【摘要】 目前基于深度学习的精神分裂症脑网络的研究，大多忽略了高阶脑功能网络对精神分裂症的影响和存在获取的样本量较少的问题，为了解决这些问题，提出一种基于高阶拓扑持续图像变分自编码器生成对抗网络预测模型，该模型结合了VAE和GAN的优点，生成高质量的样本并有效捕捉数据的潜在分布，提取数据的高阶拓扑特征，在高维空间中捕捉信号的复杂结构.其中VAE对高阶拓扑持续图像的分布进行建模，GAN中采用GAT和LSTM结合捕捉空间特征和时间特征.在103例精分患者和92例健康被试的精神分裂症数据集上进行实验，结果表明，与现有模型相比，所提出的模型在识别时空特征和功能脑网络方面表现优秀.在精神分裂症脑电信号的五个频段分析中，Gamma频段和Theta频段模型的高阶特征性能最佳，准确率分别达到96.1%和95.7%，均优于所选的对比方法 .为精神分裂症的早期诊断和预测提供了一种新的研究思路，具有重要的临床参考价值.更多还原

【Abstract】 Most of the current studies on deep learning–based brain networks for schizophrenia have neglected the influence of high–order brain function networks on schizophrenia and the existence of the problem of obtaining a small sample size,in order to solve these problems,a predictive model based on high – order persistence imagevariational autoencoder generative adversarial network is proposed.The model combines the advantages of VAE and GAN to generate high–quality samples and efficiently capture the underlying distribution of the data,extract high–order topological features of the data,and capture the complex structure of the signal in a high–dimensional space.Where VAE models the distribution of high–order topologically persistence images,a combination of GAT and LSTM is used in GAN to capture spatial and temporal features.Experiments on a schizophrenia dataset of 103 seminal patients and 92 healthy subjects show that the proposed model performs excellent in recognizing spatiotemporal features and functional brain networks compared to existing models.Among the five frequency bands analyzed for schizophrenia EEG signals,the Gamma and Theta band models had the best performance in terms of high–order features,with accuracies of 96.1% and 95.7%,respectively,which were superior to those of the selected comparison methods.It provides a new research idea for the early diagnosis and prediction of schizophrenia,which has important clinical reference value.更多还原