Rich-club organization and functional brain network metrics during facial emotional processing: a task-based fMRI study Authors: Maryam Gholam Tamimi, Mohammad Reza Daliri Published: 2026 ABSTRACT BackgroundEmotion is a complex psychological phenomenon involving both arousal and valence. emotional processing (EP) refers to the ability to perceive and interpret emotional stimuli, such as facial expressions or vocal cues.MethodsIn this study, we investigated functional connectivity (FC), graph-theoretical network measures, and rich-club organization during EP using task-based functional magnetic resonance imaging (fMRI) data from 100 healthy participants from the Human Connectome Project (HCP). Mean time series were extracted from 264 regions of interest (ROIs) defined by the Power Atlas, encompassing 10 large-scale functional networks. Pairwise Pearson correlation coefficients were computed to generate FC matrices, which were then thresholded using the orthogonal minimal spanning tree (OMST) method to form adjacency matrices. These matrices served as the basis for calculating global and local network metrics and analyzing rich-club organization. Permutation-based paired t-tests (p < 0.05, 1000 permutations) and family wise error (FWE)-corrected were used to identify significant differences between face and shape conditions.ResultsOur findings indicate a significant modulation of neural activity between the face and shape conditions during facial EP. Significant differences in FC, network metrics, and rich-club organization were observed at both ROI and network levels. High-level cognitive networks exhibited stronger positive correlations, whereas low-level perceptual networks showed increased anticorrelations. Global network measures, including modularity, mean local efficiency, and clustering coefficient, were increased, indicating enhanced functional segregation. Simultaneously, selective rich-club connectivity among hubs in dorsal attention, frontoparietal, visual, somatomotor, and subcortical networks suggests preserved network integration at the mesoscale level. These findings uniquely combine rich-club organization with graph-theoretical measures across 10 large-scale networks, providing novel insights into hub coordination under emotional demands.ConclusionsEP induces reorganization of brain networks, enhancing functional specialization while selectively modulating hub regions to maintain efficient integration. These results offer deeper insight into neural mechanisms underlying emotional cognition and may help explain connectivity alterations in emotional and affective disorders. Introduction Emotions have played an essential role throughout evolution, helping organisms adapt to dynamic environments and enhance survival [1]. Among these emotions, fear and anger represent two fundamental affective states that often drive motivated behavior in opposite directions [2]. Emotional processing (EP) is intricately linked to social cognition, emotional regulation, and the evaluation of behavioral interactions [3, 4]. Over the past decades, the neural mechanisms underlying EP have been extensively investigated in both healthy and clinical populations [5–8], leading to the identification of multiple brain systems involved in emotional perception and expression [9–11]. Task-based fMRI studies using naturalistic stimuli—such as facial expressions of anger, fear, or happiness—have revealed that facial emotion perception engages a distributed network of cortical and subcortical regions, including the inferior occipital gyrus, superior temporal gyrus, inferior and middle frontal gyrus, fusiform gyrus, amygdala, and thalamus [12–14]. Collectively, these findings suggest that distinct brain regions contribute to different components of EP [15, 16]. Functional connectivity (FC) quantifies the temporal interaction between the different brain regions and provides a means to characterize the coordination and integration of neural activity [17]. Strong FC between regions implies efficient communication and functional integration within the brain. Recent task-based fMRI study, has focused on FC between the amygdala and occipital face area, highlighting emotion-related changes in neural synchronization [18]. Prior researches have reported FC alterations within and between large-scale brain networks during emotion perception [19, 20]. In recent years, graph-theoretical analysis (GTA) has emerged as a powerful framework for quantifying the organization and topology of brain networks [21–23]. Network science offers a range of measures—such as global and local efficiency, clustering coefficient, and centrality—that capture different aspects of network architecture [24–26]. Studies applying GTA to EP tasks have reported significant modulations in these metrics [27–29]. In particular, analyses of nodal properties (e.g., degree, betweenness centrality, participation coefficient, path length) have demonstrated that certain brain regions act as affective hubs, supporting communication within and across networks [30]. Despite these advances, previous research has primarily focused on regional activations or pairwise connectivity, without comprehensively examining hub-to-hub interactions across large-scale functional networks. Moreover, most studies have not evaluated how network-level reorganization balances functional integration and segregation during emotional face processing. These limitations hinder a full understanding of the mesoscale coordination underlying EP. Building on this framework, the present study focuses on rich-club organization, which identifies highly interconnected hub regions forming the functional core of the brain network. Unlike traditional hub metrics, rich-club analysis captures coordinated interactions among hubs, providing a mesoscale perspective on how core networks maintain integration [31, 32]. Specifically, this study aims to: (1) Identify changes in FC at both regional and network levels in response to negative emotional stimuli; (2) Evaluate local and global network metrics, characterizing differences within and between 10 large-scale functional networks; and (3) Examine the rich-club organization across these networks and compare its properties between experimental conditions. Together, these analyses of FC, graph-theoretical metrics, and rich-club topology provide a comprehensive framework for understanding the large-scale network reconfiguration underlying emotion perception. Materials and methods Dataset Data were obtained from 100 healthy young adults (55 females; mean age = 27.7 ± 3.7 years) provided by the Human Connectome Project (HCP). The datasets are publicly available through the ConnectomeDB database (Open Access Data; https://db.humanconnectome.org). All participants were physically healthy and free from neurological or psychiatric disorders [33, 34]. Specifically, participants underwent screening via structured clinical interviews as part of the HCP protocol to rule out any neurological or psychiatric conditions. Their physical health was assessed using standardized health questionnaires and medical history reviews. Furthermore, to minimize variability stemming from hemispheric dominance, only