BiModule-ECA Network: A Dual-Module Fusion Strategy for Robust EEG Denoising Using Efficient Channel Attention

Electroencephalography (EEG) signals are critical in clinical and research applications but are often contaminated by artifacts such as electrooculography (EOG) and electromyography (EMG). Traditional convolutional neural networks (CNNs) typically excel at removing either EOG or EMG noise but struggle to handle both effectively. To address this, we propose the BiModule-ECA Network (BME-Net), a deep learning architecture designed to denoise both EOG and EMG artifacts in EEG signals. BME-Net combines residual and non-residual convolutional blocks with Efficient Channel Attention (ECA) mechanisms, effectively addressing both stable EOG and variable EMG noise. We trained and evaluated BME-Net on the EEGdenoiseNet dataset, demonstrating that it outperforms six existing deep-learning-based denoising methods across various signal-to-noise ratio (SNR) levels. BME-Net achieved superior Relative Root Mean Square Error (RRMSE) and Correlation Coefficient (CC) metrics for both EOG and EMG noise removal, highlighting its effectiveness and potential for real-time EEG applications requiring reliable denoising.