摘要
复杂网络分析与机器学习方法相结合的阿尔茨海默病辅助诊断研究受到了越来越多的关注,其通常采用脑功能网络的方法来描述大脑活动的信息.然而,现有的成果大多基于时域信号匹配构建脑功能网络,忽略了脑活动信息在各个频段下的差异.因此,本文提出了脑网络多频融合图核的阿尔茨海默病诊断方法.首先,将功能磁共振成像产生的图像通过小波变换的方法进行分频段处理;其次,分别计算得到的各频段图像中任意两个脑区间的互信息,并设定阈值与互信息值进行比较进而构造出多频脑网络模型;然后,基于此提出面向多频脑网络模型的融合图核;最后,基于多频融合图核、采用核极限学习机在ADNI(Alzheimer’s Disease Neuroimaging Initiative)公开数据库中获取的一组数据以及在OASIS(Open Access Series of Imaging Studies)公开数据库上获取的一组数据进行阿尔茨海默病的诊断.同时,还通过实验验证了不同参数设置对诊断结果的影响.两组数据集的实验结果表明,提出的多频融合图核的辅助诊断方法能够取得最佳性能,且该方法的辅助诊断准确率在两种数据集上比对比方法的最好结果分别提高了13.79%和15.29%.
In recent years,Alzheimer’s disease aided diagnosis research combining complex network analysis and machine learning method has received more and more attention.Usually,brain functional networks are used to describe the information of brain activity,and they are becoming one of the most important ways to diagnose some mental diseases.However,in the existing solutions,brain networks are mostly constructed based on signal matching in the time domain,which ignores the differences of brain activity information at various frequency bands.Furthermore,machine learning method cannot be applied to the data in the form of graphs.Based on it,we define the similarity between two graphs by the method of graph kernel,and then realize the classification by using the graph kernel for machine learning.Aiming at dealing with the problems,we propose a diagnosis method for multi-frequency fused graph kernel of brain network,which is used to classify the patients with Alzheimer’s disease and the normal people combined with the machine learning method.This method not only retains the multi-band information of brain activity,but also considers the unique topological characteristics of the brain network itself.Specifically,the resting state functional magnetic resonance imaging is first processed to different frequency bands by the wavelet transform technique,in order to obtain multiple frequency bands and construct the brain networks and calculate the graph kernels in different frequency bands of the image.In this way,it can express the activity information reflected by the brain at different frequencies.Secondly,the mutual information values between any two brain regions in each frequency band are calculated and the multi-frequency brain networks are obtained by comparing mutual information values and thresholds in different frequency bands.Then,the brain networks in all frequency bands together form a multi-frequency brain functional network model.After constructing the multi-frequency brain network model,the multi-frequency fuse
作者
汪新蕾
王之琼
王中阳
信俊昌
谷峪
WANG Xin-Lei;WANG Zhi-Qiong;WANG Zhong-Yang;XIN Jun-Chang;GU Yu)(School of Computer Science and Engineering,Northeastern University,Shenyang 110169;College of Medicine and Biological Information Engineering,Northeastern University,Shenyang 110169;Neusoft Institute of Intelligent Healthcare Technology,Co.Ltd.,Shenyang 110000)
出处
《计算机学报》
EI
CSCD
北大核心
2020年第1期64-77,共14页
Chinese Journal of Computers
基金
国家自然科学基金（61472069,61402089,61872070）
中央高校基本科研业务费（N180101028,N180408019,N171605001,N161602003,N160601001）
中国博士后科学基金（2019T120216,2018M641705）
十三五装备预研中国电科联合基金和沈阳东软智能医疗科技研究院有限公司开放课题基金（NRIHTOP1802）资助~~
关键词
阿尔茨海默病
功能磁共振成像
脑功能网络
多频融合图核
核极限学习机
Alzheimer’s disease
functional magnetic resonance imaging
brain functional network
multi-frequency fused graph kernel
kernel-based extreme learning machine
