Limits...
A data-driven method to reduce the impact of region size on degree metrics in voxel-wise functional brain networks.

Liu C, Tian X - Front Neurol (2014)

Bottom Line: In this study, we designed a data-driven method to reduce this impact of the region size in degree metrics by (1) disregarding all self-connections among voxels within the same region and (2) regulating connections from voxels of other regions by the sizes of those regions.In addition, our modified method improved test-retest reliability of degree metrics as well as the sensitivity in group-statistic comparisons.As a promising new tool, our method may reveal network properties that better represent true brain architecture without compromising its data-driven advantage.

View Article: PubMed Central - PubMed

Affiliation: Queensland Brain Institute, The University of Queensland , St Lucia, QLD , Australia.

ABSTRACT
Degree, which is the number of connections incident upon a node, measures the relative importance of the node within a network. By computing degree metrics in voxel-wise functional brain networks, many studies performed high-resolution mapping of brain network hubs using resting-state functional magnetic resonance imaging. Despite its extensive applications, defining nodes as voxels without considering the different sizes of brain regions may result in a network where the degree cannot accurately represent the importance of nodes. In this study, we designed a data-driven method to reduce this impact of the region size in degree metrics by (1) disregarding all self-connections among voxels within the same region and (2) regulating connections from voxels of other regions by the sizes of those regions. The modified method that we proposed allowed direct evaluation of the impact of the region size, showing that traditional degree metrics overestimated the degree of previous identified hubs in humans, including the visual cortex, precuneus/posterior cingulate cortex, and posterior parietal cortex, and underestimated the degree of regions including the insular cortex, anterior cingulate cortex, parahippocampus, sensory and motor cortex, and supplementary motor area. However, the locations of prominent hubs were stable even after correcting the impact. These findings were robust under different connectivity thresholds, degree metrics, data-preprocessing procedures, and datasets. In addition, our modified method improved test-retest reliability of degree metrics as well as the sensitivity in group-statistic comparisons. As a promising new tool, our method may reveal network properties that better represent true brain architecture without compromising its data-driven advantage.

No MeSH data available.


Evaluation of adaptive thresholds is shown. (A) Correlation distribution of all nearest neighbors to their target voxels before and after two-step filtering. (B) Adaptive thresholds based on temporal and spatial correlation of three scans. Bars represent mean ± SEM, ***p < 0.001.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4195315&req=5

Figure 4: Evaluation of adaptive thresholds is shown. (A) Correlation distribution of all nearest neighbors to their target voxels before and after two-step filtering. (B) Adaptive thresholds based on temporal and spatial correlation of three scans. Bars represent mean ± SEM, ***p < 0.001.

Mentions: Figure 4A shows the temporal correlation distribution of all nearest neighbors to their target voxels of one subject. The right tail of the distribution contains high correlations from highly homogeneous functional regions. These high correlations should be preserved for the estimation of Ta because they represent the correlations from the same functional regions. On the contrary, the left tail contains negative and low correlations, which should be excluded for the estimation of Ta, because they may represent voxels located near the boundary of functional regions, but not in the same region as the target voxel. The first filtering eliminates all negative and most low correlations (shown in purple in Figure 4A), and the second filtering further reduces the number of low correlations and increases the value of Ta.


A data-driven method to reduce the impact of region size on degree metrics in voxel-wise functional brain networks.

Liu C, Tian X - Front Neurol (2014)

Evaluation of adaptive thresholds is shown. (A) Correlation distribution of all nearest neighbors to their target voxels before and after two-step filtering. (B) Adaptive thresholds based on temporal and spatial correlation of three scans. Bars represent mean ± SEM, ***p < 0.001.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4195315&req=5

Figure 4: Evaluation of adaptive thresholds is shown. (A) Correlation distribution of all nearest neighbors to their target voxels before and after two-step filtering. (B) Adaptive thresholds based on temporal and spatial correlation of three scans. Bars represent mean ± SEM, ***p < 0.001.
Mentions: Figure 4A shows the temporal correlation distribution of all nearest neighbors to their target voxels of one subject. The right tail of the distribution contains high correlations from highly homogeneous functional regions. These high correlations should be preserved for the estimation of Ta because they represent the correlations from the same functional regions. On the contrary, the left tail contains negative and low correlations, which should be excluded for the estimation of Ta, because they may represent voxels located near the boundary of functional regions, but not in the same region as the target voxel. The first filtering eliminates all negative and most low correlations (shown in purple in Figure 4A), and the second filtering further reduces the number of low correlations and increases the value of Ta.

Bottom Line: In this study, we designed a data-driven method to reduce this impact of the region size in degree metrics by (1) disregarding all self-connections among voxels within the same region and (2) regulating connections from voxels of other regions by the sizes of those regions.In addition, our modified method improved test-retest reliability of degree metrics as well as the sensitivity in group-statistic comparisons.As a promising new tool, our method may reveal network properties that better represent true brain architecture without compromising its data-driven advantage.

View Article: PubMed Central - PubMed

Affiliation: Queensland Brain Institute, The University of Queensland , St Lucia, QLD , Australia.

ABSTRACT
Degree, which is the number of connections incident upon a node, measures the relative importance of the node within a network. By computing degree metrics in voxel-wise functional brain networks, many studies performed high-resolution mapping of brain network hubs using resting-state functional magnetic resonance imaging. Despite its extensive applications, defining nodes as voxels without considering the different sizes of brain regions may result in a network where the degree cannot accurately represent the importance of nodes. In this study, we designed a data-driven method to reduce this impact of the region size in degree metrics by (1) disregarding all self-connections among voxels within the same region and (2) regulating connections from voxels of other regions by the sizes of those regions. The modified method that we proposed allowed direct evaluation of the impact of the region size, showing that traditional degree metrics overestimated the degree of previous identified hubs in humans, including the visual cortex, precuneus/posterior cingulate cortex, and posterior parietal cortex, and underestimated the degree of regions including the insular cortex, anterior cingulate cortex, parahippocampus, sensory and motor cortex, and supplementary motor area. However, the locations of prominent hubs were stable even after correcting the impact. These findings were robust under different connectivity thresholds, degree metrics, data-preprocessing procedures, and datasets. In addition, our modified method improved test-retest reliability of degree metrics as well as the sensitivity in group-statistic comparisons. As a promising new tool, our method may reveal network properties that better represent true brain architecture without compromising its data-driven advantage.

No MeSH data available.