High-Order Brain Functional Network Learning And Its Applications

Brain functional network(BFN)is an important tool for understanding the cerebral working mechanism and exploring brain diseases and its biomarkers.In the past decades,many BFN estimation methods have been proposed,such as Pearson’s correlation(PC)and sparse representation(SR).However,these methods only encode the direct associations between brain regions or spatial region of interests(ROIs)(low-order relationship),while the relationship between associations,i.e.,the high-order relationships of BFN,are ignored.In fact,the brain is a complex network,and the high-order relationships between ROIs can easily capture the global and the complementary information of ROIs.Thus,this thesis attempts to construct high-order BFNs.The specific research contents and results are as follows:1.Learning high-order BFNs based on multiple sequential PCs operation.The study shows that the method of constructing high-order BFN via conducting two sequential PCs can provide complementary information for group difference analysis.This inspires us to continue the correlation operation to explore the characteristics and properties of higherorder correlation.Therefore,we applied the constructed BFNs to conduct the gender and brain disease classification task.Through experiments,we have the following findings:(1)with the increase of PC operation,the discriminative ability of the constructed BFNs tends to decrease,and the sequence of BFN adjacency matrices will converge;(2)fusing the highorder BFNs with the low-order BFN constructed by traditional PC can improve the sensitivity for mild cognitive impairment(MCI)identification.2.Learning high-order BFNs based on the signed random walk(SRW).BFNs constructed by traditional methods only capture the low-order relationship(i.e.,the direct connectivity strength between ROIs).ignoring the high-order information in our brain(e.g.,the global network structure).To solve this issue,we introduced the signed random walk(SRW)to estimate high-order BFNs to capture the global network structure information.Not only this method can measure the global network structure,but also naturally consider the positive and negative edge weights in the functional connections of the brain,and reasonably deal with the cooperative and inhibitory relationships between ROIs through structural balance theory.To illustrate the effectiveness of the proposed method,we identify subjects with MCI from normal controls(HCs)based on the estimated high-order BFN.Experimental results show our scheme tends to achieve higher classification performance than baseline method in the sense of classification accuracy(ACC),sensitivity(SEN),specificity(SPE),positive predictive value(PPV)and negative predictive value(NPV).