An FMRI-based Modeling And Classification System For Brain Dynamic Network Sparsity

Nowadays,the diagnosis of brain diseases through brain imaging technology has received more and more attention.The exploration of mechanism in brain functional connectivity based on functional magnetic resonance imaging(fMRI)data is critical for the study of brain diseases.Because attention-deficit/hyperactivity disorder(ADHD)is difficult to diagnose,there is still much space for improvement.In this paper,we make several improvements in classification algorithms and brain dynamic network modeling based on hypergraph to improve the classification accuracy of children with ADHD and normal children(NC).In terms of classification algorithm research,we propose a novel ADHD classification architecture.Considering the dynamic characteristics of brain functional connectivity,Bayesian Connected Change Point Model(BCCPM)is used to detect the dynamics of the resting state brain.After that,local Binary Encoding Method(LBEM)is used to extract local features.Finally,the classification part is based on the Extreme Learning Machine.In the classification part,experiments are performed on the proposed classification framework by using Kernel Extreme Learning Machine(KELM),Hierarchical Extreme Learning Machine(HELM),and Kernel Hierarchical Extreme Learning Machine(KH-ELM).The experimental results show that the single-layer KH-ELM algorithm achieves more stable and high-precision classification results comparing to the existing methods.In terms of sparseness modeling of brain dynamic network,we propose a novel sparseness modeling method of brain dynamic network.Firstly,based on hypergraph theory,a brain functional hyper-network is constructed.Secondly,because the brain functional hyper-network has subnets,we use the sparse representation method and the graph embedding method to extract the sparse features of hyper-networks.We do the experiments by using BCCPM to detect the dynamics of brain.Then we construct the hyper-network and extract features from it.Finally,the features are classified by KELM and SVM respectively.The experimental results show that the two classification algorithms can obtain more stable and high-precision classification results.At the same time,by visualizing the sparse features,we find that although the hyper-networks have similar structures,the weights of the hyper-edges are different.It further reflects the dynamic of brain functional network.Finally,we develop a system,which contains fMRI data pre-processing,dynamic detection,local feature extraction,sparseness modeling of the brain dynamic network and classification.