Mapping Individual Brain Networks Using Statistical Similarity in Regional Morphology from MRI

Abstract

Representing brain morphology as a network has the advantage that the regional morphology of 'isolated' structures can be described statistically based on graph theory. However, very few studies have investigated brain morphology from the holistic perspective of complex networks, particularly in individual brains. We proposed a new network framework for individual brain morphology. Technically, in the new network, nodes are defined as regions based on a brain atlas, and edges are estimated using our newly-developed inter-regional relation measure based on regional morphological distributions. This implementation allows nodes in the brain network to be functionally/anatomically homogeneous but different with respect to shape and size. We first demonstrated the new network framework in a healthy sample. Thereafter, we studied the graph-theoretical properties of the networks obtained and compared the results with previous morphological, anatomical, and functional networks. The robustness of the method was assessed via measurement of the reliability of the network metrics using a test-retest dataset. Finally, to illustrate potential applications, the networks were used to measure age-related changes in commonly used network metrics. Results suggest that the proposed method could provide a concise description of brain organization at a network level and be used to investigate interindividual variability in brain morphology from the perspective of complex networks. Furthermore, the method could open a new window into modeling the complexly distributed brain and facilitate the emerging field of human connectomics.

Conflict of interest statement

Competing Interests: The authors have declared that no competing interests exist.

Figures

Fig 1. General workflow for the construction of an individual morphological network using gray matter measurements from MRI.

(1) The estimation of gray matter volume as a morphological measure using a routine VBM procedure. (2) Brain parcellation with the AAL atlas and the estimation of morphological distribution for each region. (3) Repeated quantification of the similarity between morphological distributions for pairs of regions and formation of the similarity matrix by filling in corresponding similarity values. (4) Extract the binarized matrix at a specific sparsity threshold. (5) Represent individual brain network as a graph. (6) Calculate the network metrics (e.g., γ, λ, and σ). Note that, in this study steps 4–6 were repeated across a range of different sparsity thresholds, from 10% to 40% with an interval of 1%. HIP: Hippocampus; FFG: Fusiform gyrus; L: left; R: right; KLS: Kullback-Leibler divergence-based similarity; MRI: magnetic resonance imaging; VBM: voxel-based morphometry; AAL: automated anatomical labeling.

Fig 2. The averaged map of the connectivity matrices.

Red and blue color indicates high and low similarity between regions, respectively. Main diagonal (i.e., self-connection) is indicated in white and excluded from following analyses. L, Left; R, Right.

Fig 3. Coefficient of variation (CV) map of the connectivity matrices.

Red and blue color indicates high and low dispersion of that connection across participants, respectively. Most of the connections possessed relative low CV and in particular the connections with relative high similarities showed low CV, suggesting relative high consistency across subjects. Main diagonal (i.e., self-connection) is indicated in white and excluded from following analyses. L, Left; R, Right.

Fig 4. Small-world properties of the morphometric networks as a function of network sparsity thresholds.

The error bar indicates the standard deviation.

Fig 5. Spatial distribution of hubs within the morphometric network.

(A) Three examples of the spatial distributions for betweenness. (B) The similarity (left) and uniqueness (right) of individual spatial distributions. (C) Hubs identified in this study. L: left; R: right. PreCG: precentral gyrus; IFGoper: inferior frontal gyrus opercularis; IFGtri: inferior frontal gyrus triangularis; SMA: supplementary motor area; SMG: supramarginal gyrus; PCUN: precuneus; TPOsup: superior temporal gyrus of temporal pole; MTG: middle temporal gyrus; ITG: inferior temporal gyrus.

Fig 6. The Intraclass Correlation Coefficient (ICC) map of the connectivity matrices.

More than 97% of the edges showed excellent reliability (i.e., ICC > 0.75).

Fig 7. Test-retest reliability with intraclass correlation coefficient (ICC) for each of the network metrics as a function of network sparsity thresholds.

The smallest ICC at various sparsity thresholds for each network metric is listed for the following: Cp: min ICC = 0.605, p = 0.0018; Lp: min ICC = 0.518, p = 0.0060; γ: min ICC = 0.636, p = 0.0020; λ: min ICC = 0.391, p = 0.035; σ: min ICC = 0.611, p = 0.00092; Eloc: min ICC = 0.377, p = 0.047; Eg: min ICC = 0.611, p = 0.0014; meanBet: min ICC = 0.485, p = 0.012.

Fig 8. Age-related changes in each of the network metrics at a predefined sparsity threshold (i.e., 23%).

Fig 9. Age-related changes in each of the network metrics over the range of sparsity thresholds.

The star markers (correlations falling outside the shaded area) indicate significant correlations (p