MSAttNet: Multi-scale attention convolutional neural network for motor imagery classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Convolutional neural networks (CNNs) are standard for motor imagery (MI) classification.
• Small, noisy, and non-stationary EEG datasets pose challenges for CNN-based decoding algorithms.

Purpose of the Study:

• To introduce a novel method, MSAttNet, for enhanced feature extraction from limited MI-EEG data.
• To improve the performance of MI classification algorithms on small datasets.

Main Methods:

• MSAttNet integrates multi-band segmentation, attention spatial convolution, and multi-scale temporal convolution modules.
• A filter bank enhances frequency domain features, while an attention mechanism adaptively adjusts convolutional kernels.
• Bilinear pooling extracts temporal features and eliminates noise for classification.

Main Results:

• MSAttNet achieved high accuracies across four diverse MI-EEG datasets (BCI Competition IV IIa, IIb, OpenBMI, ECUST-MI).
• Cross-session accuracies ranged from 75.94% to 84.52%, demonstrating robust performance.
• The method outperformed existing state-of-the-art algorithms in MI decoding.

Conclusions:

• MSAttNet effectively addresses the challenges posed by small MI-EEG datasets.
• The proposed network improves decoding performance through robust feature extraction.
• This advancement holds promise for more effective brain-computer interfaces.