Xu Fangzhou, Rong Fenqi, Leng Jiancai, Sun Tao, Zhang Yang, Siddharth Siddharth, Jung Tzyy-Ping
IEEE Trans Neural Syst Rehabil Eng. 2021;29:2417-2424. doi: 10.1109/TNSRE.2021.3123969. Epub 2021 Nov 25.
Acquiring Electroencephalography (EEG) data is often time-consuming, laborious, and costly, posing practical challenges to train powerful but data-demanding deep learning models. This study proposes a surrogate EEG data-generation system based on cycle-consistent adversarial networks (CycleGAN) that can expand the number of training data. This study used EEG2Image based on a modified S-transform (MST) to convert EEG data into EEG-topography. This method retains the frequency-domain characteristics and spatial information of the EEG signals. Then, the CycleGAN is used to learn and generate motor-imagery EEG data of stroke patients. From the visual inspection, there is no difference between the EEG topographies of the generated and original EEG data collected from the stroke patients. Finally, we used convolutional neural networks (CNN) to evaluate and analyze the generated EEG data. The experimental results show that the generated data effectively improved the classification accuracy.
获取脑电图（EEG）数据通常既耗时、费力又成本高昂，这给训练强大但对数据要求较高的深度学习模型带来了实际挑战。本研究提出了一种基于循环一致对抗网络（CycleGAN）的替代EEG数据生成系统，该系统可以扩充训练数据的数量。本研究使用基于改进型S变换（MST）的EEG2Image将EEG数据转换为脑电地形图。该方法保留了EEG信号的频域特征和空间信息。然后，使用CycleGAN学习并生成中风患者的运动想象EEG数据。通过目视检查，生成的EEG地形图与从中风患者收集的原始EEG数据的地形图之间没有差异。最后，我们使用卷积神经网络（CNN）对生成的EEG数据进行评估和分析。实验结果表明，生成的数据有效地提高了分类准确率。
