A U-Net-Based Multi-Scale Attention Neural Network for Removing EEG Artifacts

Electroencephalography (EEG) is a highly random and nonlinear time series signal, and it is easily affected by other physiological artifacts. The interference from various physiological artifacts is detrimental to subsequent EEG signal analysis. Therefore, denoising EEG signals is a crucial step in EEG analysis. In recent years, deep learning-based methods have shown superior performance compared to traditional methods in removing EEG artifacts. This paper proposes a U-Net-based multi-scale attention neural network to eliminate ocular and electromyographic artifacts in single-channel EEG signals. This network is divided into multiple modules based on different sizes of convolution kernels, utilizing our designed multi-scale gated convolution interaction module, allowing our model to focus on the reconstruction information of the contaminated segments. After passing through the multi-scale gated convolution interaction module, the data is downsampled by channel summation to enhance features. Then, through a U-Net-based convolution interaction module, which consists of several denoising modules connected in parallel at different scales and built on one-dimensional convolutions, we achieve the final reconstruction of the EEG signal through feature fusion and dimensionality reduction using a linear layer. Various experiments validate that our proposed network model can reconstruct clean EEG time series, achieving good performance on two quantitative metrics (relative root mean square error and correlation coefficient) compared to other representative models.