Texture-based brain networks for characterization of healthy subjects from MRI

Brain networks have been widely used to study the relationships between brain regions based on their dynamics using, e.g. fMRI or EEG, and to characterize their real physical connections using DTI. However, few studies have investigated brain networks derived from structural properties; and those have been based on cortical thickness or gray matter volume. The main objective of this work was to investigate the feasibility of obtaining useful information from brain networks derived from structural MRI, using texture features. We also wanted to verify if texture brain networks had any relation with established functional networks. T1-MR images were segmented using AAL and texture parameters from the gray-level co-occurrence matrix were computed for each region, for 760 subjects. Individual texture networks were used to evaluate the structural connections between regions of well-established functional networks; assess possible gender differences; investigate the dependence of texture network measures with age; and single out brain regions with different texture-network characteristics. Although around 70% of texture connections between regions belonging to the default mode, attention, and visual network were greater than the mean connection value, this effect was small (only between 7 and 15% of these connections were larger than one standard deviation), implying that texture-based morphology does not seem to subside function. This differs from cortical thickness-based morphology, which has been shown to relate to functional networks. Seventy-five out of 86 evaluated regions showed significant (ANCOVA, p < 0.05) differences between genders. Forty-four out of 86 regions showed significant (ANCOVA, p < 0.05) dependence with age; however, the R2 indicates that this is not a linear relation. Thalamus and putamen showed a very unique texture-wise structure compared to other analyzed regions. Texture networks were able to provide useful information regarding gender and age-related differences, as well as for singling out specific brain regions. We did not find a morphological texture-based subsidy for the evaluated functional brain networks. In the future, this approach will be extended to neurological patients to investigate the possibility of extracting biomarkers to help monitor disease evolution or treatment effectiveness.