right-handed individuals were included in this analysis. The negative affect was quantified using the National Institutes of Health (NIH) Toolbox Emotion Measures [35]. All adult participants (≥ 18 years) completed the 18+ battery, a Computer Adaptive Test (CAT) designed to assess emotional function. This battery incorporates items from the Patient-Reported Outcomes Measurement Information System (PROMIS) Anger (22 items) and Fear (29 items) banks. Participants rated each item on a 5-point Likert scale: Never, Rarely, Sometimes, Often, and Always. A higher total score indicated greater intensity of the emotion trait. In this study, the average scores for Anger Affect were 48.9 ± 8.4, and for Fear Affect were 50.2 ± 7.1. These measures were included to capture individual differences in negative emotional traits within the healthy population. Even in non-clinical samples, variability in negative affect has been shown to relate to emotional regulation processes and individual differences in brain functional organization [36, 37]. Therefore, incorporating negative affect scores allowed us to examine whether variability in affective traits is associated with differences in brain network architecture across participants. The original data collection was approved by the Washington University Institutional Review Board, and permission to use the HCP Open Access data was obtained for the present study. Behavioral task Task-based fMRI studies are used to determine which brain regions are activated by specific behavioral or cognitive tasks, making it essential to design an appropriate experimental task. Hariri and colleagues [38, 39], created a facial EP task suitable for use during fMRI scanning. This task follows a block-design format, consisting of three emotion blocks where participants match faces showing negative emotions, and three control blocks involving matching neutral shapes. Each participant performs two runs with six blocks in a fixed sequence. During the face blocks, an emotion face (angry or fearful) was presented at the top of the screen and subjects were to choose one of two faces presented at the left or right bottom. During the shape blocks, a neutral oval shape is shown at the top, and participants choose the matching shape based on orientation from two options presented at the bottom. Each block begins with a 3 s (sec) cue indicating the task (“Face” or “Shape”), followed by six trials—each consisting of a 2 s stimulus presentation and a 1 s inter-trial interval—lasting 21 s in total. Each run ends with an 8 s fixation period [40]. Mean accuracy rate (AR) was 98.2 ± 0.4% for the face condition and 96.4 ± 0.5% for the shape condition. Mean reaction time (RT) was 800 ± 13 msec for the face condition and 777 ± 12 msec for the shape condition. Data acquisition and preprocessing EP task-based fMRI data were acquired using a Siemens 3 Tesla Connectome Skyra scanner with gradient-echo echo-planar imaging (EPI). Each run consisted of 176 frames with an approximately duration of 2 min (min) and 16 s. The imaging parameters included a slice thickness of 2.0 mm, a field of view (FOV) measuring 208 × 180 mm², a resolution matrix of 104 × 90, and 72 slices. The repetition time (TR) was set at 720 msec, echo time (TE) at 33.1 msec, flip angle (FA) at 52 degrees, and a multiband acceleration factor of 8. The HCP task-fMRI data was preprocessing using the HCP “fMRIVolume” pipeline (version 3.4) [41]. This pipeline includes several stages: correcting gradient distortions, adjusting for motion and geometric distortions, aligning the functional images to the corresponding structural scans, performing nonlinear registration to the MNI152 (Montreal Neurological Institute) standard brain space, and normalizing image intensities. More information about the HCP “fMRIVolume” pipeline procedures can be found in [29]. Additional preprocessing procedures involved physiological noise reduction using the Component-based Noise Correction method (CompCor), which extracts principal components from signals within the white matter (WM) and cerebrospinal fluid (CSF) [42]. Furthermore, six rigid-body motion parameters—three for translation and three for rotation—along with their temporal derivatives, were regressed out. Spatial smoothing was applied using a Gaussian kernel with 4 × 4 × 4 mm³ full width at half maximum (FWHM), and temporal filtering was performed with a band-pass range of 0.01–0.1 Hz. To ensure high data quality, we calculated mean frame-wise displacement (FD) [43] and the derivative of root-mean-square (rms) variance over voxels (DVARS) [44]. All participants met strict quality-control (QC) thresholds, with average values of FD = 0.169 ± 0.058 mm (any single FD < 0.25 mm) and voxel-wise DVARS = 0.99 ± 0.12. Furthermore, all participants were unrelated, minimizing potential confounding effects of familial or genetic similarity. Functional connectivity (FC) analysis Functional connectivity (FC) describes the statistical relationships—whether linear or nonlinear—between brain regions that are spatially separated. These dependencies are typically estimated from correlations between blood oxygen level dependent (BOLD) signal time series extracted from single voxels or predefined regions of interest (ROIs) [45]. In this study, ROIs were defined using the Power atlas, which comprises 264 functionally defined nodes assigned to 10 large-scale networks: auditory (AU), cingulo-opercular (CO), dorsal attention (DA), default mode (DM), frontoparietal (FP), somatomotor (SOM), visual (VIS), salience (SAL), subcortical (SUB), and ventral attention (VA) [46]. The Power atlas was constructed through a combination of meta-analytic and resting-state connectivity approaches and has been widely applied in task-based network studies [47, 48]. Compared with anatomically based atlases, functionally defined parcellations have shown improved sensitivity for detecting task-related connectivity changes. Nevertheless, because the Power atlas does not include anatomical labels for its nodes, we incorporated the AAL atlas as an auxiliary anatomical reference. The AAL parcellation divides the cortex and subcortical regions into 116 defined areas [49], allowing us to assign meaningful anatomical descriptions to the nodes when interpreting local network measures. For each participant, connectivity was estimated using fMRI time points acquired during task performance [50]. Our procedure involved using a total of 75 timepoint per subject for each condition to compute the correlations. This count aligns with the parameters of 18 s duration and 0.72 s TR, applied across the number of trials. To isolate condition-specific connectivity during the EP task, we first removed task-evoked coactivation effects from the time series. Specifically, a subject-level general linear model (GLM) was implemented with separate regressors for the face and shape blocks, each convolved with the canonical hemodynamic response function (HRF). The model also included six motion parameters and additional nuisance regressors. Residual BOLD time series were then extracted from each ROI [51]. Condition-specific time series were obtained by masking residuals corresponding to the face and shape