A Research Of Regularization Framework Based Functional Brain Network Estimate And Application

In the past decades,a wide research is focused on the brain network.These researches on the functional or structural brain network have been made a great progress.It shows that the functional or structural bran network covers lots of topological properties.In particular,functional brain network(FBN)has been becoming an increasingly important measurement for exploring the cerebral working mechanism and mining informative biomarkers for assisting diagnosis of some neurodegenerative disorders(such as Parkinson’s disease,autism spectrum disease and Alzheimer’s disease).Despite its potential performance in discovering the valuable patterns hidden in the brains,the existing model are inflexible.Therefore,It is the research trend to discover how to incorporating the prior information into the brain network model.Aiming to overcome these issues,this paper incorporates the brain network model into a regularized framework,which proposed a very flexible model.The regularized framework can introduce the prior information by setting the regularization term and set the statistical model by setting the data-fitting term.Specifically,this paper extent the traditional Pearson’s model and Sparse Representation model by incorporating the regularization terms,which are based on the biological/physical prior,data quality prior and group constraint prior,and obtain a new functional brain network which combines the prior information.Based on the regularization framework,the contribution of this paper can be given as follows:(1)We remodel the traditional Pearson based FBN estimation method to an optimization model and introduced it into the regularization framework.Meanwhile,we introduced the regularization term by the sparsity and scale-free prior information and obtain the network with these topologies.The results show that this information can help the improvement to the discriminability for the estimated FBN.(2)We introduce a regularization term into the FBN estimation model by the data quality prior.In this way,we can simultaneously estimate the FBN and scrub the ‘dirty’ points,which can alleviate the lack of the independence between these two tasks.The results reveal that the scrubbed ‘dirty’ data points are significantly related to some special resting states.(3)In order to incorporate the group constraint,we extend our model to a tensor format.In particular,we use the tensor low-rank regularizer for a simply trying to approximate the group similarity prior information.The low-rank regularizer is estimated by the PARAFAC decompose.The accuracy of the results significantly improved with the group information introduced.In the end,in order to further verifying the constructed brain network,we use these networks for a simple neuro disease diagnosis task.For alleviating the confounding effect on the feature selection and classification model,we adopt the simplest feature selection and classification model.The result of experiment shows that the brain network with the prior information can be mode discriminability on the neuro disease diagnosis.the proposed FBN modeling method can achieve higher classification accuracy,outperforming the baseline methods.