CT-DCENet: Deep EEG Denoising via CNN-Transformer-Based Dual-Stage Collaborative Ensemble Learning

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Electroencephalogram (EEG) artifact removal is crucial for accurate signal analysis.
• Existing denoising methods struggle with complex, mixed artifacts and temporal dependencies.
• A need exists for advanced techniques that can reconstruct artifact-free EEG without prior artifact knowledge.

Purpose of the Study:

• To propose a novel CNN-Transformer-based dual-stage collaborative ensemble learning framework (CT-DCENet) for robust EEG artifact removal.
• To enhance the reconstruction of clean EEG signals, preserving fine-grained temporal details.
• To improve the reliability of EEG data for subsequent analytical tasks.

Main Methods:

• Developed CT-DCENet, a three-module framework: randomized collaboration, linear ensemble, and information complementation.
• Employed a dual-stage training approach to successively learn morphological and detailed characteristics of artifact-free EEG.
• Utilized a CNN-Transformer architecture for feature extraction and denoising in the information complementation module.

Main Results:

• CT-DCENet significantly outperformed state-of-the-art methods (DuoCL, GCTNet) in removing mixed EMG, ECG, and EOG artifacts.
• Achieved improvements in Signal-to-Noise Ratio (SNR) by 0.79 dB and Peak-to-Peak Correlation (PCC) by 0.6%, with a 1.9% decrease in Root Mean Square Error (RMSE).
• Reconstructed EEG signals closely matched clean EEG, accurately capturing peak amplitudes, high-frequency components, and waveform boundaries.

Conclusions:

• CT-DCENet offers a superior solution for EEG artifact removal, particularly for complex artifact scenarios.
• The framework's ability to reconstruct detailed EEG characteristics enhances data quality for downstream analysis.
• This advanced denoising technique holds significant promise for improving the utility of EEG in various research and clinical applications.