blocks. Pairwise FC was calculated as the Pearson correlation coefficient between all ROI pairs, separately for each condition, resulting in both positive and negative correlation values. Correlation coefficients were Fisher Z-transformed to improve normality. This procedure yielded a 264 × 264 symmetric connectivity matrix per subject for each condition, comprising 34,716 unique edges. The overall workflow for the FC analysis is illustrated in Fig. 1. Fig. 1The work flow for FC, graph and rich-club analysis The work flow for FC, graph and rich-club analysis Graph analysis The functional brain network for each participant was represented as a graph, in which nodes corresponded to the 264 ROIs and edges represented the Pearson correlation coefficients between the BOLD time series of each ROI pair. Both positive and negative FC values were retained in the connectivity matrices and were incorporated in graph metric computations whenever the corresponding metric supported signed networks. To minimize the influence of spurious or noisy connections while preserving the overall network topology, we applied the orthogonal minimal spanning tree (OMST) filtering method [52, 53]. Because the original OMST framework was developed for positive-weight networks, a three-step procedure was adopted. First, the signed FC matrices were transformed by taking the absolute value of the correlation coefficients, yielding non-negative weighted matrices that reflected connection strength irrespective of correlation sign. Second, OMST filtering was applied to these absolute-weight matrices to identify a sparse yet topologically optimized network backbone that maximized global efficiency relative to wiring cost. Third, after the OMST topology had been determined, the original signs of the retained connections were restored using the corresponding entries from the original signed FC matrices. This procedure allowed the network backbone to be identified based on connection strength while preserving the sign information for subsequent analyses designed for signed networks. OMST iteratively extracts multiple minimal spanning trees (MSTs) from the fully connected network and integrates them into a sparse yet topologically optimized network that maximizes global efficiency relative to wiring cost. Unlike conventional thresholding approaches that require arbitrary density or correlation cutoffs, OMST is parameter-free and has been shown to produce more reliable and reproducible network structures at both the individual and group levels. Following OMST filtering and sign restoration, graph-theoretical analyses were performed on the resulting weighted networks. The global measures include characteristic path length (L), global efficiency (GE), mean local efficiency \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( {\overline{{{\mathrm{LE}}}} } \right)$$\end{document}, mean clustering coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( {\overline{{{\mathrm{CC}}}} } \right)$$\end{document}, modularity (M), and small-world (SW). Among these global metrics, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{L}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{G}\mathrm{E}$$\end{document} are utilized to assess network integration, which refers to how efficiently nodes communicate across the entire network. Conversely, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{{{\mathrm{CC}}}} }$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{{{\mathrm{LE}}}} }$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{M}$$\end{document} are used to quantify the degree of network segregation, indicating the formation of densely connected groups of nodes that enable specialized processing and facilitate communication within local clusters [54]. At the nodal level, local efficiency (LE) and clustering coefficient (CC) were used as measures of local segregation [55–57]. Because shortest-path measures are not straightforwardly defined in signed networks and may require specialized formulations such as signed Laplacian approaches [58], path-based metrics (L, GE, and LE ) were computed using edge lengths derived from the absolute magnitude of the connection weights, thereby ensuring non-negative distances and compatibility with conventional shortest-path algorithms. This approach is commonly adopted when graph-theoretical path metrics are derived from correlation-based FC networks [59]. Consequently, no shortest-path computations were performed on negative edge weights. In contrast, measures specifically designed for signed networks were computed using the restored signed adjacency matrices. In particular, the CC was calculated according to the signed weighted formulation proposed by Costantini and Perugini [60], which extends the classical CC by incorporating the balance of positive and negative relationships within network triangles. Similarly, M was estimated using the signed Louvain algorithm [61], which optimizes a signed M quality function appropriate for networks containing both positive and negative connections. All network metrics were calculated in Python using the bctpy [59] and NetworkX [62] libraries. The overall workflow for the graph analysis is illustrated in Fig. 1. Rich-club organization The rich-club organization in the brain is a critical network structure that facilitates the integration and coordination of information across different brain systems, particularly during EP. This organization is characterized by a set of highly interconnected hub nodes that are essential for efficient communication and functional integration. Due to the definition of the rich-club coefficient for weighted networks, the rich-club analysis exclusively incorporated positive weights. To analyze the rich-club organization, the rich-club coefficient, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varnothing\left(k\right)$$\end{document}, was calculated for each degree level k, ranging from 1 to the network’s maximum degree. This coefficient is defined by the formula:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varnothing \left(k\right)=\frac{2{E}_{>k}}{{N}_{>k}({N}_{>k}-1)}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{>k}$$\end{document} represents the number of existing connections among nodes whose degree is greater than k, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{>k}$$\end{document} is the count of such nodes after excluding those with degrees less than or equal to k [31]. To determine whether a rich club actually exists, the value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varnothing\left(k\right)$$\end{document} is normalized by comparing it to the average rich-club coefficient of a set of randomized networks with the same size and degree distribution. For the normalization of the rich-club index, we employed degree-preserving randomization using the Maslov–Sneppen method to establish the null model [63]. These null networks are generated by randomly rewiring connections while preserving the degree sequence throughout:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varnothing}_{norm}\left(k\right)=\frac{\varnothing \left(k\right)}{{\varnothing}_{random}\left(k\right)}$$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varnothing}_{random}\left(k\right)$$\end{document} is