Research On Alzheimer's Disease Related Multimodal Brain Network Based On Persistent Homology And Sparse Representation Modeling

Alzheimer's disease(AD)is a neurodegenerative brain disease.Patients with AD suffer from severe memory impairment and loss of daily behavior,which affects the normal life seriously.The identification of early AD helps to delay the deterioration of the disease by the drug or non-drug intervention in the early stage of onset.However,early AD involves very subtle brain tissue changes,which is extremely challenging for the early diagnosis of AD disease.With the development of neuroimaging technology,a growing body of research shows that the progressive process of AD is associated with the damage in functional and structural brain network.Hence,the exploration with respect to early diagnosis and pathological characteristics of AD from the perspective of brain connectome has become a research hotspot in the AD-related brain science field.Currently,most brain network modeling methods are based on the pair-wise correlation framework.However,the neurophysiological process of the human brain often involves the interaction within multiple co-excited brain regions,rather than the pairwise correlation pattern between brain regions.The traditional correlation brain modeling has shown the deficiencies.In addition,the feature extraction and the precise analysis for complex brain networks are the basis for identifying abnormal connections from the normal brain network.Nevertheless,the popular brain network analysis methods based on a single threshold scale cannot effectively extract all the useful information from the neuroimaging data.Therefore,how to construct a brain network structure and quantify the network features accurately is the key for AD disease-related brain network research.This paper aims to explore the brain network characterization method using multi-dimensional features based on persistent homology.We integrate the correlation network,sparse representation modeling,multimodal analysis,and subject-level structure into an entire framework for in-depth and innovative study on the AD-related brain network mechanism.The main contributions of this paper are as following:(1)We propose a brain network quantification method by combining the multi-dimensional features of persistent homology,aiming to address the problem with poor interpretability and poor stability in current brain network analysis methods.On the one hand,we use the 0-dimensional feature of persistent homology(PH-0)to explore the persistence of connected components during graph filtration;on the other hand,we further apply the 1-dimensional feature of persistent homology(PH-1)for evaluating the persistence of cycle structures during graph filtration.Additionally,the difference between subject-level brain networks is computed using a distance-based measurement,in order to evaluate the effectiveness of the multi-dimensional features in AD disease identification.Finally,we design a permutation test process based on the multi-dimensional features to verify the statistical power of the features in disease-related analysis.The proposed method explores the whole brain connectivity patterns from low-dimensional to high-dimensional perspective.The experimental results show that the multi-dimensional features are superior to the other commonly used graph theory measures in terms of subject-level classification and statistical performance evaluation.(2)The single-modality-based brain network modeling often lacks the supplementary information of multi-source resolution,and shows a poor temporal-spatial sensitivity.We hence propose a multimodal fusion based sparse Lasso brain network modeling technique.First,we use the Diffusion Tensor Imaging(DTI)data as the regularization item of Lasso,for guiding functional magnetic resonance imaging(fMRI)modeling.After that,the modality information from DTI and fMRI can be integrated into the Lasso-based brain network modeling framework.Second,we employ functional connectivity strength to construct generalized fused Lasso,aiming to limit the influence of group constraint and achieve the subject-level multimodal brain network.The experimental results demonstrate the proposed multimodal brain network can reflect the brain structure well,and is outperformed the traditional single-modality-based brain models and Pearson's correlation model in terms of classification performance.(3)During the process of the Lasso brain network modeling,the selection of regularization parameters is uncertain.Moreover,it is necessary to accurately evaluate the connectivity strength between brain regions in the sparse brain model,which is helpful to extract the complete network connectivity patterns.However,the modeling complexity is also increased naturally.In order to address this limitation,we propose a persistent homology based sparse brain network analysis framework under multi-scale regularization.First,we integrate the brain networks with the different sparsity corresponding to multi-scale regularization parameters by the edge weighting and the grid search.That can avoid the selection of single regularization parameter.Second,persistent homology method is introduced to analyze the computational topology features of the integration brain network.This way can transfer connectivity strength evaluation to connectivity structure measure.The experiment is designed by the two different groups of data,and they include APOE ?4 genotype data and EMCI data.The results verify the effectiveness of the proposed method from the perspective of statistics and classification performance.(4)For brain network analysis of neurological diseases such as Alzheimer's disease,the main purpose is to identify disease-related functional or structural network disconnections,and to explore the pathological mechanism of the disease from the brain connectomics level.We propose a disease-specific brain connectivity pattern identification method based on the multi-dimensional features of persistent homology.Specifically,the PH-0 is used to identify the disease-related hub connections of sparse brain network,and the AD-specific brain connectivity patterns can be further extracted by the way of group analysis and difference evaluation.Moreover,for exploring the effectiveness of PH-1 in identifying brain connectivity pattern,we evaluate the frequency cycle structures of correlation brain network model,which helps to find more neurobiologically meaningful connectivity patterns.The experiment uses the sparse network and the correlation network to verifiy the effectiveness of the multi-dimensional persistent features in identify disease-related brain connectivity patterns.