The use of functional brain networks to characterize cognitive processes in both healthy subjects and patients with various neurological diseases has been widely spread (for a review, see e.g. 1 ).Data obtained from techniques such as electroencephalography (EEG) [2][3][4][5][6][7] , magnetoencephalography 8,9 , near-infrared spectroscopy [10][11][12][13] , positron emission tomography (PET) 14,15 and mainly, functional magnetic resonance imaging (fMRI) 16,17 , have been used for this purpose.Indeed, it has been argued that, since most brain functions depend on several areas, the network approach gives a more complete picture of brain processes and function, compared to looking at isolated brain regions (see, e.g., 18 ).
One of the first functional networks found was the default mode network (DMN); discovered by Raichle's group using PET scans 19,20 and later confirmed by Greicius and colleagues using resting-state fMRI 21 .It is composed of a set of regions that are thought to be involved in mind wandering or unfocused mental tasks; these regions are only active during passive rest and are 'turned off ' when an individual engages in an externally goal-directed task.This network has been 'measured' in all sorts of populations, from children to various types of neurological patients, and therefore has been extensively studied.Another important resting-state functional www.nature.com/scientificreports/fmri.wfubmc.edu/ softw are/ PickA tlas).This atlas corresponds to the most cited paper in the neuroimaging field 71 , and therefore, it was deemed as a good first choice for this exploratory study.Future studies may explore this method with other more modern atlases (e.g.Desikan 72 , Destrieux 73 , and Glaser 74 ).
The AAL is in the standard MNI space 75 , therefore, subjects' images were first converted to this space using the SPM12 software (https:// www.fil.ion.ucl.ac.uk/ spm/ softw are/ spm12/).To do this, the anterior commissure was set to be the origin of the reference system of the images, and these were realigned according to the MNI space orientation.Next, the images were segmented into gray matter, white matter, and cerebrospinal fluid; normalized to the standard MNI space, and registered to the AAL atlas.The brain and intracranial volumes were calculated during this step using the UF2C software 76 .Some of these pre-processing steps involved interpolation methods, which can lead to changes in the image's texture.When possible, precautions were taken to minimize this effect, such as employing a 4th-degree B-spline, a smoother interpolation method.Finally, the AAL regions were used as masks to identify the regions of interest (ROIs) in the patients' images.
Texture parameters from each ROI were then extracted from the images, using the gray level co-occurrence matrix (GLCM) approach 77 , implemented through homemade Matlab routines.The GLCM is a symmetrical square matrix, where each element (i, j) represents the (normalized) number of pixel pairs in the ROI with gray levels i and j, separated by a given distance in a given direction.Usually, the GLCM is computed for 2D images, for distance values ranging from 1 to 5 pixels, and directions horizontal, vertical, and diagonals (45° or 135°).For 3D images, texture parameters from the corresponding 2D slices are combined using simple or weighted average [78][79][80] .
In this work, an isotropic GLCM (i.e., not considering any particular direction), computed directly from the 3D images, was used.For this, a cubic layer centered in the reference voxel was defined, as outlined in Fig. 1 for a distance of 2 voxels (represented in yellow).In the ROIs border, only voxels contained in the ROI were accounted for the GLCM.
Distance values ranging from 1 to 5 voxels were used so that, for each ROI, five 3D GLCMs were obtained.Initially, GLCMs were calculated for 128 gray levels.This led to very sparse GLCM matrices (~ 90% zero entries for the 1-voxel distance to ~ 87% zero entries for the 5-voxel distance).Therefore, the number of gray levels was lowered to 64, and finally to 32, which decreased sparsity while still keeping some gray-level information (~ 86 to ~ 83% zero entries for the 1 and 5-voxel distance respectively).Although reducing the number of gray levels decreases the sparsity of the GLCM, this also reduces its size, and consequently, the information contained in it.Therefore, it is necessary to balance the trade-off between the information content and the matrix sparsity.
At first, texture parameters from all 116 AAL regions were used to obtain the brain networks.However, the small size of some of these regions led again to an increased number of zeros in their respective GLCMs.Therefore, a selection criterion was applied: regions had to be larger than 900 voxels or, if they were part of a homologous pair, at least one of the regions had to be larger than 1000 voxels.This resulted in a set of 86 AAL regions (these are shown in Table S1 of the Supplementary Material).
From each GLCM, 11 texture parameters were computed: uniformity, contrast, correlation, variance, homogeneity, entropy, sum average, sum variance, sum entropy, difference variance, and difference entropy 77 .Therefore, since there were five GLCMs (for distances from 1 to 5 pixels) for each anatomic region, this resulted in a total of 55 parameters per region.Since these parameters have largely different ranges, they were all normalized to be within the [0, 1] range.Finally, a feature vector consisting of these 55 parameters was used to characterize each network node (ROI) of a given individual.Figure 2 shows an explanatory scheme for the generation of the feature vectors.
To compare the feature vectors from every pair of nodes, the inverse of the Euclidean distance was calculated, and those values were normalized to the [0,1] interval.In this way, nodes with more similar texture values (and Figure 1.Example of a cubic layer (voxels in yellow) used to compute an isotropic gray level co-occurrence matrix (GLCM) directly from a 3D image.therefore smaller Euclidian distance between them) would have a stronger link.These values were then employed to generate the brain network links of the weighted graphs of the subjects.
Once the networks were generated, the next step was to compare them.This was done through mathematical measures that describe the topological characteristics of the networks-the network measures 53 .Five network measures were computed for each brain network generated: strength (ST), betweenness centrality (BC), eigenvector centrality (EC), clustering coefficient (CC), and local efficiency (LE).Regarding the measures' choice, this was driven by two factors: easiness of interpretation and popularity of the measure.The chosen measures have both been used in other works on brain networks using graphs [81][82][83] and have a fairly straightforward interpretation.
Four studies were conducted to explore the potential of texture-based networks to characterize a given population.These were: 1. Analysis of structural texture-based networks and comparison with well-established functional networks-For the five functional networks evaluated (DMN, sensory-motor, attention, visual, and subcortical), the mean value of each connection across all individuals and the mean value over all connections and individuals were computed.Then, for each network, the mean connection values (over individuals) were compared with the mean value of all connections.Table S2 in the Supplementary Material shows all the regions belonging to each functional network selected.2. Comparison between texture-based networks obtained from male and female populations-For each individual brain network, five network measures were extracted-ST, BC, EC, CC, and LE.A statistical test of the ANCOVA type was performed, in which gender was employed as the independent variable and the dependent variables were set as the values of the network measure of each region.Age, brain volume, and intracranial volume were selected as covariates and a Bonferroni correction for multiple comparisons was applied.3. Analysis of age dependence of texture-based networks-The same five network measures used to evaluate gender differences were extracted from each individual network to investigate age dependence-ST, BC, EC, CC, and LE.A statistical test of the ANCOVA type was performed, in which age was selected as the independent variable and the dependent variable was the network measure of each region.Sex, brain volume, and intracranial volume were selected as covariates, and correction for multiple comparisons was performed using Bonferroni.4. Analysis of network measure variation for different brain regions-The same five network measures were extracted from each individual network-ST, BC, EC, CC, and LE.For each network measure, the mean of each region's network measure over the entire population was calculated, as well as the global mean (over all regions and individuals).Then, the difference between each region's mean and the global mean was computed.

