ADFCNN: Attention-Based Dual-Scale Fusion Convolutional Neural Network for Motor Imagery Brain-Computer Interface

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are effective for motor imagery (MI)-based brain-computer interfaces (BCIs).
• Single-scale CNNs struggle with extracting comprehensive spectral information from EEG.
• Existing multi-scale CNNs often fail to effectively fuse features from different scales.

Purpose of the Study:

• To propose an attention-based dual-scale fusion convolutional neural network (ADFCNN) for enhanced MI-BCI performance.
• To jointly extract and fuse EEG spectral and spatial information across multiple scales.
• To leverage self-attention for improved feature fusion.

Main Methods:

• Developed an ADFCNN integrating temporal and spatial convolutions at dual scales.
• Employed two different kernel sizes for temporal convolutions to capture μ and β rhythms.
• Utilized a self-attention mechanism for feature fusion based on internal similarity.

Main Results:

• Achieved superior subject-specific motor imagery recognition performance on BCI Competition IV datasets 2a, 2b, and OpenBMI.
• Demonstrated significant improvements in cross-session average classification accuracies (e.g., 9.14% on BCI-IV2a).
• Showcased significant improvements in within-session average classification accuracies (e.g., 10.89% on BCI-IV2a).

Conclusions:

• The proposed ADFCNN effectively extracts and fuses multi-scale spectral and spatial EEG features.
• Self-attention mechanism enhances feature fusion, leading to state-of-the-art performance in MI-BCI.
• Ablation studies and visualizations confirm the effectiveness of the dual-scale CNN and attention modules.