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

Simultaneous functional MRI and neurophysiology recordings provide insights into the relationship between neural activity and hemodynamic responses. However, gradient switching and radiofrequency (RF) pulse transmission induce large artifacts in neurophysiology signals, severely masking neural activity. Existing artifact removal methods, such as average artifact subtraction (AAS) and principal component analysis (PCA)-based approaches, result in significant residual artifacts and potential signal loss. In this study, we propose a Generative Adversarial Network (GAN) based blind source separation model to remove gradient and RF artifacts without requiring ground truth denoised data. The model incorporates identity loss to preserve neural signals, while GAN loss and frequency loss constrain the denoising process in both time and frequency domains. We validated our approach using both simulated and empirical data. Results demonstrate that our method effectively removes artifacts while maintaining neural signal integrity, outperforming existing approaches.