Using Convolutional Neural Networks to Automatically Detect Eye-Blink Artifacts in Magnetoencephalography Without...

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Machine Learning

Background:

• Magnetoencephalography (MEG) records brain activity via magnetic fields but is susceptible to muscle artifacts, particularly eye-blinks.
• Current methods for artifact removal, like electrooculography (EOG), complicate setup and patient comfort.
• Existing Independent Component Analysis (ICA) methods for MEG lack automated artifact identification.

Purpose of the Study:

• To develop an electrooculography (EOG)-free, data-driven approach for identifying and removing eye-blink artifacts in MEG data.
• To implement a Convolutional Neural Network (CNN) for automated classification of eye-blink components derived from ICA.
• To validate the CNN's performance and interpret its learned features for artifact detection.

Main Methods:

• Applied Independent Component Analysis (ICA) to magnetoencephalography (MEG) data to separate neuronal and non-neuronal signals.
• Developed a 10-layer Convolutional Neural Network (CNN) to classify independent components as either eye-blink or non-eye-blink artifacts.
• Utilized attention mapping to visualize learned spatial features and validate the CNN's decision-making process.

Main Results:

• The CNN achieved high classification accuracy (99.67%), sensitivity (97.62%), specificity (99.77%), and ROC AUC (98.69%) on a test dataset of 30 subjects.
• Visualized spatial features learned by the CNN corresponded to expert-identified eye-blink artifact patterns.
• The proposed method successfully removed eye-blink artifacts without requiring additional electrooculography (EOG) electrodes.

Conclusions:

• The developed CNN-based method offers an effective and automated solution for removing eye-blink artifacts in MEG.
• This approach enhances the feasibility of fully automated MEG processing pipelines by addressing motion artifacts.
• The EOG-free method improves patient comfort and simplifies experimental procedures in MEG studies.