EEG-CDILNet: a lightweight and accurate CNN network using circular dilated convolution for motor imagery...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) classification using electroencephalography (EEG) signals is crucial for brain-computer interfaces (BCIs).
• Deep learning models offer high accuracy but often require substantial computational resources, posing a challenge for practical applications.
• Balancing decoding performance and computational cost in deep learning for MI classification remains a significant research hurdle.

Purpose of the Study:

• To develop a novel, lightweight end-to-end convolutional neural network (CNN) for accurate motor imagery (MI) classification.
• To address the trade-off between classification accuracy and computational cost in deep learning-based MI analysis.
• To validate the proposed model's effectiveness on multiple public EEG datasets.

Main Methods:

• Proposed an end-to-end CNN model named EEG-circular dilated convolution (CDIL) network.
• Utilized depth-separable convolution to reduce parameters and extract spatio-temporal features from EEG signals.
• Employed circular dilated convolution (CDIL) for time-varying deep feature extraction and global average pooling for parameter reduction.

Main Results:

• Achieved average classification accuracies of 79.63% (BCIIV2a), 94.53% (HGD 4-class), and 87.82% (BCIIV2b).
• Demonstrated a superior balance between decoding performance and computational cost compared to other lightweight models.
• Ablation experiments and feature visualization confirmed the model's structural feasibility and effectiveness.

Conclusions:

• The proposed CNN model offers high MI classification accuracy with significantly reduced computing resources.
• The EEG-CDIL network presents a viable solution for practical applications in motor imagery classification research.
• This lightweight deep learning approach effectively addresses the challenge of computational cost in EEG-based BCIs.