the mean rich-club coefficient calculated across these random networks. A normalized coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varnothing}_{norm}\left(k\right),$$\end{document} greater than one indicates that nodes with degree above k form a rich-club, meaning they are densely connected [64, 65]. In this study, 1000 such randomized networks were created to establish a baseline for each average randomized networks. Because the choice of the degree k is somewhat subjective and varies depending on the analysis goals [66, 67], we selected a range 9 ≤ k ≤ 26. Nodes with degree values within this range were classified as part of the rich-club organization. Then, connections were categorized as: rich-club links, connections between two rich-club nodes, feeder links, connections rich-club nodes to non-rich-club nodes, and local links, connections between non-rich-club nodes [68]. In rich-club analysis, it is important to identify the network hubs. In network science, hubs are nodes characterized by high centrality, exerting a strong influence on the overall network structure. In this study, hubs are defined as nodes whose degree (k) surpasses the statistically significant rich-club threshold range (9 ≤ k ≤ 26) and that rank within the top 15% of the degree distribution [57]. Statistical analysis We investigated differences in brain network properties between the face and shape conditions at multiple levels, including edgewise FC, nodal and global graph measures, and rich-club organization. Condition-specific differences in connectivity were assessed using permutation-based paired t-tests. Within each subject, condition labels were randomly shuffled 1000 times, and the maximum t-value (Tmax) across all edges was recorded in each iteration to construct a null distribution. Observed t-values were then compared to this distribution, providing family-wise error (FWE) control for multiple comparisons [69]. Local and global network metrics were calculated for each participant. Between-condition differences were evaluated using non-parametric permutation tests (1000 permutations), with the maximum t-value approach applied to control for multiple comparisons across nodes. All p-values were two-tailed, with statistical significance set at p < 0.05 FWE-corrected. Rich-club coefficients φ(k) were computed as a function of node degree for each condition. Statistical significance was determined by generating 1000 randomized networks per participant and calculating their corresponding φrandom(k). A rich-club was deemed significant if φnorm(k) > 1 and the observed φ(k) exceeded the 95th percentile of the null distribution (p < 0.05) [70]. To quantify the magnitude of differences between conditions, effect size (Cohen’s d ± standard deviation (SD)) for paired samples was calculated for each ROI-to-ROI edge and network metric as the mean difference divided by the standard deviation of the differences [71]. Results Significant FC between ROIs and large-scale functional networks We first assessed condition-specific differences in FC between the face and shape conditions using permutation-based paired t-tests. Out of 34,716 possible edges, 187 exhibited significant differences. Figure 2 displays the significant matrix highlights edges that differed significantly between conditions (p < 0.05, 1000 permutations, FWE-corrected), for all participants based on 264 ROIs defined by the Power atlas. The maximal observed t-value across all edges was Tmax = 5.1283 corresponding to FWE-corrected p-value (pFWE = 0.002). Significant edges were distributed across multiple large-scale networks, with the majority involving the CO, DA, DM, FP, SOM, and VIS networks, while the remaining edges spanned other networks. Fig. 2Significant matrix highlighting edges that differed significantly between shape and face conditions (p < 0.05, 1000 permutations, FWE-corrected). The x- and y-axes correspond to ROIs belong to 10 large-scale functional networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, and VIS). Red and blue colors indicate high and low correlation values between node pairs, respectively Significant matrix highlighting edges that differed significantly between shape and face conditions (p < 0.05, 1000 permutations, FWE-corrected). The x- and y-axes correspond to ROIs belong to 10 large-scale functional networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, and VIS). Red and blue colors indicate high and low correlation values between node pairs, respectively We next investigated condition-specific connectivity differences at the level of large-scale functional networks. Pairwise mean correlations between the 10 networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, and VIS) were computed by averaging all ROI-to-ROI connections linking each network pair. Significant differences between conditions revealed a clear reorganization of large-scale interactions during face processing. Specifically, higher positive connectivity was observed among high-level cognitive control networks, particularly between the FP and DA networks and between the FP and CO networks, suggesting enhanced coupling of executive and attentional systems during the processing of emotional facial stimuli. In contrast, several low-level sensory networks exhibited stronger negative correlations (anticorrelations) in the face versus shape condition. These included connections between the SOM and AU networks and SOM and VIS networks. The increased anticorrelation indicates greater functional segregation among sensory processing systems during facial emotion processing. Figure 3 shows the significant matrix highlights significant differences in network-level connectivity (p < 0.05, 1000 permutations, FWE-corrected, Tmax = 5.837, pFWE = 0.002), with each element representing the mean pairwise FC between ROIs of two networks. The mean effect size (Cohen’s d ± SD) for these edges was 0.27 ± 0.15, reflecting moderate differences in FC. Fig. 3Significant matrix highlights significant differences between face and shape conditions (p < 0.05, 1000 permutations, FWE-corrected). The x- and y-axes correspond to 10 large-scale functional networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, and VIS). Red and blue colors indicate high and low correlation values between network pairs, respectively Significant matrix highlights significant differences between face and shape conditions (p < 0.05, 1000 permutations, FWE-corrected). The x- and y-axes correspond to 10 large-scale functional networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, and VIS). Red and blue colors indicate high and low correlation values between network pairs, respectively FC values were computed by averaging pairwise ROI-to-ROI correlations within each network. Statistical comparisons were performed using non-parametric permutation tests (p < 0.05,1000 permutations, FWE-corrected). Figure 4 shows the median and interquartile ranges for the FC values within 10 large-scale functional networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA and VIS) under two conditions. Across conditions, high-level cognitive networks—including