基于双范式功能磁共振成像的毕生常模在年龄和认知能力上的预测

Norm atlas of lifespan for age and cognitive ability prediction based on dual-paradigm fMRI

在世界毕生功能磁共振成像公开数据集中,如何评估整体人口的大脑老化程度和认知能力是研究焦点.以往的相关研究大多依赖于对静息态功能磁共振成像中不同指标的统计分析,如功能连接、功能连接密度等.本文提出基于双范式功能磁共振成像的毕生常模图谱,旨在反映被试于静息态和自然刺激范式间大脑神经活动的异步、同步.随后,设计一种基于局部感受野模块的模型,预测被试的年龄及认知评分.首先,将功能连接输入到基于变分自编码器的生成器中,构建常模.其次,利用常模对比计算每个被试的毕生功能连接模式,以形成特征.最后,将这些特征以五折交叉验证的方式拆分训练集、测试集,并输入到所提出的预测模型中,以预测年龄、流体智力和反应时评分.该方法在剑桥老龄化神经科学中心数据集(样本量:424;年龄:18~88岁)中进行测试,年龄、流体智力和反应时评分的绝对误差分别为5.97、3.72和0.108,优于其他相关研究.结果表明,所提出的常模图谱可以评估个人的认知能力,而结合所提出的方法能够降低预测年龄及认知评分的误差,这将有助于揭示被试在整个生命周期的大脑神经发育的演化模式.

With extensive attention to human brain health worldwide,researchers are devoted to discovering the development and aging mechanisms of individual brains to assess the degree of brain aging and cognitive ability.The rapid development of functional Magnetic Resonance Imaging(fMRI)has provided robust technical support for neuroscience research.Different measures(functional connectivity and functional connectivity density,etc.)from resting-state fMRI have been widely used in related works.However,these studies do not provide sufficient coverage of the entire lifespan brain change owing to the utilization of a single diagram.Therefore,we propose a norm atlas of lifespan for age and cognitive ability prediction based on dual-paradigm fMRI,which combines unsupervised and supervised learning.Specifically,functional connectivity derived from resting-state and natural stimulation diagrams are integrated simultaneously,aiming at responding to the synchronous and asynchronous states of brain neural activity of both paradigms.The whole workflow is divided into the following steps.Firstly,the functional connectivity data of both paradigms are fed into a variational auto-encoder(VAE)as the input of the encoder and the output of the decoder,respectively.Then,the asynchronous and synchronous states of brain neural activity in the resting-state and natural stimulation paradigm can be revived by the latent variables(norm).Secondly,the functional connectivity patterns for each subject are calculated using the lifespan norm as a feature representation for the downstream prediction tasks.Thirdly,all subjects are split into training and testing sets using five-fold cross-validation,and the corresponding norm features of each subject are fed into the prediction module based on the local receptive field block.To achieve the learning and fusion of features,the prediction module combines the mechanisms of multi-scale learning and residual connection.Finally,the prediction of age and cognitive ability scores(e.g.,fluid intelligence,reaction time scores)is performed.The proposed method has been validated on the Cambridge Center for Neuroscience of Aging dataset(sample size:424,age range:18 to 88 years).The mean absolute errors for age,fluid intelligence,and reaction time scores are 5.97,3.72,and 0.108,respectively,which are superior to other related studies.Moreover,the trend of fluctuations in prediction errors across lifespan is consistent with cognitive behavior scores.To the best of our knowledge,it is the first time for the discovery of such a trend revealed by a data-driven approach.The results indicate that the combination of norm atlas and local receptive field module can make predictions for age and cognitive ability at high accuracy,emphasizing feature modeling’s effectivity of proposed norm atlas.The complex parameter optimization for VAE and prediction module are not performed,instead,empirical settings are adopted as much as possible.This strategy not only highlights the generalization of the proposed method,but also explains the norms from a biomedical aspect,and explores the“black-box”effect of deep learning.The proposed method is benefit to reveal the patterns of brain neurodevelopment throughout lifespan and has the potential for clinical applications.