DeepReducer: A linear transformer-based model for MEG denoising

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Magnetoencephalography (MEG) measures event-related magnetic fields (ERFs) for cognitive and perceptual research.
• ERFs are often obscured by noise in single trials, necessitating lengthy data acquisition.
• Efficiently isolating ERFs is critical for advancing neuroscience and clinical applications.

Purpose of the Study:

• To introduce DeepReducer, a novel deep learning model for denoising MEG event-related magnetic fields.
• To reduce the number of trials required for reliable ERF analysis.
• To enhance the efficiency and practicality of MEG data acquisition.

Main Methods:

• Developed DeepReducer, a linear transformer-based deep learning model.
• Trained the model on a combination of limited-trial and multi-trial averaged ERFs.
• Utilized mean squared error as the loss function to capture signal fluctuations in MEG data.

Main Results:

• DeepReducer demonstrated superior performance compared to conventional trial-averaging methods.
• Significantly improved the signal-to-noise ratio (SNR) of event-related magnetic fields.
• Reduced source localization errors in both semi-synthetic and experimental MEG data.

Conclusions:

• DeepReducer offers a reliable and efficient method for denoising ERFs in MEG.
• The model optimizes MEG data acquisition, reducing participant burden and artifacts.
• Facilitates more accessible and less time-consuming neuroimaging studies.