DA, CO, FP, and DM networks—exhibited significantly lower FC in the face condition compared to shape. Conversely, the SOM showed significantly higher FC during the face condition, reflecting increased within-network integration during facial EP. Fig. 4Boxplots displaying the median and interquartile ranges of the within-network FC values for each 10 large-scale functional networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, VIS) under two conditions (shape: blue; face: orange). The red stars indicate statistically significant differences (p < 0.05, 1000 permutations, FWE-corrected) Boxplots displaying the median and interquartile ranges of the within-network FC values for each 10 large-scale functional networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, VIS) under two conditions (shape: blue; face: orange). The red stars indicate statistically significant differences (p < 0.05, 1000 permutations, FWE-corrected) Local and global graph metrics relative to ROIs and large-scale functional networks We examined local network topology during face and shape conditions using graph-theoretical metrics, including clustering coefficient (CC) and local efficiency (LE). Significant differences in local network measures were observed between the face and shape conditions. Table 1 summarizes nodes exhibiting significant condition differences (p < 0.05, 1000 repetitions, FWE-corrected), mean ± SD for the shape and the face conditions, Tmax and pFWE for local graph metrics. Also, the corresponding anatomical locations of ROIs was reported by the AAL Atlas. Key regions with altered local network metrics spanned both high-level cognitive networks (DM, FP, DA) and low-level perceptual/sensorimotor networks (SOM, VIS, SUB).These results indicate that facial EP is associated with widespread local network reorganization, affecting cortical and subcortical regions involved in sensorimotor integration, attention, visual perception, and cognitive control. Table 1ROIs exhibiting significant differences in local network measures (clustering coefficient (CC) and local efficiency (LE)), between the face and shape conditions (p < 0.05, 1000 permutations, FWE-corrected)ROIMNI coordination coordinationNetworkBrain region AALLocal measuresMean ± SD (shape)Mean ± SD (face)TmaxpFWE2129 − 17 71SOMRight Precentral gyrusCC0.0133 + 0.00350.012 + 0.00382.53010.012223− 23 − 30 72SOMLeft Postcentral gyrusCC0.0134 + 0.0030.0118 + 0.00293.73920.00023110 − 17 74SOMRight Supplementary motor areaCC0.0136 + 0.0040.0123 + 0.00342.53040.0122403 − 17 58SOMRight Supplementary motor areaCC0.0136 + 0.00380.0122 + 0.00342.64830.0087928 − 48 31DMRight Posterior cingulate gyrusCC0.0146 + 0.00510.0159 + 0.00282.19550.0296113− 3 42 16DMLeft Anterior cingulate gyrusCC0.0135 + 0.00350.0123 + 0.00292.47840.0141124− 26 − 40 − 8DMLeft ParaHippocampalLE0.0145 + 0.00530.0159 + 0.00312.25350.025614318 − 47 − 10VISRight LingualLE0.015 + 0.0040.0165 + 0.0023.57260.000514440 − 72 14VISRight Middle occipital gyrusLE0.0148 + 0.00410.0163 + 0.00123.57920.0005146− 8 − 81 7VISLeft CalcarineLE0.0146 + 0.00420.0163 + 0.00123.79640.0002147− 28 − 79 19VISLeft Middle occipital gyrusLE0.015 + 0.00390.0167 + 0.00134.12980.0001151− 15 − 72 − 8VISLeft LingualCC0.0105 + 0.0050.0132 + 0.0044.17780.000115820 − 86 − 2VISRight LingualLE0.0145 + 0.00510.0161 + 0.00282.81260.0056164− 42 − 74 0VISLeft Middle occipital gyrusLE0.0147 + 0.00410.0165 + 0.00193.88190.0002172− 33 − 79 − 13VISLeft Inferior occipital gyrusCC0.0134 + 0.00350.011 + 0.00444.16710.0002177− 53 − 49 43FPLeft Inferior parietal gyrusLE0.0165 + 0.0010.0159 + 0.00143.68450.0003201− 42 25 30FPLeft Inferior frontal gyrus, triangular partCC0.0111 + 0.00440.013 + 0.00343.44610.000723129 1 4SUBRight putamenLE0.0133 + 0.00620.0149 + 0.0052.0810.0388232− 31 − 11 0SUBLeft putamenLE0.0138 + 0.00610.0153 + 0.00392.06090.040825746 − 59 4DARight Middle Temporal gyrusLE0.0152 + 0.00350.0163 + 0.0022.76240.0064262− 42 − 60 − 9DALeft Inferior temporal gyrusLE0.0143 + 0.00440.016 + 0.00323.15840.0019263− 17 − 59 64DALeft PrecuneusCC0.0136 + 0.00380.0125 + 0.00312.2530.0254SOM somatomotor network, DA dorsal attention network, DM default mode network, FP frontoparietal network, SUB subcortical network, VIS visual networkFor each ROI, the table lists its number, MNI coordinates, corresponding Power atlas network, anatomical location, local measures, mean ± SD for the shape and the face conditions, Tmax and pFWE Table 1ROIs exhibiting significant differences in local network measures (clustering coefficient (CC) and local efficiency (LE)), between the face and shape conditions (p < 0.05, 1000 permutations, FWE-corrected)ROIMNI coordination coordinationNetworkBrain region AALLocal measuresMean ± SD (shape)Mean ± SD (face)TmaxpFWE2129 − 17 71SOMRight Precentral gyrusCC0.0133 + 0.00350.012 + 0.00382.53010.012223− 23 − 30 72SOMLeft Postcentral gyrusCC0.0134 + 0.0030.0118 + 0.00293.73920.00023110 − 17 74SOMRight Supplementary motor areaCC0.0136 + 0.0040.0123 + 0.00342.53040.0122403 − 17 58SOMRight Supplementary motor areaCC0.0136 + 0.00380.0122 + 0.00342.64830.0087928 − 48 31DMRight Posterior cingulate gyrusCC0.0146 + 0.00510.0159 + 0.00282.19550.0296113− 3 42 16DMLeft Anterior cingulate gyrusCC0.0135 + 0.00350.0123 + 0.00292.47840.0141124− 26 − 40 − 8DMLeft ParaHippocampalLE0.0145 + 0.00530.0159 + 0.00312.25350.025614318 − 47 − 10VISRight LingualLE0.015 + 0.0040.0165 + 0.0023.57260.000514440 − 72 14VISRight Middle occipital gyrusLE0.0148 + 0.00410.0163 + 0.00123.57920.0005146− 8 − 81 7VISLeft CalcarineLE0.0146 + 0.00420.0163 + 0.00123.79640.0002147− 28 − 79 19VISLeft Middle occipital gyrusLE0.015 + 0.00390.0167 + 0.00134.12980.0001151− 15 − 72 − 8VISLeft LingualCC0.0105 + 0.0050.0132 + 0.0044.17780.000115820 − 86 − 2VISRight LingualLE0.0145 + 0.00510.0161 + 0.00282.81260.0056164− 42 − 74 0VISLeft Middle occipital gyrusLE0.0147 + 0.00410.0165 + 0.00193.88190.0002172− 33 − 79 − 13VISLeft Inferior occipital gyrusCC0.0134 + 0.00350.011 + 0.00444.16710.0002177− 53 − 49 43FPLeft Inferior parietal gyrusLE0.0165 + 0.0010.0159 + 0.00143.68450.0003201− 42 25 30FPLeft Inferior frontal gyrus, triangular partCC0.0111 + 0.00440.013 + 0.00343.44610.000723129 1 4SUBRight putamenLE0.0133 + 0.00620.0149 + 0.0052.0810.0388232− 31 − 11 0SUBLeft putamenLE0.0138 + 0.00610.0153 + 0.00392.06090.040825746 − 59 4DARight Middle Temporal gyrusLE0.0152 + 0.00350.0163 + 0.0022.76240.0064262− 42 − 60 − 9DALeft Inferior temporal gyrusLE0.0143 + 0.00440.016 + 0.00323.15840.0019263− 17 − 59 64DALeft PrecuneusCC0.0136 + 0.00380.0125 + 0.00312.2530.0254SOM somatomotor network, DA dorsal attention network, DM default mode network, FP frontoparietal network, SUB subcortical network, VIS visual networkFor each ROI, the table lists its number, MNI coordinates, corresponding Power atlas network, anatomical location, local measures, mean ± SD for the shape and the face conditions, Tmax and pFWE ROIs exhibiting significant differences in local network measures (clustering coefficient (CC) and local efficiency (LE)), between the face and shape conditions (p < 0.05, 1000 permutations, FWE-corrected) SOM somatomotor network, DA dorsal attention network, DM default mode network, FP frontoparietal