A total of 65 healthy subjects (29 males, average age: 17.17 ± 1.59 years old) were recruited in this study. All subjects were Chinese native speakers, with no history of neurological or psychological illness and no abnormal findings in conventional brain MRI, corresponding to Diagnostic and Statistical Manual of Mental Disorders (DSM-IV). All subjects were right-handed measured with the Edinburgh handedness inventory (Oldfield, 1971). The study was approved by the Human Research Ethics Committee of Shandong University School of Medicine. Written informed consents were obtained from all the participants as well as their parents.

A version of ANT devised by Fan et al. (2005) was used as a cognitive task to examine the efficiency of the alerting, orienting, and EC networks involved in attention. Participants were instructed to press a button as accurately and quickly as possible to identify the direction of the target, which was a leftward or rightward arrow at the center and flanked on either side by two arrows in the same direction (congruent condition), or the opposite direction (incongruent condition). The target and flankers were presented until the participant made a response or 2000 ms elapsed. A cue (an asterisk) was presented for 200 ms before the target appeared. The task used three cue conditions: no cue (baseline), center cue (at the fixation for alerting), and spatial cue (at the target location for alerting and orienting).

Each subject performed six blocks in this experiment, each block lasting 5 min 42 s and consisting of 36 trials plus 2 buffer trials at the beginning. In each block, the six trial types (three cue conditions by two target conditions) were presented in a predetermined counterbalanced order. All subjects were trained before the formal experiment. Stimulus presentation and behavioral response collection were performed using E-Prime (Psychology Software Tools, Pittsburgh, PA, USA) on an experimental control computer.

The total accuracy of each subject was calculated and those with high error rates (>20%) should be excluded from this study. The trials with incorrect responses or with response time (RT) longer than 1500 ms or shorter than 200 ms were also excluded to avoid the influence of the outliers. In addition, we need remove responses following erroneous ones to avoid post-error slowing effect. Since RTs were not normally distributed, we used median RT per condition as raw scores (Adolfsdottir et al., 2008). The accuracy for each of the six trial types was also calculated. Finally, instead of the conventional subtraction measure (Fan et al., 2005, 2007), we used ratio scores of alerting, orienting, and EC to define the effects of three attention networks. The formulas were as follows:

The executive attention effect was calculated by subtracting the mean RT of congruent conditions from the mean RT of incongruent conditions; thus, a higher EC score reflected a relatively poorer executive function.

MR imaging data were acquired using a 3.0 Tesla GE Signa scanner (General Electric Medical Systems, Milwaukee, WI, USA) with a standard eight-channel head coil. DTI images were obtained using spin-echo, single shot echo planar imaging sequence [TR/TE = 14000/75.1 ms, acquisition matrix = 96 × 96, field of view (FOV) = 250 mm, slice thickness = 2.6 mm, no gap]. The DTI scheme included 30 non-linear diffusion gradients directions with b = 1000 s/mm² and 3 non-diffusion-weighted images (b = 0 s/mm²). Array spatial sensitivity encoding technique (ASSET) was adapted with an acceleration factor of 2 to reduce acquisition time with less image distortion from susceptibility artifacts (Yu et al., 2008). For each subject 56 axial slices were collected and the diffusion sequence was repeated twice to increase signal-to-noise ratio (SNR). At the end of the DTI scans, a three-dimensional volume spoiled gradient-echo (SPGR) pulse sequence with 174 slices (TR = 6.5 ms, TE = 2.0 ms, matrix = 256 × 256, FOV = 256 mm, FA = 15°, slice thickness = 1.0 mm, no gap) was used to acquire the structural images for anatomical reference.

The DTI data were pre-processed using FSL (University of Oxford, UK) to correct for eddy currents and head motion. After two acquisitions being averaged, the averaged images were masked to remove the skull and non-brain tissue using the FSL Brain Extraction Tool (BET) (Smith, 2002). Afterward, the fractional anisotropy (FA) was calculated using the diffusion tensor analysis toolkit (FDT; Smith et al., 2004).