Analysis of structural texture-based networks and comparison with well-established functional networks
The similarity measure was calculated between all regions in the network and the mean of these measures was computed.We evaluated the number of edges between the N nodes of each functional network as N*(N-1)/2.We then looked at the value of the similarity measure corresponding to each edge and compared it to the previous mean.Table 1 shows the number of edges that are greater than the mean, as well as the number of edges that are at least one standard deviation greater than the mean.Although three out of the five evaluated functional networks presented around 70% of their texture-based connections stronger than the mean connection value, when we look at how strong these values were, we see that only between 7 and 15% of the values were actually larger than one standard deviation.Therefore, this seems to indicate that there isn't a morphological texture-based subsidy for these functional networks.This result differs from other results in the literature, such as the study by Park and colleagues, who found a high agreement between brain parcellations based on fMRI networks and cortical thickness networks for the medial frontal cortex 84 , or the study by Chen and coworkers, who found that functional domains such as auditory/language, strategic/executive, sensorimotor, visual, and mnemonic processing had a close overlap with cortical thickness network modules 43 .However, these differences with our results can be explained because although texture measures are expected to reflect the underlying tissue structure, these are usually evaluated from volumetric ROIs that encompass sulci and gyri, and therefore manifest different morphological properties than those disclosed by cortical thickness.