network, SUB subcortical network, VIS visual network For each ROI, the table lists its number, MNI coordinates, corresponding Power atlas network, anatomical location, local measures, mean ± SD for the shape and the face conditions, Tmax and pFWE We next examined global network topology during face and shape conditions using graph-theoretical metrics, including modularity (M), global efficiency (GE), mean local efficiency (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{{\mathrm{LE}}}$$\end{document}), characteristic path length (L), mean clustering coefficient (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{CC}}$$\end{document}), and small-worldness (SW). Metrics were computed on OMST-filtered networks, and statistical comparisons between conditions were performed using permutation-based paired t-tests (p < 0.05, 1000 permutations, FWE-corrected). Table 2 summarizes mean ± SD for the shape and the face conditions, Tmax and pFWE for global graph metrics. Table 2Comparison of global network metrics between face and shape conditionsGlobal graph metricsMean ± SD (shape)Mean ± SD (face)TmaxpFWEModularity (M)0.5705 ± 0.02470.5862 ± 0.02223.23010.0002Global Efficiency(GE)0.3398 ± 0.00810.3331 ± 0.00784.34240.0002Mean Local Efficiency(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{LE}}$$\end{document})0.5428 ± 0.02670.5681 ± 0.02364.87520.0002Mean Clustering Coefficients(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{CC}}$$\end{document})0.3454 ± 0.02220.3636 ± 0.01974.08750.0004Characteristic Path Length(L)1.2580 ± 0.04461.2904 ± 0.04774.60470.0004Small_Worldness (SW)3.01949 ± 0.5533.2762 ± 0.51683.83930.0006Reported values include mean ± SD for the shape and the face conditions, Tmax and pFWE Table 2Comparison of global network metrics between face and shape conditionsGlobal graph metricsMean ± SD (shape)Mean ± SD (face)TmaxpFWEModularity (M)0.5705 ± 0.02470.5862 ± 0.02223.23010.0002Global Efficiency(GE)0.3398 ± 0.00810.3331 ± 0.00784.34240.0002Mean Local Efficiency(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{LE}}$$\end{document})0.5428 ± 0.02670.5681 ± 0.02364.87520.0002Mean Clustering Coefficients(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{CC}}$$\end{document})0.3454 ± 0.02220.3636 ± 0.01974.08750.0004Characteristic Path Length(L)1.2580 ± 0.04461.2904 ± 0.04774.60470.0004Small_Worldness (SW)3.01949 ± 0.5533.2762 ± 0.51683.83930.0006Reported values include mean ± SD for the shape and the face conditions, Tmax and pFWE Comparison of global network metrics between face and shape conditions Reported values include mean ± SD for the shape and the face conditions, Tmax and pFWE Figure 5 shows the boxplot comparisons of global network measures between face and shape conditions. Results revealed that emotional face processing modulated global network organization. Specifically: (1) GE decreased and L increased during the face condition, indicating reduced functional integration, (2) M, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{LE}}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{CC}}$$\end{document} increased, reflecting enhanced functional segregation and more cohesive community structure, (3) SW exceeded 1 in both conditions, confirming SW topology, and was slightly higher during the face condition, suggesting a more segregated yet efficient network organization under facial EP. Fig. 5Boxplots display the median and interquartile range of six global graph-theoretical metrics: modularity (M), global efficiency (GE), mean local efficiency (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{LE}}$$\end{document}), characteristic path length (L), mean clustering coefficient (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{CC}}$$\end{document}), and small-worldness (SW), computed for the shape (blue) and face (orange) conditions. Asterisks indicate significant differences between conditions (p < 0.05, 1000 permutation, FWE-corrected) Boxplots display the median and interquartile range of six global graph-theoretical metrics: modularity (M), global efficiency (GE), mean local efficiency (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{LE}}$$\end{document}), characteristic path length (L), mean clustering coefficient (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{CC}}$$\end{document}), and small-worldness (SW), computed for the shape (blue) and face (orange) conditions. Asterisks indicate significant differences between conditions (p < 0.05, 1000 permutation, FWE-corrected) The mean effect size (Cohen’s d ± SD) for the global graph metrics were as follows: M (0.548 ± 0.21), GE (− 0.717 ± 0.25), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{LE}}$$\end{document} (0.814 ± 0.18), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{CC}}$$\end{document} (0.698 ± 0.20), L (0.590 ± 0.16), and for SW (0.363 ± 0.14). We further examined global network metrics across 10 large-scale functional networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, and VIS) for the face vs. shape contrast (p < 0.05, 1000 permutations, FWE-corrected; Fig. 6). Significant changes in global metrics were found primarily within the CO, DA, DM, FP, and VIS networks. Specifically: (1) DA, DM, FP, and VIS networks showed decreased GE and increased L, indicating reduced functional integration, (2) The DA network exhibited the highest\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{CC}$$\end{document}, GE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{LE}$$\end{document} values, along with relatively lower M and L, suggesting that this network is more functionally segregated and specialized during emotional face processing. Fig. 6The boxplots displaying the median and interquartile ranges of global network measures, (modularity, global efficiency, mean local efficiency, characteristic path length, mean clustering coefficient, and small-world), for 10 large-scale networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, VIS) and between two conditions (face: orange, shape: blue). The red stars indicate statistically significant differences (p < 0.05, 1000 permutations, FWE-corrected) The boxplots displaying the median and interquartile ranges of global network measures, (modularity, global efficiency, mean local efficiency, characteristic path length, mean clustering coefficient, and small-world), for 10 large-scale networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, VIS) and between two conditions (face: orange, shape: blue). The red stars indicate statistically significant differences (p < 0.05, 1000 permutations, FWE-corrected) Table 3 presents the results of statistical comparisons of global network metrics across 10 large-scale functional networks between the face and shape conditions. These findings indicate a network-level reorganization characterized by increased segregation among local communities and reduced global integration during emotional face processing. This pattern aligns with selective processing demands, complementing the previously observed local network and FC changes, and provides a basis for examining rich-club hub connectivity as a mechanism supporting integration among core networks despite globally increased segregation. Table 3Statistical comparisons of global network metrics across 10 large-scale functional networks between face and shape conditionsGlobal graph metricsNetworkmean ± SD (shape)mean ± SD (face)TmaxpFWEModularity (M)DA0.0271 ± 0.01820.0378 ± 0.02063.89650.002FP0.2675 ± 0.047620.2861 ± 0.04962.56720.008VIS0.3186 ± 0.04460.3337 ± 0.04652.28440.0213Global efficiency (GE)DA0.9195 ± 0.01650.9080 ± 0.01674. 