EEGDnet: Fusing Non-Local and Local Self-Similarity for 1-D EEG Signal Denoising with 2-D Transformer

TL;DR

This work tackles denoising of 1-D EEG signals corrupted by ocular and muscle artifacts for non-invasive BCI. It introduces EEGDnet, a 2-D transformer-based denoising network that fuses non-local self-similarity in self-attention with local self-similarity in feed-forward blocks by reshaping 1-D EEG into 2-D patches and applying a multi-depth transformer with residual connections. Evaluated on a dataset with extensive artifact types, EEGDnet achieves state-of-the-art denoising performance with lower parameter count and FLOPs, demonstrating robustness across SNRs and superior time- and frequency-domain metrics $\text{RRMSE}_{temporal}$, $\text{RRMSE}_{spectral}$, and $CC$. The approach is extensible to other 1-D signals and presents practical advantages for embedded BCI devices due to reduced computation and memory requirements.

Abstract

Electroencephalogram (EEG) has shown a useful approach to produce a brain-computer interface (BCI). One-dimensional (1-D) EEG signal is yet easily disturbed by certain artifacts (a.k.a. noise) due to the high temporal resolution. Thus, it is crucial to remove the noise in received EEG signal. Recently, deep learning-based EEG signal denoising approaches have achieved impressive performance compared with traditional ones. It is well known that the characteristics of self-similarity (including non-local and local ones) of data (e.g., natural images and time-domain signals) are widely leveraged for denoising. However, existing deep learning-based EEG signal denoising methods ignore either the non-local self-similarity (e.g., 1-D convolutional neural network) or local one (e.g., fully connected network and recurrent neural network). To address this issue, we propose a novel 1-D EEG signal denoising network with 2-D transformer, namely EEGDnet. Specifically, we comprehensively take into account the non-local and local self-similarity of EEG signal through the transformer module. By fusing non-local self-similarity in self-attention blocks and local self-similarity in feed forward blocks, the negative impact caused by noises and outliers can be reduced significantly. Extensive experiments show that, compared with other state-of-the-art models, EEGDnet achieves much better performance in terms of both quantitative and qualitative metrics.

Figures (3)

• Figure 1: The architecture of the training phase of the EEG denoising model based on the EEGDnet model. First, the noise-free EEG signal and the artifact (ocular artifact or muscle artifact) are pixel-wise summed according to a certain signal-to-noise ratio. The result, namely noisy EEG signal, and corresponding noise-free one form the input and the label of EEGDnet, respectively. Inside the EEGDnet, the signal is transformed into a 2-D matrix through the reshape layer at the beginning. The opposition is true at the end of the network. And then sequentially through the self-attention block, normalization layer, feed forward block and normalization layer. Note that there are residual connections in the network. $N Depths$ indicates the number of self-attention blocks as well as other components. Finally, these features that can distinguish EEG signal and noise are automatically reconstructed by minimizing the objective function (MSE) to get the denoised EEG signal.
• Figure 2: Visualized examples of denoising results with different state-of-the-art methods. a) Denoising on Ocular Artifact b) Denoising on Muscle Artifact. For each sub-figure, an easily-observed window (red box and corresponding zoomed-in view) is selected for better comparison. Note that the amplitude is normalized and the time domain sampling rate is 256 SPS. Regarding two examples, we can observe that the denoising results of our proposed EEGDnet are closer to the ground-truth signals. Please zoom in for better view.
• Figure : EEGDnet