Regional radiomics similarity networks (R2SNs) in the human brain: Reproducibility, small-world properties and a biological basis

A structural covariance network (SCN) has been used successfully in structural magnetic resonance imaging (sMRI) studies. However, most SCNs have been constructed by a unitary marker that is insensitive for discriminating different disease phases. The aim of this study was to devise a novel regional radiomics similarity network (R2SN) that could provide more comprehensive information in morphological network analysis. R2SNs were constructed by computing the Pearson correlations between the radiomics features extracted from any pair of regions for each subject (AAL atlas). We further assessed the small-world property of R2SNs, and we evaluated the reproducibility in different datasets and through test-retest analysis. The relationships between the R2SNs and general intelligence/interregional coexpression of genes were also explored. R2SNs could be replicated in different datasets, regardless of the use of different feature subsets. R2SNs showed high reproducibility in the test-retest analysis (intraclass correlation coefficient > 0.7). In addition, the small-word property (σ > 2) and the high correlation between gene expression (R = 0.29, p < 0.001) and general intelligence were determined for R2SNs. Furthermore, the results have also been repeated in the Brainnetome atlas. R2SNs provide a novel, reliable, and biologically plausible method to understand human morphological covariance based on sMRI.


S01: Definitions of the radiomics features
A total of 47 MRI imaging features, including intensity and textural features, were extracted in this study (Aerts et al., 2014). We provided this information in previous studies (Feng et al., 2018;Zhao et al., 2020). To maintain the integrity of this study, we also list it here. Table S1. Intensity features describe the distribution of voxel intensities within an MRI image through commonly used and basic metrics. Because some intensity features have a close relationship with the voxel intensity, such as the maximum and mean, we first normalized those features by the image intensity using the min-max method in each center.

Symmetry of intensity values in an image
Standard Deviation Measures the homogeneity of the intensity value distribution in an image The spread or variation around the mean (sum of squares) X denotes the three-dimensional image matrix. N is the number of voxels. P is the first-order histogram with Nl discrete intensity levels. X ̅ is the mean of X. The number of histogram bins is 100. Table S2. Textural features describe the patterns or spatial distribution of voxel intensities.

Image feature Equation Definition
Textural features (33) Cluster Tendency Represents the similarity of intensity values in an image P(i, j) is the co-occurrence matrix for an arbitrary δ and α is the number of discrete intensity levels in the image ( , | ) is the ( , )th entry in the given run-length matrix for a direction

Nr is the number of different run lengths
Np is the number of voxels in the image u is the mean of P(i, j) is the marginal row probabilities is the marginal column probabilities μ x is the mean of p x μ y is the mean of p y σ x is the standard deviation of p x σ y is the standard deviation of p y

S02: Supplementary method for R2SN construction
We first performed a common min-max method to normalize the radiomics feature among different brain regions in individuals, and the redundancy feature was defined as those features that had a high correlation with other features (R>0.9). Briefly, we computed the correlation between the i th (i=1.2…47) feature and the rest of the features, and the feature was redundant if an R value>0.9 appeared once or more. These superfluous features were removed before subsequent analysis. As a result, a final feature matrix of 25 × 90 for each subject was obtained for further analysis (Table1). The detailed steps of the R2SN construction are shown in Figure S1. Figure S1. The detailed steps of the R2SN construction.

S03: Supplementary results for the replicability of the result in Brainnetome Atlas
Overall, a high consistency was found in any two mR2SNs that were constructed by different datasets (1000 times randomly simulation), and the Pearson coefficient ranged from 0.9997 to 1 ( Figure S1b). In addition, a high consistency was also obtained by mR2SNs, which were constructed with different numbers of features (20 features for randomly selected and all features), and the R-value ranged from 0.90 to 1 ( Figure S1c). More importantly, the R2SN achieved a high ICC value (ICC > 0.7) within more than 95% of edges by test-retest analysis ( Figure S1d). The clustering coefficients, shortest path length, and sparseness are shown in Figure S2a-c. In addition, lambda (λ) indicates the ratio of the shortest path length of R2SN and the random network, and gamma (γ) indicates the ratio of the clustering coefficient of R2SN and the random network. As a result, the value of λ was close to 1 ( Figure S2d), the value of γ was significantly larger than 1 (Figure S2e), and the small world index sigma (σ) was also significantly larger than 1 by different thresholds of binarization (from 0.5 to 0.75, and the step size =0.01) ( Figure S2f). To further explore the network structure, we also computed the hub nodes of the R2SN. In this study, the hub nodes were defined as those nodes that had many connections with other nodes. As Figure S4 shows, the hub nodes of the R2SN based on the AAL atlas were located mainly in the PrCG, SFGdor, ORBsup, MFG, ORBmid, IFGoperc, IFGtriang, SMA, CAL, CUN, SOG, PoCG, SMG, and HIP (Degree>10). Meanwhile, the hub nodes of the R2SN based on the BN atlas were located mainly in the A8m, A6m, A14m, A11m, A11m, A12_47l, A12_47l, A4tl, A6cvl, and A1_2_3tonLa. The PrCG, SFG, MFG, IFG and PoCG have been reported as the hub regions in previous studies (He et al., 2007;Tijms et al., 2012;Kim et al., 2016). Figure S4. The hub nodes of the R2SN are based on the AAL atlas and BN atlas.
We also computed the correlation between the degree of the node and the cognitive score. PrCG, ORBsup, ORBmid, IFGoperc, SMA, CAL, CUN, SOG, PoCG, SMG, and HIP showed a significant correlation with the cognitive score. The hub regions are important for the connectome's efficiency and are preferentially affected by neurodegenerative disease (Seidlitz et al., 2018). These hub regions require more attention in future studies.
The mean connective strengths of each node of mR2SN and mGSN are shown in Figure S5a (mR2SN), Figure   S5b (mGSN), and Figure S5c. We also computed the similarity between mR2SN and mGSN (edge-based measure), and the Euclidean distance between each pair of ROIs was employed as a concomitant variable. A significant correlation can be found between the two networks with R=0.25 (P<0.001), meaning that the brain region with high morphometric similarity also tended to have a high transcriptional similarity of the gene. The Pearson correlation showed that approximately 1% of connections have a significant correlation with the fluid intelligence score (Bonferroni-corrected P<0.05, with N=30315) ( Figure S6a). As Figure S6 shows, the neighbor degree of 7 nodes ( Figure S6b) The Pearson correlation showed that approximately 9% of connections have a significant correlation with fluid intelligence score (Bonferroni-corrected P<0.05, with N=30315) ( Figure S6f). As Figure S6 shows, the neighbor degree of 59 nodes ( Figure S6g  We also computed the correlations between betweenness centrality/shortest path length and gF/FICA. However, weak significance was obtained as shown in Figure S7. Figure S7. The correlation between R2SN and cognitive ability. The brain regions based on the AAL atlas in which (a) the between betweenness centrality and (b) the shortest path length showed a significant association with the gF score (Bonferroni-corrected P<0.05); the brain regions based on the AAL atlas in which (c) the between betweenness centrality and (d) the shortest path length showed a significant association with the FICA score (Bonferroni-corrected P<0.05). The brain regions based on the BN atlas in which (e) the between betweenness centrality and (f) the shortest path length showed a significant association with the gF score (Bonferroni-corrected P<0.05); and the brain regions based on the BN atlas in which (g) the between betweenness centrality and (h) the shortest path length showed a significant association with the FICA score (Bonferroni-corrected P<0.05).