Generation of Individual Whole-Brain Atlases With Resting-State fMRI Data Using Simultaneous Graph Computation and Parcellation

ORIGINAL RESEARCH published: 04 May 2018 doi: 10.3389/fnhum.2018.00166 Generation of Individual Whole-Brain Atlases With Resting-State fMRI Data Using Simultaneous Graph Computation and Parcellation J. Wang 1,2 , Z. Hao 1 and H. Wang 1* 1 School of Mathematics and Big Data, Foshan University, Foshan, China, 2 Key Laboratory of Child Development and Learning Science of Ministry of Education, Research Center for Learning Science, Southeast University, Nanjing, China Edited by: Muthuraman Muthuraman, Universitätsmedizin der Johannes Gutenberg, Universität Mainz, Germany Reviewed by: Gabriel Gonzalez-Escamilla, Universitätsmedizin der Johannes Gutenberg, Universität Mainz, Germany Abdul Rauf Anwar, University of Engineering and Technology, Pakistan *Correspondence: H. Wang Received: 20 December 2017 Accepted: 10 April 2018 Published: 04 May 2018 Citation: Wang J, Hao Z and Wang H (2018) Generation of Individual Whole-Brain Atlases With Resting-State fMRI Data Using Simultaneous Graph Computation and Parcellation. Front. Hum. Neurosci. 12:166. doi: 10.3389/fnhum.2018.00166 The human brain can be characterized as functional networks. Therefore, it is important to subdivide the brain appropriately in order to construct reliable networks. Resting-state functional connectivity-based parcellation is a commonly used technique to fulfill this goal. Here we propose a novel individual subject-level parcellation approach based on whole-brain resting-state functional magnetic resonance imaging (fMRI) data. We first used a supervoxel method known as simple linear iterative clustering directly on resting-state fMRI time series to generate supervoxels, and then combined similar supervoxels to generate clusters using a clustering method known as graph-without-cut (GWC). The GWC approach incorporates spatial information and multiple features of the supervoxels by energy minimization, simultaneously yielding an optimal graph and brain parcellation. Meanwhile, it theoretically guarantees that the actual cluster number is exactly equal to the initialized cluster number. By comparing the results of the GWC approach and those of the random GWC approach, we demonstrated that GWC does not rely heavily on spatial structures, thus avoiding the challenges encountered in some previous whole-brain parcellation approaches. In addition, by comparing the GWC approach to two competing approaches, we showed that GWC achieved better parcellation performances in terms of different evaluation metrics. The proposed approach can be used to generate individualized brain atlases for applications related to cognition, development, aging, disease, personalized medicine, etc. The major source codes of this study have been made publicly available at https://github.com/ yuzhounh/GWC. Keywords: whole-brain parcellation, resting-state fMRI, supervoxel, graph-without-cut, random parcellation INTRODUCTION Since the first manifestation that specific brain areas are functionally connected in resting brain (Biswal et al., 1995), neuroscientists have been characterizing the human brain as networks (Sporns et al., 2005; Bullmore and Sporns, 2009). To construct brain networks, a critical step is to parcellate the brain into a specific number of functional units (Wig et al., 2011). However, no agreement has been reached on how the brain should be parcellated (Hallquist and Hillary, 2018). Frontiers in Human Neuroscience | www.frontiersin.org 1 May 2018 | Volume 12 | Article 166 Wang et al. Individual Subject Level Parcellation by GWC designed to segment 2D images. Traditional graph-based approaches organize the elements of an image into a graph and then partition the image based on the graph. GWC merges the two steps, i.e., calculating the graph and partitioning the image, into a single optimization problem. This algorithm design generates the optimal graph for segmentation. Both spatial information and multiple visual features of the image are considered in GWC. Additionally, GWC restricts the number of connected components in the obtained graph so that it is exactly equal to the initialized cluster number. Gao et al. (2016) have reported that GWC achieves better clustering performances than some existing image segmentation approaches. Therefore, we extended GWC to 3D space and applied it to perform wholebrain parcellation for individuals in this study. After generating a brain atlas, it is important to ensure that the brain atlas does not rely heavily on spatial structures. Different parcellation approaches incorporate spatial structures in different ways. In the normalized cuts (Ncut) approach (Craddock et al., 2012), spatial structure is introduced by the spatial constraint in weight definition. In the SLIC approach (Wang and Wang, 2016), spatial structures are introduced by initializing an ideal geometric pattern, integrating the spatial distance into the unified distance, and searching in a local space. As Wang and Wang (2016) have shown, incorporating suitable spatial structures in whole-brain parcellation approaches is quite necessary to guarantee the spatial contiguity of the resultant clusters. However, parcellation approaches with excessive spatial structures would encounter three major problems (Craddock et al., 2012; Blumensath et al., 2013; Shen et al., 2013; Gordon et al., 2016; Wang and Wang, 2016). First, they tend to generate clusters with comparable shapes and sizes, which are unlikely to be the functional units in the brain (Glasser et al., 2016). Second, when applying these approaches, random parcellation would be visually similar to functional parcellation (Craddock et al., 2012). Third, when applying these approaches, random parcellation and functional parcellation tend to achieve nearly identical performances under different evaluation metrics, such as Dice coefficient and silhouette width (Craddock et al., 2012; Wang and Wang, 2016). The utility of such approaches is limited due to the above three problems. Therefore, to justify a parcellation approach, besides visually inspecting the generated clusters, it is necessary to compare the results obtained based on functional magnetic resonance imaging (fMRI) data to those obtained based on random data obtained using the same approach. If the two results are very close, then the parcellation approach encounters the above problems and is not reasonable, and vice versa. To our knowledge, only few studies (Gordon et al., 2016; Parisot et al., 2016; Arslan et al., 2017; Gallardo et al., 2017) have demonstrated that their parcellations are better than corresponding random parcellations. Among these studies, Gordon et al. (2016) created null models by randomly rotating each hemisphere of the original parcellation, Parisot et al. (2016) and Arslan et al. (2017) created random parcellations by Poisson disk sampling, and Gallardo et al. (2017) created random parcellations by random region growing and random hierarchical clustering. All of these studies fo (...truncated)