature embeddings for improved MI decoding.

Main Methods:

• MSVTNet utilizes Convolutional Neural Networks (CNNs) to extract local spatio-temporal features at multiple filtered scales.
• Features are concatenated to form multi-scale spatio-temporal tokens, processed by Transformers for cross-scale and global temporal information.
• An auxiliary branch loss is employed for intermediate supervision, ensuring effective CNN-Transformer integration.

Main Results:

• MSVTNet demonstrated state-of-the-art performance across subject-dependent and subject-independent experiments.
• Evaluations were conducted on the BCI competition IV 2a, 2b, and OpenBMI datasets.
• The proposed network achieved superior results in all tested MI decoding scenarios.

Conclusions:

• MSVTNet exhibits significant superiority and robustness in enhancing MI decoding performance.
• The network effectively captures crucial cross-frequency coupling and global temporal correlations.
• This approach offers a promising advancement for sophisticated EEG-based BCI applications.