IC-U-Net

<iframe width="560" height="315" src="https://www.youtube.com/embed/0tHadL3kRjc" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>

Introduction

This is the Computational NeuroErgonomics x NeuroEducation (CNElab) on EEG artifacts removal end-to-end process: IC-U-Net: A U-Net based Denoising Autoencoders using Mixtures of Independent Components for Automatic EEG Artifact Removal, written in Pytorch. The aim of this project is to

• A novel EEG artifact removal method, IC-U-Net, is proposed.
• IC-U-Net is built based on the U-Net architecture with a loss function ensemble.
• IC-U-Net is trained using mixtures of EEG sources decomposed by independent component analysis.
• IC-U-Net does not require parameter tuning and can simultaneously remove various types of artifacts.
• IC-U-Net can facilitate the extraction of relatively more brain sources from noisy EEG signals.

Requirements

• Python == 3.6
• Pytorch == 1.6.0
• numpy >= 1.19.2

Evaluated Dataset

We evaluate the model with lane-keeping drive data collected and walking experiment from scientific data and mygooglecloud.

Ethics approval All participants completed informed consent forms after receiving a complete explanation of the study. The Institutional Review Board of Taipei Veterans General Hospital approved the study protocol.

Consent to participate All of the participants provided written informed consent prior to participation. The consent regarding publishing their data as a scientific report was also included.

Prerequisites

Below we assume the working directory is the repository root.

Install dependencies

• Using pipenv (recommended)

  Make sure pipenv is installed. (If not, simply run pip install pipenv.)

  # Install the dependencies
  pipenv install
  # Activate the virtual environment
  pipenv shell

• Using pip

  # Install the dependencies
  pip install -r requirements.txt

(Sim) Prepare training

The training data can be download from ./UNet1D_sim/datalog/train/1-51log.csv Each data was randomly generated from the frequency(1-51Hz), amplitude(0-1), phase(0-2pi)

# Generate the raw data in 
# (function) dataRestore(name)
# you can run the code
python train.py

(Real) Prepare training

Real EEG training data could not be open right now. But you can prepare for yourself with 30 channel and 1024 points (4sec, 256 sample rate)

# you can run the code
python train.py

Scripts

Use pretrained models

1. Download pretrained models

   # Download the pretrained models
   ./UNet1D_real/final_RealEEG_5/modelsave/BEST_checkpoint.pth.tar
   #or
   ./UNet1D_real/final_RealEEG_5/modelsave/checkpoint.pth.tar

2. You can either perform inference from a trained model:

   # Run inference from a pretrained model
   (writing)

   or perform interpolation from a trained model:

   # Run interpolation from a pretrained model
   (writing)

Sample Results

(writing)

Papers

Under review on NeuroImage

Acknowledgement

This work was supported by the Ministry of Science and Technology, Taiwan (project numbers: MOST 110-2636-E-007-018 and 109-2636-E-007-022), and by the Research Center for Education and Mind Sciences, National Tsing Hua University. No funding source had involved in any of the research procedures.