Enhancing EEG Signal Reconstruction in Cross-Domain Adaptation Using CycleGAN

This paper explores the applicability of CycleGAN, to reconstruct as well as improve brain-computer interface (BCI) signals. The experimental protocol used in the study involves a real-world table tennis dataset that includes multiple-channel high-density electroencephalography (EEG) data and structural magnetic resonance imaging (MRI) scans from individuals. The overall purpose is to develop a system that would allow enhancing the low-quality or partially damaged EEG signals by using information obtained from MRI, which is represented by Continuous Wavelet Transform (CWT). This technique solves the problem of converting MRI data into effective EEG signals using CycleGAN. The approach may present some practical benefits in cases where signal quality becomes a problem, for example, in wireless or portable BCI equipment. Data preprocessing involves channel filtering of the data, segmentation, and conversion of the data to continuous wavelet transform format to be used by the model. This paper compares three models of CycleGAN, focusing on the various losses and using identity loss to improve them.