Research On Methods Of Brain Functional Connectivity Based On Capsule Network

Resting-state functional magnetic resonance imaging(rs-f MRI)shows great promise in neuroimaging studies as a powerful diagnostic tool to identify abnormal neural functions in brain regions,the study of rs-f MRI brain functional connectivity is helpful to explore the differences of coactivation levels among different brain regions.The emergence of deep learning makes it possible to develop an end-to-end computer-aided diagnosis methods and find abnormal biomarkers scientifically.In this thesis,based on rs-f MRI,a capsule network deep learning method is used to study the high-performance auxiliary identification algorithm for autism disease diagnosis and sex classification,and explore the significant differences of brain functional connectivity in ASD disease and sex.The main contributions are as follows:(1)Aiming at the problem that the traditional complex network architecture is easy to overfit and ignore the spatial position information of features,this thesis utilizes the feature representation capability of dynamic routing algorithm to design a disease diagnosis method based on shallow capsule network for the study of disease-related brain functional connectivity.The experimental results show that the proposed method can achieve effective performance improvements in multi-site ASD disease datasets from different scanning equipment and protocols.(2)In order to solve the limitation of visualization method of the reconstructed input image of the original capsule network for finding biomarkers,in this thesis we propose a method on the model internal logical interpretability of gradient-weighted class activation map for the capsule network(Caps-Grad-CAM)according to the characteristics of the vector output of the capsule network.By the proposed Caps-Grad-CAM method to statistically interpret the feature contributions of sample groups in recognition judgments,the most distinguished abnormal functional connectivity is found among brain regions in the diagnosis of ASD disease and healthy subjects.(3)In order to overcome the limitations of single scale features and linear stacked deep-layer networks,a multi-scale residual capsule network is proposed to study the sex brain functional connectivity.Features in different scales can be adaptively detected with multi-branches and convolution kernels of different sizes,and the residual network is used to obtain the hierarchical features,and according to different feature fusion strategies,multiple features interact with each other to enrich learning information and the final classification is achieved with the capsule network.The results show that model can achieve higher classification accuracy of 90.60%.Meanwhile,the sex related significant functional connectivity is found with the proposed Caps-Grad-CAM method.The proposed methods in this thesis can help effectively identify and find significant ASD disease and sex related functional connectivity differences,which is of great significance for the early diagnosis of neurological diseases and the exploration of potential biomarkers in the future.