Brain Network Analysis Of Magnetic Resonance Images For Classification Of Early Alzheimer's Disease

Alzheimer's disease(AD)is a neurodegenerative disease.Because of its irreversibility and huge social burden,modeling the early stages of AD for timely intervention will ease this burden.Magnetic resonance imaging is a safe,high spatial and temporal resolution imaging.Currently,it has been widely used in AD diagnosis.However,manual diagnosis is often a subjective,time-consuming and labor-intensive process.Also,the imaging characteristics of the early stages of AD are not obvious,and human eyes cannot detect it.To solve this problem,the machine learning method is an effective way.However,existing research still has many limitations.For example,the existing model fails to consider the dynamic characteristics of the brain network,the demographic information,and the multimodal complementarity.Hence,the dynamic high-order network model and graph convolution network model are studied based on machine learning techniques,respectively.The Alzheimer's Disease Neuroimaging Initiative(ADNI)database collects subjects at different stages,including late mild cognitive impairment(LMCI),early mild cognitive impairment(EMCI),significant memory concern(SMC),and normal control(NC).In this thesis,the ADNI database is used for experimental verification.Overall,the main research contents are as follows:(1)In order to accurately classify the early stages of AD,this study considers the dynamic characteristics of the brain,analyzes the dynamic and high-order properties of the resting-state functional magnetic resonance imaging(R-fMRI)scanning time series,and improves the functional expression of brain function networks.Firstly,the sliding window is used to divide the time series to obtain the dynamic functional networks,and the covariance matrix between the dynamic functional networks is constructed to obtain the high-order networks.Then,the weight local clustering method is used to extract the effective features of the network,and then the Least Absolute Shrinkage and Selection Operation(LASSO)method is used to select the discriminative features.Finally,the leave-one-out cross-validation method is used for support vector machine classification model training.We have achieved the six different classification tasks of LMCI vs.NC,EMCI vs.NC,SMC vs.NC,LMCI vs.EMCI,LMCI vs.SMC and EMCI vs.SMC,and the classification average accuracy is 85.2%,80.3%,78.9%,78.8%,84.3% and 80.2%,respectively.(2)In order to further improve the accurate diagnosis of the early stages of AD,we introduce a novel method based on graph convolution network model,which combines with population phenotype information to learn discriminative features from multimodal neuroimaging.Firstly,we construct the functional connectivity networks using the R-fMRI time series,and the structural connectivity networks using diffusion tensor imaging.Then,the network features are combined with the collection device information and the subject gender information to construct a population phenotypic feature graph.Finally,the feature graph is learned by using the graph convolutional network model.The method obtains an average accuracy of 79.4%,85.6% and 86.7% for the classification tasks of EMCI vs.NC,LMCI vs.NC and LMCI vs.EMCI.In summary,the research work is mainly based on magnetic resonance imaging brain network(including structural brain network and functional brain network)for learning model in early classification of AD.The main contents include the construction of dynamic high-order functional network model and the construction of graph convolutional network model,and verify the validity of the model on ADNI database.