Related Experiment Video
Updated: Oct 28, 2025
08:23
A Multimodal Imaging- and Stimulation-based Method of Evaluating Connectivity-related Brain Excitability in Patients with Epilepsy
Published on: November 13, 2016
11.4K
MEGnet: Automatic ICA-based artifact removal for MEG using spatiotemporal convolutional neural networks.
Alex H Treacher1, Prabhat Garg2, Elizabeth Davenport3
1Lyda Hill Department of Bioinformatics, UT Southwestern Medical Center, Dallas, TX, United States.
Neuroimage
|July 18, 2021
Summary
This study introduces an automated method using deep learning to detect and remove non-neuronal artifacts in magnetoencephalography (MEG) data without electrooculography (EOG) or electrocardiography (ECG) sensors. The approach achieves high accuracy, simplifying MEG processing for research and clinical applications.
Area of Science:
- Neuroimaging
- Biomedical Engineering
- Machine Learning
Background:
- Magnetoencephalography (MEG) records neuronal activity via magnetic fields.
- Non-neuronal artifacts, such as eye-blinks, saccades, and cardiac activity, contaminate MEG data.
- Current artifact detection often relies on electrooculography (EOG) and electrocardiography (ECG), complicating setup and patient comfort.
Purpose of the Study:
- To develop an Electrooculography (EOG)- and Electrocardiography (ECG)-free approach for automated artifact detection and suppression in MEG data.
- To improve the efficiency and accessibility of MEG data processing for both clinical and research settings.
Main Methods:
- Utilized a data-driven, multivariate decomposition approach based on Independent Component Analysis (ICA).
- Developed a highly accurate artifact classifier using a hybrid deep learning model combining 1-D and 2-D Convolutional Neural Networks (CNNs).
- Optimized the CNN architecture via an unbiased, computer-based hyperparameter random search and employed visualization methods to interpret model features.
Main Results:
- Achieved state-of-the-art artifact detection accuracy of 98.95% on a dataset of 217 subjects.
- Demonstrated high sensitivity (96.74%) and specificity (99.34%) for identifying eye-blink, saccade, and cardiac artifacts.
- Validated the model on both resting-state and task-based MEG data, showing robustness and generalizability.
Conclusions:
- The proposed EOG- and ECG-free method effectively automates the identification and removal of common non-neuronal artifacts in MEG.
- This approach enhances MEG data processing, reduces setup complexity, and improves patient comfort.
- The findings support the broader adoption of automated artifact suppression in clinical and research neuroimaging.
