📃 A Novel End-to-end Network Based on a bidirectional GRU and a Self-Attention Mechanism for Denoising of Elctroencephalography Signals
🔍 개발 과정 및 논문에 대한 정보 (2023/7/4 ~ 2023/8/9)
👪 CNA
- 양한열
- 지우근
Deep learning networks have been increasingly attracting attention in many fields. Recently, the application of deep learning models has been brought to the field of electroencephalography denoising, and has provided performance that is comparable to that of traditional techniques. Howerver, the lack of well-structured, standardized dataset with benchmark limits the development of deep learning solutions for EEG denoising. Therefore, we present EEGdenoiseNet, a benchmark dataset, that is suited for training and testing deep learning-based EEG denoising models, as well as for comparing the performance across different models. Our EEGdenoiseNet dataset contains 4514 clean EEG epochs with ground truth for model training and testing. EEGdenoiseNet also offers a set of benchmarks generated by evaluating the performance of four classical deep learning networks (a fully-connected network, a simple convolution network, a complex convolution network and a recurrent neural network). Our benchmark dataset would hopefully accelerate the development of the emerging field of deep learning-based EEG denoising.
For more information, the paper of this dataset is publicly available on arXiv(https://arxiv.org/abs/2009.11662).
Single-Channel-EEG-Denoise tool box could be find in Github (https://github.com/ncclabsustech/Single-Channel-EEG-Denoise)
Our laboratory also proposed an deep learning framework to separate neural single and artifacts in the embedding space and reconstruct the denoised signal, which is called DeepSeparator. Could be find in Github(https://github.com/ncclabsustech/DeepSeparator)
(1) BG-Attention algorithm is the first time the RNN and the self-attention network have been used to denoise EEG signals.
(2) The network is an end-to-end structure. Without preprocessing or feature extraction, the incoming EEG data is processed directly.