Using Generative Adversarial Networks to eliminate RF and gradient interference in neurophysiology signals recorded...

Area of Science:

• Neuroimaging
• Neurophysiology
• Biomedical Engineering

Background:

• Simultaneous functional MRI (fMRI) and neurophysiology recordings are crucial for understanding brain function.
• Gradient switching and radiofrequency (RF) pulses in fMRI introduce significant artifacts into neurophysiology signals.
• Current artifact removal methods like average artifact subtraction (AAS) and principal component analysis (PCA) often leave residual artifacts and cause signal loss.

Purpose of the Study:

• To develop a novel blind source separation model for removing gradient and RF artifacts from simultaneous fMRI-neurophysiology data.
• To improve the quality of neurophysiology signals recorded during fMRI acquisition.
• To overcome limitations of existing artifact removal techniques.

Main Methods:

• Proposed a Generative Adversarial Network (GAN) based blind source separation model.
• Incorporated identity loss to preserve neural signal integrity.
• Utilized GAN loss and frequency loss for time and frequency domain denoising.

Main Results:

• The GAN model effectively removed gradient and RF artifacts from both simulated and empirical data.
• Neural signal integrity was maintained, outperforming traditional AAS and PCA methods.
• Demonstrated successful blind source separation without requiring ground truth denoised data.

Conclusions:

• The proposed GAN-based method offers a superior approach for artifact removal in simultaneous fMRI-neurophysiology recordings.
• This technique enhances the reliability of neurophysiological data acquired during fMRI.
• The method holds significant potential for advancing research into neural activity and hemodynamic responses.