Limits...
Effects of different correlation metrics and preprocessing factors on small-world brain functional networks: a resting-state functional MRI study.

Liang X, Wang J, Yan C, Shu N, Xu K, Gong G, He Y - PLoS ONE (2012)

Bottom Line: Our results show significant differences in global network metrics associated with both correlation metrics and global signals.Moreover, we found that there were significant frequency-related differences in topological properties of WOGR-PEAR networks, and brain networks derived in the 0.027-0.073 Hz band exhibited greater reliability than those in the 0.01-0.027 Hz band.Taken together, our results provide direct evidence regarding the influences of correlation metrics and specific preprocessing choices on both the global and nodal topological properties of functional brain networks.

View Article: PubMed Central - PubMed

Affiliation: State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University, Beijing, China.

ABSTRACT
Graph theoretical analysis of brain networks based on resting-state functional MRI (R-fMRI) has attracted a great deal of attention in recent years. These analyses often involve the selection of correlation metrics and specific preprocessing steps. However, the influence of these factors on the topological properties of functional brain networks has not been systematically examined. Here, we investigated the influences of correlation metric choice (Pearson's correlation versus partial correlation), global signal presence (regressed or not) and frequency band selection [slow-5 (0.01-0.027 Hz) versus slow-4 (0.027-0.073 Hz)] on the topological properties of both binary and weighted brain networks derived from them, and we employed test-retest (TRT) analyses for further guidance on how to choose the "best" network modeling strategy from the reliability perspective. Our results show significant differences in global network metrics associated with both correlation metrics and global signals. Analysis of nodal degree revealed differing hub distributions for brain networks derived from Pearson's correlation versus partial correlation. TRT analysis revealed that the reliability of both global and local topological properties are modulated by correlation metrics and the global signal, with the highest reliability observed for Pearson's-correlation-based brain networks without global signal removal (WOGR-PEAR). The nodal reliability exhibited a spatially heterogeneous distribution wherein regions in association and limbic/paralimbic cortices showed moderate TRT reliability in Pearson's-correlation-based brain networks. Moreover, we found that there were significant frequency-related differences in topological properties of WOGR-PEAR networks, and brain networks derived in the 0.027-0.073 Hz band exhibited greater reliability than those in the 0.01-0.027 Hz band. Taken together, our results provide direct evidence regarding the influences of correlation metrics and specific preprocessing choices on both the global and nodal topological properties of functional brain networks. This study also has important implications for how to choose reliable analytical schemes in brain network studies.

Show MeSH
Global topological properties of Pearson's-correlation and partial-correlation-based networks with and without global signal regression.Plots show the changes in small-world parameters (Cp, Lp, γ, λ and σ), network efficiency (Local efficiency and Global efficiency), assortativity coefficient (α) and hierarchy coefficient (β) in functional brain networks dependent on both correlation metrics (Pearson's correlation or partial correlation) and global signal (regressed or not) as a function of sparsity thresholds. Local and global efficiency of random and regular networks with the same number of nodes and edges as the real networks were shown in gray lines in the network efficiency plots.
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC3295769&req=5

pone-0032766-g002: Global topological properties of Pearson's-correlation and partial-correlation-based networks with and without global signal regression.Plots show the changes in small-world parameters (Cp, Lp, γ, λ and σ), network efficiency (Local efficiency and Global efficiency), assortativity coefficient (α) and hierarchy coefficient (β) in functional brain networks dependent on both correlation metrics (Pearson's correlation or partial correlation) and global signal (regressed or not) as a function of sparsity thresholds. Local and global efficiency of random and regular networks with the same number of nodes and edges as the real networks were shown in gray lines in the network efficiency plots.

Mentions: Graph theoretical analysis revealed that functional brain networks derived from R-fMRI data show prominent small-world architecture across a wide sparsity range, Figure 2 showed the global network parameters within a sparsity range from 0.1 to 0.4, where the networks are sparse and their small-world attributes are estimable [4]. Compared with random networks, brain networks are highly clustered (i.e., γ>1) and have approximately equivalent path lengths (λ∼1). We also compared the global and local efficiency of the brain networks with those of comparable random networks and regular lattices. The results show that the efficiency curves of the brain networks are generally intermediate between the two extreme cases: the brain networks have greater global efficiency than the lattices but less than the random networks, and they have greater local efficiency than the random networks but less than the lattices (Fig. 2). Taken together, our observations indicate that human functional brain networks have efficient small-world properties regardless of the correlation metric selected or the application of global signal regression. Furthermore, functional brain networks were found to be assortative (assortative coefficients ) and hierarchical (hierarchy coefficients ) over a wide sparsity range (Fig. 2). Under the two subdivided frequency bands (i.e., slow-5 and slow-4), all of the above-mentioned global topological characteristics (small-world, network efficiency, assortative and hierarchy) were also found to be present in the functional brain networks constructed under all four different conditions: WGR-PEAR (Fig. S1), WGR-PAR (Fig. S2), WOGR-PEAR (Fig. S3) and WOGR-PAR (Fig. S4).