81180.0002DM0.5179 ± 0.01480.5116 ± 0.01323.64130.0005FP0.6907 ± 0.01800.6795 ± 0.01714.61310.0002VIS0.6378 ± 0.01550.6298 ± 0.01663.34530.002Mean local efficiency (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{LE}}$$\end{document})CO0.8834 ± 0.030310.8602 ± 0.03383.83930.0002DA0.9248 ± 0.02490.9069 ± 0.02634.08750.0002FP0.7739 ± 0.04030.7625 ± 0.03812.15370.0002Mean clustering coefficient (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{CC}}$$\end{document}) CO0.7688 ± 0.05850.7235 ± 0.06484.60470.0002DA0.8497 ± 0.04990.8137 0.05264.34240.0002Characteristic path length (L)DA1.1609 ± 0.03311.1840 ± 0.03344. 44840.0002DM2.2624 ± 0.10602.2894 ± 0.09042.35180.0192FP1.6607 ± 0.06761.6929 ± 0.07013.56620.0007VIS1.8067 ± 0.06301.8414 ± 0.07793.38910.0007CO cingulo-opercular network, DA dorsal attention network, DM default mode network, FP frontoparietal network, VIS visual networkReported values include mean ± SD for the shape and the face conditions, Tmax and pFWE Table 3Statistical comparisons of global network metrics across 10 large-scale functional networks between face and shape conditionsGlobal graph metricsNetworkmean ± SD (shape)mean ± SD (face)TmaxpFWEModularity (M)DA0.0271 ± 0.01820.0378 ± 0.02063.89650.002FP0.2675 ± 0.047620.2861 ± 0.04962.56720.008VIS0.3186 ± 0.04460.3337 ± 0.04652.28440.0213Global efficiency (GE)DA0.9195 ± 0.01650.9080 ± 0.01674. 81180.0002DM0.5179 ± 0.01480.5116 ± 0.01323.64130.0005FP0.6907 ± 0.01800.6795 ± 0.01714.61310.0002VIS0.6378 ± 0.01550.6298 ± 0.01663.34530.002Mean local efficiency (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{LE}}$$\end{document})CO0.8834 ± 0.030310.8602 ± 0.03383.83930.0002DA0.9248 ± 0.02490.9069 ± 0.02634.08750.0002FP0.7739 ± 0.04030.7625 ± 0.03812.15370.0002Mean clustering coefficient (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\mathrm{CC}}$$\end{document}) CO0.7688 ± 0.05850.7235 ± 0.06484.60470.0002DA0.8497 ± 0.04990.8137 0.05264.34240.0002Characteristic path length (L)DA1.1609 ± 0.03311.1840 ± 0.03344. 44840.0002DM2.2624 ± 0.10602.2894 ± 0.09042.35180.0192FP1.6607 ± 0.06761.6929 ± 0.07013.56620.0007VIS1.8067 ± 0.06301.8414 ± 0.07793.38910.0007CO cingulo-opercular network, DA dorsal attention network, DM default mode network, FP frontoparietal network, VIS visual networkReported values include mean ± SD for the shape and the face conditions, Tmax and pFWE Statistical comparisons of global network metrics across 10 large-scale functional networks between face and shape conditions CO cingulo-opercular network, DA dorsal attention network, DM default mode network, FP frontoparietal network, VIS visual network Reported values include mean ± SD for the shape and the face conditions, Tmax and pFWE Rich-club organization Rich-club organization was analyzed by computing the normalized rich-club coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varnothing}_{norm}\left(k\right)$$\end{document} across node degrees, with significant rich-club identified for 9 ≤ k ≤ 26 where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varnothing}_{norm}\left(k\right)$$\end{document} exceeds 1. Figure 7 illustrates the distribution of rich-club nodes for the face and shape conditions across 10 large-scale functional networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, and VIS). Significant rich-club nodes were primarily distributed in the DA, FP, SOM, SUB, and VIS networks (p < 0.05, 1000 permutations, FWE-corrected). Among these networks, VIS and SOM hubs likely support detailed visual and sensorimotor integration of facial features, whereas FP and DA hubs facilitate executive control and attentional processes. Importantly, the inclusion of SUB nodes within the rich-club core suggests that subcortical structures contribute directly to the integrative backbone of the network. Rather than functioning solely as peripheral affective modules, subcortical regions appear embedded within the high-degree core architecture, potentially facilitating rapid salience signaling and efficient coordination between perceptual and higher-order control systems. This cortico–subcortical configuration indicates that emotional face processing promotes the formation of a task-dependent, highly interconnected core that supports large-scale information integration. Fig. 7The boxplots displaying the median and interquartile ranges of the rich-club nodes for face and shape conditions in 10 large-scale functional networks. The red stars indicate significant differences (p < 0.05, 1000 permutations, FWE-corrected) between two conditions (shape: blue; face: orange) The boxplots displaying the median and interquartile ranges of the rich-club nodes for face and shape conditions in 10 large-scale functional networks. The red stars indicate significant differences (p < 0.05, 1000 permutations, FWE-corrected) between two conditions (shape: blue; face: orange) Moreover, we determined the network hubs associated with the face and shape conditions across 10 large-scale functional networks (AU, CO, DA, DM, FP, SAL, SOM, SUB, VA, and VIS). The analysis revealed that significant hub nodes were primarily located within the DA, VIS, and SOM networks (p < 0.05; 1000 permutations, FWE-corrected). These identified integration regions demonstrate high participation across multiple large-scale brain networks and exhibit significant overlap with the rich-club core, thereby underscoring their pivotal integrative functions within the network architecture. To further characterize network organization, we analyzed the number of rich-club, feeder, and local links under face and shape conditions. Permutation-based tests (p < 0.05, 1000 permutations, FWE-corrected) revealed significant differences in rich-club and feeder links. Figure 8 illustrates the number of rich-club, feeder and local links for the face and shape conditions. The face condition showed increased rich-club links (Tmax = 3.8393, pFWE = 0.004) and decreased feeder links (Tmax = 4.0875, pFWE = 0.006), while local links did not differ significantly. Effect size analysis revealed a modest increase in rich-club links (d = 0.23 ± 0.12) and decreases in feeder (d = − 0.24 ± 0.10) and local links (d = − 0.20 ± 0.09), indicating a shift toward direct hub-to-hub connectivity at the expense of peripheral connections. Fig. 8The boxplots displaying the median and interquartile ranges of the number of rich-club, feeder and local links for two conditions. The stars indicate significant differences (p < 0.05, 1000 permutations, FWE-corrected) between two conditions (shape: blue; face: orange) The boxplots displaying the median and interquartile ranges of the number of rich-club, feeder and local links for two conditions. The stars indicate significant differences (p < 0.05, 1000 permutations, FWE-corrected) between two conditions (shape: blue; face: