Early Diagnosis Of Alzheimer’s Disease Based On Functional Magnetic Resonance Imaging

Alzheimer’s disease（AD）is one of irreversible neurodegenerative diseases of the brain,the prevalence of which increases significantly with age.It is considered the seventh leading cause of death.In the absence of clear and effective biomarkers,the progression of AD can only be delayed through early intervention.Functional magnetic resonance imaging（fMRI）is an important medical tool for studying brain diseases,which can observe functional changes in related regions of the brain.This may be a potential biomarker of AD.It plays an important role in the auxiliary diagnosis and early intervention of AD by capturing this biomarker.Nonimaging differences between individuals,such as gender,age,education,etc.,also affect the prevalence of AD.How to integrate imaging information and non-imaging information is also an important challenge for computer-aided AD diagnosis.Due to the powerful data learning ability of deep learning,it has been favored by more and more researchers and has been widely used in the medical field.Therefore,this thesis will use deep learning technology to integrate fMRI image information and phenotype information to classify AD and normal controls（NC）.Then,this thesis will identify biomarkers associated with AD.The main research contents of this thesis are as follows:（1）At present,most of the related studies based on fMRI construct the brain functional connection network by Pearson correlation coefficient.But this method cannot reflect the dynamic trend of the correlation between the functions of different brain regions over time in the neurophysiological process.Therefore,this thesis proposes a dynamic brain function network sequence based on the K-S（Kolmogorov-Smirnov）test method.Based on this graph sequence data,this thesis proposes a graph convolutional network（GCN）and temporal convolutional network（TCN）diagnostic model（GCN-TCN）.We optimize the model through a multi-level loss function consisting of node loss,time slice loss and final loss.The classification results of AD and NC based on ADNI database show that the model has better classification performance.（2）Aiming at the influence of non-imaging information on AD,this thesis proposes a multimodal data learning based on graph neural network.The GCN-TCN model is used to extract imaging information features.Then the imaging information and non-imaging information are fused into a sample relationship graph,and the nodes of the sample relation graph are classified by the dynamic hypergraph neural networks（DHGNN）model.The effects of different non-imaging information on the model are verified by different construction methods.The experimental results show that the classification effect of the model is most improved when gender is used as the phenotypic data.（3）Based on the model in this thesis,we screened out the regions of interest（ROI）that the model focused on as biomarkers for AD.The brain functional connectivity network is constructed by the K-S test method,and the clustering coefficient and local efficiency of related biomarkers are calculated.It is found that the clustering coefficient and local efficiency of the AD group are lower than those of the NC group.Especially the right amygdala,the AD and NC groups have the greatest differences in the average clustering coefficient,indicating that the amygdala function is severely destructed in AD patients.