基于DTI影像图论分析的正常老化认知表现预测模型

目的 以磁共振扩散张量影像（diffusion tensor imaging,DTI）为基础进行大脑结构网络拓扑属性分析,选择与认知表现分数相关性较大的结构网络特征,并基于这些特征建立认知表现分数预测模型,藉以客观地估测老年人的大脑认知能力.方法 对94例正常老化的DTI影像进行结构脑网络构建,采用图论法分析结构连接矩阵,提取结构网络的特征,并将所有特征与受试者的简单智能状态检查量表（mini-mental status examination,MMSE）分数进行相关性分析,选取出与大脑认知高度相关的网络特征,再基于这些特征建立5种分析模型,预测受试者的认知表现分数,以进一步分析模型的预测效能.结果 通过相关性分析,在相关系数大于0.22且P值小于0.05的条件下,选取出与大脑认知高度相关的30个特征,这些特征分布在AAL（automated anatomical labeling）图谱中的12个脑区.而在模型建立与效能分析部分,以高斯回归模型的效能最佳,其训练组相关系数达0.89,预测误差最小为2.01,对受试者的认知表现分数预测较准确.结论 利用结构脑网络度量指标作为生物标记指针可建立正常老化认知功能预测模型,且能有效预测正常老年人的认知表现分数.

A prediction model for cognitive performance from diffusion tensor imaging with graph theory in health ageing

Objective On the basis of diffusion tensor imaging （DTI）,we analyzed the properties of structural brain network by selecting the characteristics of structural network which were higher correlated with cognitive performance to estimate the prediction models.Based on the estimated models,the score of cognitive performance of subjects could be evaluated objectively according to their magnetic resonance imaging （MRI）.Methods We used diffusion tensor imaging to construct brain structural network and connection matrix for 94 healthy elders.In order to obtain the characteristics of structural brain network,we analyzed connection matrices by using graph theory and diffusion tractography.We selected significant features by the correlation analysis between the characteristics of brain network and subject＇ s mini-mental state examination （MMSE） score.These features would then be used to estimate the models based on five kinds of machine learning algorithm respectively to predict the cognitive performance.Finally,the performance of the five prediction models would be analyzed and discussed.Results Under the condition of the correlation coefficient greater than 0.22 and P value less than 0.05,thirty characteristics of brain network were selected as features and the related anatomical regions located in 12 brain areas according to the automated anatomical labeling （AAL） template.Among the 5 algorithms,Gaussian processes model was more accurate due to the higher correlation coefficient of 0.89 and the lower mean absolute error of 2.01.Conclusions We successfully established prediction model based on brain structural network metrics derived from DTI which could be employed to effectively predict subject＇ s cognitive performance score.