Construction And Classification Of Brain Functional Networks Based On Pattern Decomposition

Network connectivity provides an important tool for understanding the complex functional system of the human brain.A fundamental issue in neuroscience studies is how to accurately construct the brain functional network and depict the underlying network organization.The non-invasive neuroimaging techniques,especially the functional magnetic resonance imaging?fMRI?,provides an effective channel for studying the brain functions.However,studying the brain functional network is still a challenging task due to the high complexity in time and space with low signal-to-noise ratio of the neuroimaging data.Therefore,it is of great significance to design novel methods for the effective construction and analysis of the brain functional net-work.Data-driven approaches open a new avenue for studying the brain functional network.In this work,the technique of pattern decomposition is adopted to study the brain functional network,based on resting-state fMRI data.Specifically,novel methods are proposed for the construction of the brain functional network and for the analysis of the network organization.Non-overlapping and overlapping community structures in the brain functional network are de-tected both at an individual level and at an across-subject level.Furthermore,these methods are applied to investigate the brain functional networks in populations with brain diseases.This work mainly contains the following four parts:1.For the construction of the brain functional network,a novel method based on sparse representation and clustering is proposed to calculate and analyze the functional connectivity.Specifically,a pattern decomposition approach,i.e.,adaptive sparse representation?ASR?,is applied to compute the functional connectivity for an accurate construction of the brain func-tional network.In addition,the affinity propagation?AP?clustering is applied to identify the non-overlapping community structure.Experimental results on simulated fMRI data show that ASR achieves a high sensitivity of 90.6%in depicting the functional connectivity and the frame-work of ASR with AP achieves a clustering accuracy of 74.8%,higher than those of the compet-ing methods.Experimental results on real fMRI data show that the functional network derived by ASR has a higher test-retest reliability and modularity.Besides,the community structure derived by ASR has a higher clustering quality with a better interpretation,further suggesting the rationality and effectiveness of ASR in constructing the brain functional network.2.For the community detection of the brain functional network,a novel method is pro-posed based on non-negative matrix factorization?NMF?.By adding a sparsity constraint in the form of?₁-norm regularizer on the symmetric NMF?symNMF?method,a sparse symmetric NMF?ssNMF?method is proposed to detect the overlapping community structure of the brain functional network.Furthermore,a non-negative ASR?NASR?method is proposed by adding a non-negative constraint on ASR to improve the physical meaning of the obtained functional connectivity.Simulated experimental results show that ssNMF is able to accurately detect both the overlapping and non-overlapping community structures,with higher stableness.Real ex-perimental results show that the communities derived by ssNMF have a high reproducibility and a better physiological interpretation.Besides,the communities are highly consistent with the resting-state networks discovered in the previous studies,further suggesting the rationality and effectiveness of the proposed ssNMF method.3.For the community detection of the brain functional network across multiple subjects,a novel method is proposed by adopting a collective way.The proposed method,termed as collective sparse symmetric NMF?cssNMF?,is able to identify the group-level overlapping community structure across subjects,and in the meantime to preserve the individual differences in community strengths.Simulated experimental results show that cssNMF performs better than the competing methods by achieving an accuracy of 94.4%in detecting the community structure.It also effectively saves the individual variations with a similarity of above 90.0%between the derived community strengths and the ground truth.Real experimental results show that the derived community structure has a high cross-session,cross-subjects and cross-data reproducibility,with a physiologically meaningful interpretation.Besides,most overlapping nodes identified by cssNMF are found to locate in the frontal and parietal lobes and in the fronto-parietal network,which is consistent with the findings of previous studies.Furthermore,the community strengths captured at an individual level have a high test-retest reliability.These results show that the cssNMF method provides a useful tool to study the overlapping community structure of the brain functional network with great potential for individual identification.4.A novel framework is proposed for investigating the community structure of the brain functional network in Alzheimer's disease?AD?,and for detecting AD at an individual level.Based on the NASR-derived brain functional network,the cssNMF method is combined with the agglomerative hierarchical clustering method to study the overlapping community structure and the hierarchical structure of the communities derived at different scales.Furthermore,a frame-work is proposed for detecting AD at an individual level,based on the community strengths captured by cssNMF.Experimental results show that the overlapping and hierarchical commu-nity structures at different scales are significantly altered in AD patients,with a significantly reduced reproducibility at a refined scale,declined functional segregation and reduced nodal functional diversity.Besides,the community strengths of the basal ganglia-thalamus network and the default mode network show significant positive correlations with the cognitive ability.In addition,the AD detection framework based on community strengths is able to identify the AD patients.By using the nearest neighbour classifier with the low-dimensional community strengths as features,it achieves a classification accuracy,sensitivity and specificity of 64.7%,70.0%and 60.0%,respectively,where the accuracy is statistically significant with p<0.05 by a permutation test.These findings provide complementary knowledge for identifying biomarkers for AD.