The whole cerebral cortex was parcellated into 90 areas using the automated anatomical labeling (AAL) template (Tzourio-Mazoyer et al., 2002) where each region represented a WM network node (Bullmore and Sporns, 2009). In the parcellation process, each individual T1 image was linearly coregistered to the b₀ image in the native space and then non-linearly mapped to the International Consortium for Brain Mapping 152 (ICBM 152) T1 template in the Montreal Neurological Institute (MNI) space, resulting in a non-linear transformation (Γ). Finally, one subject-specific AAL mask in the DTI native space was generated, using the inverse transformation (Γ^-1) to the AAL template in the MNI space (Figure 1).

Network edges were defined as the interregional anatomical connections determined using DTI deterministic tractography. Fibers with two end-points located in their respective masks were considered as edges to link corresponding pairs of nodes. The whole-brain fiber tracking was performed via Fiber Assignment by Continuous Tracking (FACT) algorithm (Mori et al., 1999), with the FA threshold of 0.1 and tracking turning angular threshold of 30° between two connections. PANDA (Cui et al., 2013) was used to generate a FA-weighted matrix M(90*90) with cell M(i,j) representing the averaged FA of linking fibers between node i and node j. All cells of M that exceeded the threshold of 0.1 were set to 1 and other cells that did not exceed 0.1 were set to 0. This resulted in a fully connected binary matrix B (van den Heuvel et al., 2009; Xu et al., 2015).

With the binary connectivity matrices (B), topological properties of structural brain networks were obtained using Brainnetome fMRI Toolkit¹, including small-worldness, global efficiency, local efficiency and regional efficiency.

SPSS (version 20, IBM Inc., Chicago, IL, USA) was used for statistical analyses. To explore the relationship between global brain network topology (the global efficiency, E_glob, and the local efficiency, E_loc) and behavioral performances of the three attentional components, Pearson’s linear correlations were separately performed on each network metric, using gender and age as covariates. We then apply Bonferroni correction for multiple testing to the results.

To identify specific brain regions involved in alerting, orienting and EC, multiple linear regression analyses for nodal efficiency of all 90 cortical regions and three attentional effects were performed with SPSS. As the problem of multicollinearity occurred due to correlation between the predictors, we first used multicollinearity diagnostics to assess the independence of parameters and to test whether two or more variables were highly correlated. Then, we apply ninety blocks of independent variables in a step-wise fashion. Statistical threshold at the level of p < 0.05 was considered significant.

The average accuracy of ANT performance was high (0.97 ± 0.02) and no one was excluded from the study, indicating that all participants understood the instructions and were able to perform the task reliably.

Table 1 showed the three effect ratio scores of ANT components and their correlations. Independent-sample t-tests revealed no significant gender differences in alerting, orienting and EC effects. Only a significant correlation (r = -0.35, p = 0.005) between the alerting and orienting performance was found, after controlling for gender and age.

The structural brain networks were constructed for each subject, with 90 nodes and 4005 possible edges and the global topological properties of structural networks spanning the entire brain were then examined. The constructed structural brain networks showed small-world characteristics expressed as γ > 1 and λ ≈ 1 (σ = 1.76 ± 0.10, γ = 1.94 ± 0.12, λ = 1.10 ± 0.01). We found a significant correlation (r = -0.37, p = 0.003, Bonferroni corrected) between the EC effect and global efficiency (Figure 2; Table 2). No significant associations were observed between global network properties and ratio scores of alerting and orienting.

To further identify which specific brain regions are related to any of the three attention variables, multiple linear regression analysis was performed between nodal characteristics and ratio scores of alerting, orienting and EC for individual brain networks. The correlations between these brain regions were low and no significant multicollinearity were found. Some regions were observed whose regional efficiency significantly correlated with the behavioral performances of three components of attention, as shown in Tables 3–5.

Significant correlations between EC component and regional efficiency of a set of brain areas (Table 3; Figure 3A) were discovered, including the left dorsolateral SFG (β = -0.538, p = 0.000), left parahippocampal gyrus (β = 0.292, p = 0.021), lateral thalamus (left: β = 0.268, p = 0.032; right: β = -0.632, p = 0.000), right supplementary motor area (β = 0.216, p = 0.038), lateral pallidum (left: β = -0.304, p = 0.015; right: β = -0.363, p = 0.005), right putamen (β = -0.525, p = 0.000) and right calcarine fissure (β = -0.595, p = 0.000).