Comparison between texture-based networks obtained from male and female populations
Among the 86 regions selected, two regions showed significant differences (ANCOVA, p < 0.05) among male and female populations for all five network measures, and 26 regions showed significant differences for four network measures.The remaining regions presented significant differences as follows: 9 regions for three network measures, 10 regions for two network measures, 28 regions for a single network measure, and 11 regions presented no significant differences for any network measure.Table 2 shows the regions for which at least four network measures were significantly different (ANCOVA, p < 0.05) among male and female populations.Table S3 in the Supplementary Material shows all the regions with at least one significantly different network measure.
In summary, we obtained network measures that were significantly different among men and women for most regions (75 out of 86).This indicates a strong relationship between sex and texture connections.In decreasing order, 62 regions showed significant differences for the EC, 43 for ST, 41 for both CC and LE, and 2 for BC.
The regions for which all network measures were significant were the right triangular part of the inferior frontal gyrus (Frontal Inf Tri) and the left supplementary motor area (Supp Motor Area).Along with the opercular part, the right Frontal Inf Tri is associated with Broca's area, which is involved in speech production.Specifically, Frontal Inf Tri is involved with the semantic processing of language and non-verbal communication such as gesticulation and facial expression 85 .Supp Motor Area has a role in the preparation of voluntary movements and the temporal organization of sequential movements 86 .For these regions, ST, EC, CC, and LE had larger values for men than for women.For BC, Supp Motor Area had larger values for men than for women, while the opposite happened for Frontal Inf Tri.
Since ST is the sum of connections of a node, this measures how strongly connected this node is.EC measures a node's influence in the network, while CC points to a node's tendency to form clusters.Both LE and CC analyze the shortest paths that connect nodes.LE measures the minimum number of nodes necessary for connecting a pair of nodes, while CC measures the number of shortest paths a node is part of.In the context of texture, this implies that both the right Frontal Inf Tri and left Supp Motor Area have a texture that is highly similar to the surrounding regions, forming a texture cluster with them.These regions, however, are not similar in texture among themselves.This happened more for men than for women.Conversely, BC measures how much a node behaves as a hub, i.e., how much it intermediates relationships among different parts of the network.Thus, considering texture properties, the right Frontal Inf Tri behaves more as hubs for women than for men, while the left Supp Motor Area behaves more as hubs for men than for women.
In addition, nine pairs of homologous regions (lobules IV and V of the cerebellar hemisphere (Cerebellum 4 5), crus II of the cerebellar hemisphere (Cerebellum Crus2), opercular part of the inferior frontal gyrus (Frontal and three regions in the left hemisphere (medial superior frontal gyrus (Frontal Sup Medial), orbital part of the superior frontal gyrus (Frontal Sup Orb), and Precuneus (Precuneus)) were significant for four network measures.As previously, for these regions (for the left and right Cerebellum 4 5), ST, EC, CC, and LE had larger values for men (women) than for women (men).This indicates these regions share texture similarities with other regions, with which they tend to cluster.Regarding gender, this suggests the texture of these regions (of the Cerebellum 4 5) is more similar to their neighboring regions for male (female) individuals rather than females (males).These regions (the Cerebellum 4 5) also have a higher tendency to form texture clusters for males (females) than for females (males).

Analysis of age dependence of texture-based networks
Among the 86 regions selected, one region showed significant differences (ANCOVA, p < 0.05) for the independent variable age (see Table 3) for all five network measures, and 11 regions showed significant differences for four network measures.There were significant differences for three network measures in two regions, for two network measures in five regions, for a single network measure in 25 regions, and no significant difference for any network measure in 42 regions.Table 4 shows the AAL regions with at least four significant R 2 values, obtained for the linear regression of the network measures considering the independent variable age.Table S4 (in the Supplementary Material) shows the AAL regions with at least one significant R 2 value.
Forty-four (out of 86) regions had network measures that presented a significant dependence on age.In decreasing order, there were significant differences for EC in 40 regions, ST in 17, CC and LE in 12, and BC in 9. From these, 12 regions had at least four significant network measures (out of five tested measures).There were four pairs of homologous regions (caudate nuclei (Caudate), insula (Insula), superior occipital gyri (Occipital Sup), and thalami (Thalamus)), one region in the right hemisphere (triangular part of the inferior frontal gyrus Table 2. Regions that had at least four significant (ANCOVA, p < 0.05) network measures regarding differences between male and female populations.P-values smaller than 0.05 are marked in bold font.www.nature.com/scientificreports/(Frontal Inf Tri)), and three regions in the left hemisphere (lobule VI of cerebellar hemisphere (Cerebellum 6), anterior cingulate & paracingulate gyrus (Cingulum Ant), and dorsolateral superior frontal gyrus).
The two largest values of R 2 obtained were 0.503 for the right hippocampus and 0.502 for the right supramarginal gyrus (see Table S4 in the Supplementary Material), all referring to the BC network measure.However, despite the corresponding p-values being significant (p < 0.05), the plot of age versus BC (Fig. 3) shows this result is not meaningful, since the majority of data points are zero.
On the other hand, the other network measures did show meaningful results (Fig. 4), but the R 2 values obtained in these cases were smaller than 0.5, meaning that the relationship between texture and age is not linear.For the ST, CC, and LE, all regions showed an increase with age.For the EC, only the left Frontal Sup showed a decrease with age, while all the other regions showed an increase.
An increase with age of the ST, CC, EC, and LE suggests the texture of these regions becomes more similar to their neighboring regions, while at the same time, developing a progressively higher tendency to form texture clusters.This increase in their similarity can be a result of brain degeneration, which agrees with the reported age-related reduction in brain volume 87 .Thus, it was not possible to establish a linear relationship between texture and age.Yet, this does not mean that there is no relation between these factors because several significant values were obtained.