Effects of different correlation metrics and preprocessing factors on small-world brain functional networks: a resting-state functional MRI study.

Liang X, Wang J, Yan C, Shu N, Xu K, Gong G, He Y - PLoS ONE (2012)

Global topological properties of Pearson's-correlation and partial-correlation-based networks with and without global signal regression.Plots show the changes in small-world parameters (Cp, Lp, γ, λ and σ), network efficiency (Local efficiency and Global efficiency), assortativity coefficient (α) and hierarchy coefficient (β) in functional brain networks dependent on both correlation metrics (Pearson's correlation or partial correlation) and global signal (regressed or not) as a function of sparsity thresholds. Local and global efficiency of random and regular networks with the same number of nodes and edges as the real networks were shown in gray lines in the network efficiency plots.
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC3295769&req=5

pone-0032766-g002: Global topological properties of Pearson's-correlation and partial-correlation-based networks with and without global signal regression.Plots show the changes in small-world parameters (Cp, Lp, γ, λ and σ), network efficiency (Local efficiency and Global efficiency), assortativity coefficient (α) and hierarchy coefficient (β) in functional brain networks dependent on both correlation metrics (Pearson's correlation or partial correlation) and global signal (regressed or not) as a function of sparsity thresholds. Local and global efficiency of random and regular networks with the same number of nodes and edges as the real networks were shown in gray lines in the network efficiency plots.
Mentions: Graph theoretical analysis revealed that functional brain networks derived from R-fMRI data show prominent small-world architecture across a wide sparsity range, Figure 2 showed the global network parameters within a sparsity range from 0.1 to 0.4, where the networks are sparse and their small-world attributes are estimable [4]. Compared with random networks, brain networks are highly clustered (i.e., γ>1) and have approximately equivalent path lengths (λ∼1). We also compared the global and local efficiency of the brain networks with those of comparable random networks and regular lattices. The results show that the efficiency curves of the brain networks are generally intermediate between the two extreme cases: the brain networks have greater global efficiency than the lattices but less than the random networks, and they have greater local efficiency than the random networks but less than the lattices (Fig. 2). Taken together, our observations indicate that human functional brain networks have efficient small-world properties regardless of the correlation metric selected or the application of global signal regression. Furthermore, functional brain networks were found to be assortative (assortative coefficients ) and hierarchical (hierarchy coefficients ) over a wide sparsity range (Fig. 2). Under the two subdivided frequency bands (i.e., slow-5 and slow-4), all of the above-mentioned global topological characteristics (small-world, network efficiency, assortative and hierarchy) were also found to be present in the functional brain networks constructed under all four different conditions: WGR-PEAR (Fig. S1), WGR-PAR (Fig. S2), WOGR-PEAR (Fig. S3) and WOGR-PAR (Fig. S4).

Bottom Line: Our results show significant differences in global network metrics associated with both correlation metrics and global signals.Moreover, we found that there were significant frequency-related differences in topological properties of WOGR-PEAR networks, and brain networks derived in the 0.027-0.073 Hz band exhibited greater reliability than those in the 0.01-0.027 Hz band.Taken together, our results provide direct evidence regarding the influences of correlation metrics and specific preprocessing choices on both the global and nodal topological properties of functional brain networks.

View Article: PubMed Central - PubMed

Affiliation: State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University, Beijing, China.

ABSTRACT
Graph theoretical analysis of brain networks based on resting-state functional MRI (R-fMRI) has attracted a great deal of attention in recent years. These analyses often involve the selection of correlation metrics and specific preprocessing steps. However, the influence of these factors on the topological properties of functional brain networks has not been systematically examined. Here, we investigated the influences of correlation metric choice (Pearson's correlation versus partial correlation), global signal presence (regressed or not) and frequency band selection [slow-5 (0.01-0.027 Hz) versus slow-4 (0.027-0.073 Hz)] on the topological properties of both binary and weighted brain networks derived from them, and we employed test-retest (TRT) analyses for further guidance on how to choose the "best" network modeling strategy from the reliability perspective. Our results show significant differences in global network metrics associated with both correlation metrics and global signals. Analysis of nodal degree revealed differing hub distributions for brain networks derived from Pearson's correlation versus partial correlation. TRT analysis revealed that the reliability of both global and local topological properties are modulated by correlation metrics and the global signal, with the highest reliability observed for Pearson's-correlation-based brain networks without global signal removal (WOGR-PEAR). The nodal reliability exhibited a spatially heterogeneous distribution wherein regions in association and limbic/paralimbic cortices showed moderate TRT reliability in Pearson's-correlation-based brain networks. Moreover, we found that there were significant frequency-related differences in topological properties of WOGR-PEAR networks, and brain networks derived in the 0.027-0.073 Hz band exhibited greater reliability than those in the 0.01-0.027 Hz band. Taken together, our results provide direct evidence regarding the influences of correlation metrics and specific preprocessing choices on both the global and nodal topological properties of functional brain networks. This study also has important implications for how to choose reliable analytical schemes in brain network studies.

Show MeSH