Research On Analytical Approach For Hierarchical Functional Brain Network Under Naturalistic Paradigm Based On NAS-DBN Model

Functional magnetic resonance imaging technology(fMRI)is the most effective way to explore mysteries of human brain function,where naturalistic paradigm has become an effective experimental paradigm to explore brain function activities in human daily life scenarios.Identifying meaningful brain functional networks from naturalistic paradigm fMRI(NfMRI)and characterizing important hierarchical structure of brain function networks have become a hot topic in the current field of brain science.Although there are a few deep learning models focusing on building hierarchical structure of brain functional networks,most of existing models rely on manual parameter tuning,at the same time,the inherent characteristics of brain neural activity under naturalistic paradigm and complex hierarchical spatiotemporal organization of brain spatial information are ignored by current models.For these questions,this thesis proposes two methods of functional brain network recognition and analysis based on the combination of neural architecture search and deep belief network.The main work and contributions are summarized as follows:(1)A model that VS-DBN(Volumetric Sparse DBN)based on neural architecture search(Volumetric NAS-DBN)is proposed to model group-level NfMRI volumetric images.The method focuses on the hierarchical spatiotemporal features in NfMRI volumetric image data.Experimental results demonstrate that automatically optimized VS-DBN model is able to mine meaningful spatiotemporal features in NfMRI volumetric images,further revealing the brain functional hierarchical structure in NfMRI volumetric images.(2)A two-stage deep belief network based on neural architecture search(NAS two-stage DBN)is proposed to model NfMRI time series signals.This method is based on the neurological characteristics of brain functional activity under naturalistic paradigm,taking into account group consistency and individual differences of NfMRI signals.The first stage builds the group model according to the NAS search structure,and the second stage initializes the individual model according to the group model parameters.The results show that the optimized model reveals consistency and difference of group and individual brain functional networks under naturalistic paradigm,at the same time mines meaningful brain functional hierarchical organization,highlighting the model’s good representation ability.(3)In order to evaluate the performance and robustness of model,the dataset is expanded based on work(2),NAS two-stage DBN is systematically compared with commonly brain function network recognition methods based on used data-driven.Experimental results show the proposed model in this thesis reveals the inherent properties and hierarchical structure of robust brain functional networks on both datasets,and results significantly outperform comparative methods.Finally,further analysis of the spatiotemporal features revealed that results of model characterization were in line with characteristics of neuroscience,proving effectiveness of model proposed in this thesis.