Deep Learning Based Functional Connectivity Mining Methods And Its Applications In Brain Aging

With the aggravation of the aging population,the health of the elderly is not only the focus of social attention,but also the research direction of neuroscience,clinical medicine and computer science.Functional magnetic resonance image(fMRI),as a noninvasive brain imaging technique,has been widely used to explore the changes of the brain function during aging.However,the current fMRI research aiming at mining functional changes in brain mainly apply the traditional statistical analysis and rely on relatively imited sample size,which may lead to insufficient analysis of brain aging,thus ignoring some important basic information about changes of brain function.Considering the advantages of deep learning and big data analysis compared with traditional methods,the combination of deep learning and big data in neuroimaging is expected to further explore the information of brain aging which help us understand the mechanism of brain aging.In this dissertation,new mining method and predictive model are developed based on deep learning and big data for fMRI data,and these methods are applied to study the brain mechanism in the process of healthy aging.Moreover,the analysis of local functional connectivity across lifespan reveals and infers the possible compensation mechanism of basal ganglia network(BGN)in brain aging.The work of this dissertation is summarized as following:(1)A new method of functional connectivity analysis was proposed to study the functional connectivity of brain agingMost of the present studies primarily rely on traditional statistic methods with relatively small samples,which may result in insufficient analysis of functional connectivity during brain aging,thus ignoring some important underlying information regarding functional network alterations.Aiming at mining functional connectivity sufficiently,a new analysis method,Deep neural network with Autoencoder pretrained functional connectivity analysis(DAFA),is proposed to deeply mine important functional connectivity changes in fMRI during brain aging.DAFA fully considers the potential complex relationship in the data,and combined with autoencoder and deep neural network,it can comprehensively detect the complex information of functional connectivity changes in normal aging.The results showed that compared with young group,decreased connections within cingulo-opercular network(CON),increased connections between default mode network(DMN)and cerebellum network and decreased connections between CON and occipital network were found in the elderly.We speculate that there may be two neural circuits corresponding to the loss of functional coordination and compensation mechanism in brain aging: 1)the circuit with decreased functional connectivity is located between CON and occipital\cerebellar networks;2)the circuit with increased functional connectivity is located between DMN and CON\occipital \cerebellar networks.We believe that this method has a good application prospect and is expected to mine more important information of functional connectivity changes in the study of Alzheimer's disease and schizophrenia.(2)A new predictive method and feature visualization strategy were proposed for fluid intelligence prediction and feature explainability related to brain aging.Based on autoencoder and deep neural network,a new predictive model of fluid intelligence is established by using the local functional connectivity density(lFCD)and the four-dimensional(spatial temporal)consistency of local neural(FOCA).This method takes advantage of ensemble learning and deep neural network predicting fluid intelligence from local functional connectivity.Compared with the connectome predictive model,the most popular and advanced predictive model in neuroscience,the Pearson's correlation between the predicted Cattell scores of the proposed method and the real ones has increased 10%.Furthermore,a spatial autoencoder is proposed.Moreover,the visualization and interpretation of extracted features are explored by combining with the Dropout in deep neural network training which improve the interpretability of the results.This feature visualization strategy can effectively reflect the spatial pattern of brain functional activity which could contribute to predicting fluid intelligence from lFCD and FOCA,thus demonstrating the effectiveness of the proposed spatial autoencoder in fMRI feature extraction and the feature visualization strategy.(3)The compensatory mechanism of basal ganglia network during brain aging was studied from the perspective of local functional connectivity.It is generally believed that the decrease of functional connectivity in functional networks during brain aging may be due to the extensive weakening of cognitive function,while the increase of functional connectivity between functional networks may be due to the compensation mechanism in functional networks However,the inherent mechanism of functional network changes in lifespan remains unclear,especially how this potential compensation mechanism works in lifespan.At the end of this dissertation,the lFCD and FOCA in restingstate were used to preliminarily study the changes of local functional connectivity in lifespan.We found that with the increase of age,the functional connectivity of visual network,sensorimotor network and default mode network decreased,while the functional connectivity of basal ganglia network increased.Thus,we speculate that the basal ganglia network may play a regulatory role in the aging process to compensate for the dysfunction and damage of other functional systems.