一种基于卷积神经网络的重度抑郁症辅助诊断方法  

An auxiliary diagnosis method for major depression disorder based on convolutional neural network

作　　者：王茵 郑国威 颉瑞 杨琳 姚志军[1] 胡斌[1,2,3,5] Wang Yin;Zheng Guo-wei;Xie Rui;Yang Lin;Yao Zhi-jun;Hu Bin(Key Laboratory of Wearable Equipment of Gansu Province,College of Information Science and Engineering,Lanzhou University,Lanzhou 730000,China;Joint Research Center for Cognitive Neurosensor Technology of Lanzhou University&Institute of Semiconductors,Lanzhou University,Lanzhou 730000,China;Engineering Research Center of Open Source Software and Real-Time System of the Ministry of Education,Lanzhou University,Lanzhou 730000,China;The Third People's Hospital of Tianshui,Tianshui 741000,Gansu,China;Chinese Academy of Sciences Center for Excellence in Brain Science and Intelligence Technology,Shanghai Institutes for Biological Sciences,Shanghai 200030,China)

机构地区：[1]兰州大学信息科学与工程学院,甘肃省可穿戴装备重点实验室,甘肃兰州730000 [2]兰州大学认知神经传感器技术与中国科学院半导体研究所联合研究中心,甘肃兰州730000 [3]兰州大学开源软件与实时系统教育部工程研究中心,甘肃兰州730000 [4]甘肃省天水市第三人民医院心理科,甘肃天水741000 [5]中国科学院上海生命科学研究所,中国科学院脑科学与智能技术卓越中心,上海200030

出　　处：《兰州大学学报（医学版）》2022年第8期5-10,共6页Journal of Lanzhou University（Medical Sciences）

基　　金：国家重点研发计划资助项目(2019YFA0706200);国家自然科学基金资助项目(61632014,61627808,U21A20520);甘肃省自然科学基金资助项目(20JR5RA292)。

摘　　要：目的针对多站点抑郁症数据分类的泛化能力不强,以及使用三维原始图像作为深度学习分类模型的输入容易过拟合的问题,设计了一种卷积神经网络架构用于重度抑郁症(MDD)辅助诊断。方法该模型基于静息态功能磁共振成像得到的低维功能连接矩阵作为输入,从中提取功能相关信息和高阶抽象特征从而分类。结果将该模型在多站点REST-meta-MDD数据集上验证,分类准确率为70.39%。结论通过遮挡分析描述了不同大脑区域对MDD辅助诊断的贡献,结果表明默认模式网络、视觉网络和额顶控制网络对MDD分类任务具有重要作用。Objective The generalization ability of multi-site depression data classification is not adequately strong,and the use of the 3D raw image as an input of deep learning classification model is prone to overfitting problems.Therefore,a convolutional neural network architecture was here proposed,which was designed for MDD auxiliary diagnosis.Methods In the model,the low-dimensional functional connection matrices based on resting-state fMRI were used as an input,from which functional information and higher-order abstract features were extracted for classification.Result The model was validated on multi-site REST-metaMDD datasets,and the classification accuracy was 70.39%.Besides,the contribution of different brain regions to MDD auxiliary diagnosis was described.Conclusion The results indicated that the default mode network,visual network and fronto-parietal control network could play an important role in MDD classification tasks.

关 键 词：抑郁症 功能磁共振成像 深度学习 分类 

分 类 号：R749.99[医药卫生—神经病学与精神病学]