生成式对抗网络在抑郁症分类中的应用

APPLICATION OF GENERATIVE ADVERSARIAL NETWORKS IN DEPRESSION CLASSIFICATION

深度学习领域中的条件深度卷积生成式对抗网络(CDCGAN)是一种能够生成与训练数据同分布样本的生成模型。针对抑郁症f MRI(functional Magnetic Resonance Imaging)数据难采集、用于研究的被试数远小于数据特征维数的问题,首次将CDCGAN应用于生成抑郁症f MRI数据并提出一种混合特征选择方法用于分析f MRI数据。采用组独立成分分析提取41名被试的独立成分并构建全脑动态功能连接网络;通过肯德尔排序相关系数法选出具有较强辨别能力的特征并使用CDCGAN扩充数据;采用所提出的混合特征选择法进行特征选择;对41名被试的数据进行分类。实验结果表明,采用CDCGAN的分类正确率为92.68%,明显优于不应用CDCGAN的分类结果 68.29%,同时说明了抑郁症f MRI数据采用CDCGAN方法扩充数据的可行性以及混合特征选择方法能选出更有效的特征。

The conditional deep convolutional generative adversarial network（ CDCGAN） is a generative model that can generate samples with the same distribution of training data in the deep learning domain. For the problem that the f MRI data of depression is difficult to collect and the number of subjects used for the study is much smaller than the dimension of the data features,the CDCGAN was first applied to generate f MRI data for depression and a mixed feature selection method was proposed to analyse f MRI data. Firstly,the independent components of 41 subjects were extracted using group independent component analysis and a whole brain dynamic functional connection network was constructed.Secondly,the features with strong discrimination ability were selected by the Kendall sorting correlation coefficient method and CDCGAN was used to extend the data. Then,the hybrid feature selection method proposed in this paper was used for feature selection. Finally,the data of 41 subjects were classified. The experimental results showed that the classification accuracy rate of CDCGAN was 92. 68%,which was obviously better than the classification result of68. 29% without CDCGAN. At the same time,the feasibility of using the CDCGAN method to expand the data of depressive f MRI data and the hybrid feature selection method can be used to select more effective features. At the same time,it illustrated the feasibility of using the CDCGAN method to expand data for depressive f MRI data. Mixed featureselection methods selected more effective features.