CNN - Time Frequency Representation Based Brain Wave Decoding from Magnetoencephalography Signals

Abstract

Understanding the concurrent activity of human brain is a highly crucial task in most of the brain computer interface (BCI) applications. This study exploited the potential of empirical wavelet transform and the different time frequency visualizations for interpretating the brain functionality for external visual stimuli from magnetoencephalography signals. The study examined the four types of visualizations: spectrogram, scalogram, constant Q Gabor spectrogram and Fourier synchro squeezed representation. The proficiency of the aforementioned representations of the empirical wavelet transform (EWT) decomposed modes were assessed using GoogLeNet, a prominent transfer learning architecture. The experimental results serve as an evident that mode 3 of EWT is a dominant mode and that combined with scalogram results in a promising performance with a classification accuracy of 80.79% in decoding the human brain for visual stimuli.