BSAN: A Self-Adapted Motor Imagery Decoding Framework Based on Contextual Information.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) decoding faces challenges in capturing contextual brain activity and handling cross-session variability.
• Existing brain-computer interface (BCI) systems often require extensive recalibration due to session-specific feature shifts.

Purpose of the Study:

• To develop an innovative network, the Bi-Stream Adaptation Network (BSAN), to improve the robustness of MI-based BCIs across different sessions.
• To enhance the extraction of contextual information from brain regions during MI tasks.
• To mitigate cross-session variance in neural feature distributions.

Main Methods:

• The proposed Bi-Stream Adaptation Network (BSAN) incorporates a Bi-attention module for multi-scale convolutional analysis of MI context.
• A Bi-discriminator is employed post-feature extraction to align features across different MI sessions, ensuring session-invariance.
• The framework integrates context coherence and session-invariance for effective neural pattern representation.

Main Results:

• BSAN achieved average accuracies of 78.97% and 83.79% on two public motor imagery datasets.
• The network demonstrated a fast inference time of 2.99 ms on CPU-only devices.
• The approach effectively fused contextual information and achieved session-invariance, reducing the need for redundant MI trials.

Conclusions:

• The Bi-Stream Adaptation Network (BSAN) offers a robust solution for MI-BCI, effectively addressing challenges in contextual information extraction and cross-session variability.
• BSAN shows significant potential for accelerating the practical application and deployment of motor imagery-based brain-computer interfaces.