A Supervoxel-based Whole Brain Parcellation Study With Resting-state FMRI Data

Node definition is a very important issue in human brain network analysis and functional connectivity studies.Typically,the atlases generated from meta-analysis,random criteria,and structural criteria are utilized as nodes in related applications.However,these atlases are not originally designed for such purposes and may not be suitable.To generate more appropriate nodes,we need to generate more appropriate brain atlases by parcellating the whole brain.In this paper,we introduced a supervoxel method called simple linear iterative clustering(SLIC)to parcellate whole brain resting-state fMRI data to achieve this goal.With this supervoxel method,we conducted three studies where several whole brain parcellation approaches were proposed.To demonstrate the reasonability and superiority of the proposed approaches,we compared them with several state-ofthe-art approaches under different evaluation metrics which included spatial contiguity,functional homogeneity,and reproducibility.For group level parcellation approaches,both the group-to-group reproducibility and the group-to-subject reproducibility were evaluated in our study.We varied the cluster number in a wide range in order to generate parcellations with multiple granularities.The three studies are detailed as follows.The first study applied SLIC directly on the resting-state fMRI time series to perform whole brian parcellation.With this parcellation approach,we could generate brain atlases without feature extraction.To demonstrate the reasonability of the proposed approach,we compared it with a stateof-the-art whole brain parcellation approach,i.e.,the normalized cuts(Ncut)approach,under different evaluation metrics.The experimental results showed that the proposed approach achieved satisfying performances.The second study combined Ncut and SLIC to perform whole brian parcellation.Two group level parcellation approaches,i.e.,the mean SLIC and two-level SLIC approaches were proposed.Specifically,Ncut was employed to extract features from connectivity matrices,and then SLIC was applied on the extracted features to generate the final brain atlases.The two SLIC approaches and three state-of-the-art approaches were compared under different evaluation metrics.Experimental results showed that the two SLIC approaches achieved good parcellation performances.This study also investigated several confounding factors that might influence the parcellation results,including different sparsifying schemes,global signal regression,overclustering,different weighting functions,etc.These considerations further demonstrated the reasonability and superiority of the proposed approaches.The third study utilized a clustering method called graph-without-cut(GWC)to perform whole brain parcellation based on supervoxel.Specifically,we applied SLIC directly on resting-state fMRI time series to generate supervoxels,and then aggregated similar supervoxels to generate clusters by GWC.By comparing the results of the GWC approach on fMRI data and on random data,we demonstrated that GWC does not rely heavily on spatial structures,thus avoiding the widely existing problem encountered by many previous whole brain parcellation approaches.After that,by comparing the GWC approach with the modified Ncut and SLIC approaches,we showed that GWC achieved better parcellation performances under different evaluation metrics.The proposed whole brain parcellation approaches and the generated brain atlases might find their applications in various studies related to brain network analysis,e.g.,cognition,development,aging,diseases,and personalized medicine.The generated atlases and major source codes of this study have been made publicly available online,see the appendix B for details.From the experimental results,we cannot find an optimal cluster number.Therefore,to utilize the brain atlases generated in this study,the cluster number could be chosen according to requirements.