Evaluating the repeatability and validity of community detection in functional brain networks

1722 

Abstract Submission 

Makoto Fukushima¹, Ryusuke Nakamura², Kazushi Ikeda^2,3

¹Graduate School of Advanced Science and Engineering, Hiroshima University, Hiroshima, Japan, ²Graduate School of Science and Technology, Nara Institute of Science and Technology, Nara, Japan, ³Data Science Center, Nara Institute of Science and Technology, Nara, Japan

Makoto Fukushima  
Graduate School of Advanced Science and Engineering, Hiroshima University
Hiroshima, Japan

Ryusuke Nakamura  
Graduate School of Science and Technology, Nara Institute of Science and Technology
Nara, Japan

Kazushi Ikeda  
Graduate School of Science and Technology, Nara Institute of Science and Technology|Data Science Center, Nara Institute of Science and Technology
Nara, Japan|Nara, Japan

Network analysis methods have been widely used to characterize complex connectivity patterns in functional networks in the brain. One such method is community detection, which decomposes the entire network into communities (or modules) consisting of brain regions whose activity time series are positively correlated (Sporns and Betzel, 2016). These functional communities are often detected by maximizing the modularity with the resolution parameter that controls the spatial size of the detected communities (Reichardt and Bornholdt, 2006; Rubinov and Sporns, 2011). Optimal resolution parameters were chosen so that the detected communities were repeatable across multiple trials of modularity maximization (Mišić et al., 2016) or were valid and similar to known intrinsic connectivity networks (ICNs) (Hilger et al., 2020). However, how the repeatability and validity of detected functional communities change with resolution parameters has not been comprehensively investigated. In this study, we systematically evaluated these repeatability and validity across the spectrum of resolution parameters using multiple criteria and parcellation scales with/without global signal regression (GSR) (Murphy and Fox, 2017).

Group-averaged functional brain networks were derived from resting-state fMRI data of the 100 unrelated subjects downloaded from the public database of the WU-Minn Human Connectome Project (Van Essen et al., 2013). The fMRI data of 15 high motion subjects and 1 subject categorized as age ≥ 36 were excluded in this study. Nodes in the functional brain networks were defined based on two multiscale parcellation atlases, Lausanne (Cammoun et al., 2012) and Schaefer (Schaefer et al., 2018). The functional brain networks were constructed with or without GSR. Communities in the functional brain networks were detected by modularity maximization and the consensus clustering approach (Sporns and Betzel, 2016). The repeatability of community detection was evaluated by quantifying the similarity of initial and consensus partitions and the similarity of consensus partitions in different sessions. The validity of community detection was evaluated by quantifying the similarity of consensus communities to the seven canonical ICNs (visual, somatomotor, dorsal attention, ventral attention, limbic, frontoparietal, and default mode) (Yeo et al., 2011), visualizing their spatial overlap, and checking the existence of tiny communities.

We found that when the repeatability of detected communities over resolution parameters was highest, the number of detected communities was around three or four in most cases. (Fig. 1). The similarity of these most repeatable communities to the seven ICNs was not as high as in cases where a larger number of communities were detected (Fig. 2A). When the similarity to the seven ICNs was the highest, the number of communities detected was in most cases around eight, with a few tiny communities often appearing when finer parcellation scales were used (Fig. 2A). From low to high resolution parameters, communities corresponding to the visual, somatomotor, and default mode networks appeared first, followed in most cases by the ventral attention network and then the remaining ICNs (Fig. 2B). With GSR, communities corresponding to the limbic and frontoparietal networks rarely appeared until the entire network was decomposed into about 30 communities (Fig. 2B).

Using multiple parcellation scales, we evaluated the repeatability and validity of functional brain network communities detected by modularity maximization across resolution parameters. Our results illustrate how the repeatability and validity of detected communities vary with resolution parameters and depend on the use of parcellation scales and GSR. These results may help users choose resolution parameters, parcellation scales, and whether or not to use GSR when applying modularity-based community detection to functional brain networks.

fMRI Connectivity and Network Modeling ¹

Task-Independent and Resting-State Analysis ²

Data analysis

FUNCTIONAL MRI

Other - Connectomics; Network Neuroscience