As shown in Table 4 and Figure 3B, we observed significant correlations between alerting and regional efficiency of the left thalamus (β = 0.402, p = 0.001), right inferior parietal gyrus (r = 0.250, p = 0.036), left middle temporal gyrus (β = -0.252, p = 0.015) and right lenticular nucleus, Pallidum (β = 0.462, p = 0.002). For the orienting effect, the significant associations (Table 5; Figure 3C) with the regional efficiency were found in the right paracentral lobule (β = 0.376, p = 0.020), right inferior occipital gyrus (β = 0.319, p = 0.003) and left temporal pole (middle temporal gyrus: β = -0.383, p = 0.000).

Graph theoretical analysis of structural and functional neuroimaging data have shown small-world architecture of human brain indicating high clustering of connections as well as shortest path length between brain regions, which in turn, supports the optimal balance between local specialization and global integration (Bullmore and Sporns, 2009). Consistent with these findings, the structural brain networks constructed from DTI data in this study also showed a small-world organization. Based on graph theory, recent neuroimaging studies have measured the relationships between brain network architecture and cognitive functions, like intellectual performance (van den Heuvel et al., 2009) and working memory (Gamboa et al., 2014). However, graph-theoretic studies of attention have mainly focused on functional brain networks, with few studies on structural brain networks from DTI data. Our study on structural brain networks suggested the analogous discoveries with previous studies on functional brain networks (Markett et al., 2014; Xu et al., 2015).

In this study, the global efficiency of WM networks showed a significant positive correlation with the EC performance. Global efficiency reflects the network’s capacity for parallel information transfer between brain regions via multiple parallel paths, and is believed to be the basis of integrated processing for cognitive functions. Higher scores on intelligence tests were found corresponding to a higher global efficiency of both WM networks (Li et al., 2009) and functional brain networks (van den Heuvel et al., 2009). Patients with hypertension (Li et al., 2015b), amnestic mild cognitive impairment (aMCI) (Shu et al., 2012) or traumatic brain injury (TBI) (Caeyenberghs et al., 2014) demonstrated a decreased global efficiency of the WM networks, resulting in decline of the executive function. Lo et al. (2010) found that AD patients with impairments in cognition and memory functions, exhibited a decreased global efficiency of the whole WM networks. Therefore, the executive function tends to rely on a more optimal configuration (small-worldness) of structural brain networks, and a limited capacity to integrate information across brain regions would result in a poor executive performance in cognitive deficits. Additionally, our results provided evidences that network measures could be an efficient way to quantify the structural brain system underlying cognitive functions.

We examined the correlations not only at the whole brain network level, but also at the individual regional level. From this perspective, a multiple linear regression analysis was conducted in the regional efficiency of each brain region to investigate specific regions involved in the three ANT components.

There are limitations in our study. Firstly, we employed deterministic tractography to define the structural network edges. Although this approach has been used in previous neuroimaging studies (Shu et al., 2012; Caeyenberghs et al., 2014), it might lose some existing fibers because of its limitation in tracking the crossing and long-distance fibers. Further studies based on advanced tractography methods (Zalesky and Fornito, 2009; Jeurissen et al., 2011) may identify more precise fiber pathways to solve this problem. Secondly, our study used the AAL-atlas to divide the human brain into 90 regions, which provides a macroscale view of brain networks. However, recent studies have shown that higher-spatial-resolution networks of up to 10240 parcels or of voxel-based scheme may provide an increased sensitivity to local properties (Hayasaka and Laurienti, 2010; Zalesky et al., 2010; Khundrakpam et al., 2015). Additional studies with higher-spatial-resolution networks are needed to validate our findings. Finally, the population in our study is within a narrow age range and in developing stage. Our findings may vary with time and be inconsistent with findings with children or adults. Further studies need to use wide age range to obtain more precise results.