A Deep Neural Network Classifier for Decoding Human Brain Activity Based on Magnetoencephalography

Magnetoencephalography(MEG) is an emerging medical signal processing methodology that uses the magnetic field of brain to decode internal brain activity. However, MEG signals are very complicated and usually corrupted with significant amount of noise. Therefore, it is not easy to directly understand how the human brain responds to visual stimulus by analysing the MEG signals without utilizing advanced signal processing techniques such as feature extraction and classification. The feature extraction of MEG signals can be accomplished by applying the Riemannian approach. Moreover, the extracted features can be classified by classification algorithms such as SVM and KNN to complete the decoding process. However, these  classification methods don't produce satisfying results as the number of features is very high. In this paper, the classification problem of MEG signals is addressed and a deep neural network based classifier is proposed to classify the MEG signals that were produced as the brain output for two different types of visual stimuli. The visual stimuli comprise a data base of faces and scrambled faces. Our experimental results demonstrate that the proposed classifier exhibits superior classification performance over the other competing methods used in the paper.