Conclusion
In this work, we attempted to use texture analysis to generate brain networks based on structural properties of magnetic resonance images.To the best of our knowledge, this is the first work to use image texture properties to build brain networks.
We sought to evaluate the structural connections between regions of well-established functional networks, changes in the brain networks due to gender, and the dependence of the networks with age.We also employed graph theory to attempt to characterize healthy individuals.We were able to extract meaningful information from the texture-based networks, thus proving its usefulness.
The comparison study between the structural texture-based networks and some well-established functional networks did not allow to establish a morphological relation between the texture-based networks and the functional networks.
The gender comparison study showed significant differences in the network measures for most (75/86) regions investigated.In fact, network measures obtained suggested higher similarity among neighboring regions and a higher tendency to form texture clusters for male than for female individuals, except for the Cerebellum 4 5, which presents higher similarity among neighboring regions and a higher tendency to form texture clusters for female than for male individuals.Also, this analysis found that the left Supp Motor Area (right Frontal Inf Tri) was more likely to behave as a texture hub for male (female) individuals.
Regarding the age study, around half (44/86) of the selected regions yielded significant nonlinear relations with age.Network measure changes with age for the majority of these regions showed an increase in texture similarity among them, possibly related to degeneration of the underlying tissues.www.nature.com/scientificreports/ The graph metrics' study showed that the thalamus and the putamen display weaker connections to other regions, but they function as hubs.Therefore, they have, texture-wise, quite a unique structure, which seems appropriate, due to their respective roles.
All analyses performed in this work were based on T1-weighted structural magnetic images.For future studies, T2-weighted resonance images or even multimodal images could be analyzed.This work used the cooccurrence matrix method for texture analysis, but future works could employ other techniques (e.g., wavelets 93 or local binary patterns 94,95 ).The atlas chosen for the parcellation contained many small regions-due to the great number of subdivisions-and could be replaced by an atlas with fewer (and therefore bigger) regions.On the other hand, the atlas-based parcellation could also be replaced with a learning-based parcellation.Furthermore, machine learning algorithms could be applied to the data obtained to extract meaningful information.This work's methodology was applied to healthy individuals.Further investigation is required to determine if it is possible to employ these methods for patients with anatomical alterations.Indeed, the next step for this work is to apply these techniques to patients with different brain diseases/conditions (e.g.Alzheimer's disease), to investigate the ability of the proposed method to produce biomarkers for these pathologies.

Figure 2 .
Figure 2. Scheme to obtain the feature vectors.

Figure 3 .
Figure 3. Age versus BC for the left Occipital Sup.The plot shows that even though BC presented values of R 2 greater than 0.250 and p-values lower than 0.05, those are not meaningful since only a few of the data points are non-null.

Figure 4 .
Figure 4. Age versus network measure plots.ST dispersion plot (a) for the right Caudate with linear adjustment (red line).Estimated marginal means plots of ST (b) for the right Caudate, EC (c) for the right Insula, and CC (d) and LE (e) for the right Thalamus.

Table 1 .
Number of possible connections between the regions comprising each evaluated brain functional network; and number and percentage of connections with a mean value of similarity measure between regions greater than the mean value of all connections.

Table 4 .
R squared and p-values obtained from ANCOVA for the independent variable age for the regions that had at least four significant (p < 0.05) network measures.p-values smaller than 0.05 aremarked in bold font, R 2 values greater than 0.250 are marked in italic font.

Table 5 .
Difference between each region's network measure and the global mean of the measure (in standard deviation units) for the individual networks.