RLANET: An EEG denoising network for judgemental removal of long- and short-term distribution artefacts

Abstract

The acquisition of electroencephalogram (EEG) signals is susceptible to contamination by various physiological artefacts, making the subsequent analysis of the EEG signals difficult. Deep learning (DL)-based methods of EEG denoising have achieved some effectiveness in addressing this issue. However, existing structural designs do not fully account for the randomness and waveform diversity of artefacts’ temporal distribution. Most network structures are processed directly on the samples disturbed by the mixture of artefacts, which overlooks the time-varying overlap between electromyography (EMG) and EEG in long-term distribution, as well as the mutual interference between long-and short-term artefacts. To overcome these issues, we propose an EEG denoising network for the judgemental removal of long- and short-term distribution artefacts. This network, which we call RLANET, consists of a segmentation network, a short-term denoising network, and a long-term denoising network. The segmentation network ResUNet is used to enable discrimination of the temporal distribution characteristics of the artefacts. The short-term denoising network LWTCN learns temporal correlations and capture local waveform variations to remove short-term distribution artefacts from EEG signals. The long-term denoising network ADDPM is used to reconstruct EEG signals affected by long-term distribution artefacts, improving the quality of noise removal. The experimental results demonstrate that RLANET’s denoising performance is significantly superior to that of current mainstream denoising methods. Specifically, in the removal of mixed artefacts, RLANET achieved improvements of 1.31% and 1.5316 in Correlation Coefficient (CC) and Signal-to-Noise Ratio (SNR), respectively, demonstrating its outstanding performance in handling mixed artefacts.