Classification of Finger Tapping Tasks using Convolutional Neural Network Based on Augmented Data with Deep Convolutional Generative Adversarial Network

Functional near-infrared spectroscopy (fNIRS) can help to diagnose specific diseases or distinguish motions for brain-computer interface (BCI). Also, repeating the same experiment can be uncomfortable for participants. It is difficult for researchers to obtain enough measurements to train the classification model sufficiently, which results in unstable classification accuracy. In this study, we investigated how to expand fNIRS data using the deep convolutional generative adversarial network (DCGAN) to improve classification accuracy and training stability. The data were measured using fNIRS during the finger tapping tasks, and then the proposed data augmentation method was used for generating artificial fNIRS datasets. These data were used to train a convolutional neural network (CNN), CNN model gave final classification accuracies for two classes (i. e. thumb finger tapping and little finger tapping). The AlexNet model, which is one famous model of the well-made model structure, was used as a CNN classifier model, the acquired accuracy for distinguishing the two types of tasks has been improved compared with the obtained accuracy using original data. This result suggests that the proposed deep learning model as a data generator is useful for improving classification performance.