Research On Brain Imaging Genetics Methods For Alzheimer's Disease Based On Multimodal Fusio

Alzheimer’s disease(AD)is a prevalent neurodegenerative disease that typically affects elderly individuals,and is characterized by a progressive decline in cognitive and memory function.Unfortunately,AD is irreversible and the exact pathogenic mechanisms remain unclear,with no effective treatment available to date.Additionally,the insidious nature of AD means that there are no accurate methods for early diagnosis,often resulting in missed opportunities for optimal intervention.Therefore,it is crucial to identify the pathogenic mechanisms and determine the accurate biological markers for early diagnosis of AD,as this could facilitate timely intervention and delay disease progression.This is essential for promoting healthy aging,implementing tiered healthcare system,improving relevant social security systems,and contributing to establishing the new direction in scientific technological development of the safeguarding people’s health.Neuroimaging genetics can integrate evidence chains from genetic,imaging,and behavioral data of different scales and modalities,providing a more holistic perspective on the pathogenic mechanisms of AD.This dissertation focuses on neuroimaging genetics data,utilizing machine learning and deep learning methods to model and analyze the complex relationships among ADrelated pathogenic factors,conducting three progressive studies:(1)modeling the gene-brain region binary association;(2)modeling the gene-brain region heterogeneous relationship;(3)modeling the gene-brain region multi-scale association.The specific research contents are as follows:(1)In response to the problem of high-dimensional and small-sample neuroimaging genetics data,a gene-brain region binary association extraction and disease state classification method based on ensemble learning was proposed.This method integrates multimodal data into gene-brain region binary association features through the correlation representation method.Disease state classification is performed by extracting gene-brain region binary association features with disease discriminative power using ensemble pruning method.Compared with other methods,this method has better classification performance and feature selection ability,providing strong support for the clinical diagnosis and pathogenic mechanism identification of AD.(2)In response to the problem of heterogeneity among multi-modalities data,a graph capsule convolutional network was proposed to predict the progression of mild cognitive impairment and identify the pathological mechanisms.Heterogeneous pathogenic information association graphs were constructed with genes and brain regions as nodes to integrate multi-modalities data.Heterogeneities between multi-modalities were disentangled into a set of latent components to represent capsules,and the disease-specific information flow among pathogenic factors was captured by the graph capsule convolutional network.Experimental results demonstrated the superiority of this method,and further analysis showed that it could identify more disease-relevant pathological factors for AD.(3)In response to the problem of complexity of interaction among pathogenic factors,a graph Transformer with local structure awareness was proposed and applied to AD disease state prediction and pathogenesis identification.Factor interaction graphs were constructed to integrate the microscopic genetic variations and macroscopic brain activities.Local structure perceivers and global dependency reasoning components encoded the tightly local interaction structures in the graphs and combined them into global higher-order interaction structures.These interaction patterns at different scales were finally aggregated into a classifier to predict the disease state.Experimental results show that this method has higher classification performance than existing methods and can accurately identify pathogenic factors related to AD.