Deep Learning Classification of two-class Motor Imagery EEG signals using Transfer Learning

Abstract

Motor imagery (MI) based Brain-Computer Interface (BCI) system uses Electroencephalography (EEG) signals recorded over the scalp during imagination of motor movements to control a computer or mobility device. Such systems require a method to classify the acquired MI EEG signals into commands. In this study, three pre-trained Convolutional Neural Networks (CNN) models- AlexNet, ResNet50 and InceptionV3 are studied for the classification of Left-hand and Right-hand MI EEG signals. BCI Competition IV dataset 2a and acquired MI EEG dataset of nine healthy subjects are used to study the classification performance. Classification results show that transfer learning using InceptionV3 model produces the highest classification accuracy of $82.78 \pm 4.87$% for the BCI competition dataset and $83.79 \pm 3.49$% for the acquired dataset compared to AlexNet and ResNet50. Hence, InceptionV3 CNN can be used to efficiently classify MI signals in BCI systems to aide people suffering from neuromuscular disorders by replacing or restoring motor functions.