A deep learning model for single-channel EEG artifact removal.
Code for the model in the paper DTP-Net: Learning to Reconstruct EEG signals in Time-Frequency Domain by Multi-scale Feature Reuse.
The architecture of DTP-Net:
The following setup has been used to reproduce this work:
- CUDA toolkit 10.1 and CuDNN 7.6.0
- Python == 3.7
- Tensorflow == 2.1.0
- Matplotlib == 3.5.3
- Numpy == 1.19.5
- Scipy == 1.4.1
We evaluate our DTP-Net with EEGDenoiseNet, Semi-Simulated EEG/EOG and MNE M/EEG dataset.
For EEGDenoiseNet, the dataset is publicly availabel as referenced in the article EEGdenoiseNet: A benchmark dataset for end-to-end deep learning solutions of EEG denoising.
For Semi-Simulated EEG/EOG, the dataset is publicly availabel as referenced in the article Adaptive Single-Channel EEG Artifact Removal With Applications to Clinical Monitoring.
For MNE M/EEG dataset, the dataset is using the sample dataset provided in the article EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces
For the sake of convenience, EEGDenoiseNet dataset and Semi-Simulated EEG/EOG dataset are temporarily placed in the G-node database and you can run the following scripts to download datasets.
gin get SummerBae/EEG-Denoise_Database
MNE M/EEG dataset would be download automatically when running 03-MNE_MEG_EEG_Dataset/erp.py.
Run this script to train a DTP-Net for model deployment I:
cd 01-EEGDenoiseNet_Dataset
python t1_dtpnet.py --noise_type EOG
noise_type denotes different artifacts source, including EMG, EOG and EMG_EOG.
The pre-trained models are saved in G-node database and you can run the following scripts to download checkpoints.
gin get SummerBae/EEG-Denoise_Database
Download the checkpints and save the checkpoints files SummerBae/EEG-Denoise_Database/01-EEGDenoiseNet_checkpoints/checkpoints in EEG-Denoise/01-EEGDenoiseNet_Dataset/.
Run this script to evaluate the denoise performance at different SNR levels of pre-trained models:
cd 01-EEGDenoiseNet_Dataset
python evaluation.py --noise_type EOG --plot_type metric_report
Run this script to get waveform results and PSD results for eliminating artifacts of pre-trained models:
cd 01-EEGDenoiseNet_Dataset
python evaluation.py --noise_type EOG --plot_type denoise_visualize
Run this script to train a DTP-Net for model deployment II:
cd 02-Semi_Simulated_Dataset
python t2_dtpnet.py
The pre-trained models are saved in G-node database and you can run the following scripts to download checkpoints.
gin get SummerBae/EEG-Denoise_Database
Download the checkpints and save the checkpoints files SummerBae/EEG-Denoise_Database/02-Semi_Simulated_checkpoints/checkpoints in EEG-Denoise/02-Semi_Simulated_Dataset/.
Run this script to evaluate the denoise performance of pre-trained models:
cd 02-Semi_Simulated_Dataset
python evaluation.py --plot_type metric_report
Run this script to get waveform results and PSD results for eliminating artifacts of pre-trained models:
cd 02-Semi_Simulated_Dataset
python evaluation.py --plot_type denoise_visualize
Run this script to get multi-channel waveform results for eliminating artifacts of pre-trained models:
cd 02-Semi_Simulated_Dataset
python evaluation.py --plot_type denoise_visualize_multi_channel
Run this script to get the ERP classification performance of raw data:
cd 03-MNE_MEG_EEG_Dataset
python erp.py
Run this script to get the ERP classification performance of data denoised by DTP-Net:
cd 03-MNE_MEG_EEG_Dataset
python erp.py --denoise
@article{Pei2024,
title = {DTP-Net: Learning to Reconstruct EEG Signals in Time-Frequency Domain by Multi-scale Feature Reuse},
ISSN = {2168-2208},
url = {http://dx.doi.org/10.1109/JBHI.2024.3358917},
DOI = {10.1109/jbhi.2024.3358917},
journal = {IEEE Journal of Biomedical and Health Informatics},
publisher = {Institute of Electrical and Electronics Engineers (IEEE)},
author = {Pei, Yan and Xu, Jiahui and Chen, Qianhao and Wang, Chenhao and Yu, Feng and Zhang, Lisan and Luo, Wei},
year = {2024},
pages = {1–12}
}
For academic and non-commercial use only.