orange) Discussion In this study, we investigated how emotional face processing modulates large-scale functional brain networks using task-based fMRI data from the HCP. By combining FC analyses, graph-theoretical metrics, and rich-club measures, we provide a comprehensive network-level perspective on the neural mechanisms supporting emotion perception. A major methodological strength is the use of the OMST approach, which preserves the topological backbone of the network while avoiding arbitrary thresholding. Additionally, permutation-based inference provided robust statistical validation, mitigating limitations of traditional parametric tests in network analyses. Our results demonstrate that emotional face perception induces a coordinated reorganization of brain networks. FC analyses revealed significant changes at both regional and large-scale network levels. High-level cognitive networks, including the CO, DA, and FP networks, exhibited stronger positive correlations, whereas low-level perceptual networks, such as VIS, AU, and SOM, showed increased anticorrelations. These patterns indicate that cognitive control and attentional systems become more integrated, while sensory networks operate in a more functionally segregated manner during emotional processing, potentially reflecting top-down modulation that prioritizes emotional information over sensorimotor integration. Such coordinated reconfiguration likely supports behavioral inhibition, emotional regulation, and reward processing in response to negative facial stimuli [72–75]. Graph metrics results, showed alternations global network measures. Significant decreased GE and increased L showed decreasing functional integration and, significant increased M, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{LE},$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{CC}$$\end{document} showed significant increasing functional segregation. So, the brain network showed a higher functional segregation and lower functional integration under negative emotional stimuli. Also, our findings revealed emotional stimuli promote locally specialized processing. Prior work has reported similar effects, particularly under negative emotional conditions, where decreased GE and increased LE support specialized local processing [76–78]. Examination of local metrics showed widespread alterations in both high-level cognitive (DM, FP, DA) and low-level perceptual/sensorimotor (SOM, VIS, SUB) networks. Together, these results suggest that emotional stimuli promote locally specialized processing while maintaining selective integration among key hubs [79, 80]. Analysis of global network measures across 10 large-scale networks highlighted significant changes primarily in the CO, DA, DM, FP, and VIS networks, providing insight into how specific networks contribute to emotion perception. These alterations are crucial for understanding how emotional stimuli influence brain connectivity and functionality. At the level of large-scale networks, the DA network exhibited high \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{LE}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{CC}$$\end{document}, reflecting functional specialization. Rich-club analyses revealed a preferential strengthening of hub-to-hub connections within the VIS, SOM, FP, DA, and SUB networks, while feeder links decreased and local links remained largely unchanged. This pattern indicates a subtle reorganization of network hubs, favoring direct integration among core regions at the expense of peripheral connections. Notably, VIS and SOM hubs appear critical for visual and sensorimotor integration of facial features, whereas FP and DA hubs facilitate executive control and attentional processes. The additional recruitment of subcortical structures highlights the importance of cortico-subcortical interactions, likely reflecting mechanisms of affective regulation and attentional control. These task-specific rich-club networks likely serve as a core backbone for balancing segregation and integration, supporting efficient communication across distributed networks even under globally increased segregation [81, 82]. Also, this study identified significant brain regions acting as hubs within the DA, VIS, and SOM networks. These regions are important because overlap significantly with the core of the rich-club structure. This highlights their crucial role in integrating information across the entire brain network under facial EP. These findings suggest a cross-modal mechanism, whereby emotional perception involves interplay between segregated sensory networks and integrated cognitive networks, facilitating the processing of complex affective information. Our results implicated regions across ten large-scale functional brain networks, each contributing uniquely to emotional processing (EP) [83]. The DM network comprises regions such as the anterior and posterior cingulate cortices that are typically active during rest and self-referential thought [84, 85]. Beyond their role in introspection and mind-wandering, these areas are also engaged during emotional evaluation and regulation [86, 87]. In the current study, the DM network exhibited the highest M, but the lowest GE among global metrics, suggesting increased functional integration. The FP and CO networks are critical for higher-order cognitive control, including attention shifting, inhibitory regulation, and decision-making. Consistent with prior findings, increased engagement of the dorsolateral prefrontal cortex (DLPFC), a key FP node, has been linked to the active suppression and cognitive regulation of negative emotions [88, 89]. The observed enhancement of these networks may therefore reflect top-down modulation of affective responses. The DA network responsible for voluntary allocation of attentional resources [90], demonstrated the highest FC and elevated \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{CC}$$\end{document}, GE, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{LE}$$\end{document} in our graph-theoretical results. These patterns indicate that emotionally facial stimuli strongly recruit attentional systems to facilitate goal-directed processing of affective cues. The visual (VIS) network showed robust functional coupling during EP, consistent with its fundamental role in processing complex visual and facial features. Prior meta-analytic evidence has confirmed that regions within the VIS network—including the fusiform and lingual gyri—are consistently activated by visual emotional stimuli and facial expressions [91–93]. Interestingly, activation of the AU network was also observed, possibly reflecting cross-modal engagement, as prior work has shown that even silent visual speech can activate auditory cortical regions [94]. Finally, the SOM network exhibited significant FC differences between conditions. This finding aligns with previous studies showing that emotional stimuli can elicit bodily and motor responses through enhanced activity in SOM regions [95, 96]. Such modulation likely reflects the neural preparation for behavioral or physiological reactions to emotionally salient events. Our findings extend prior work showing that emotional tasks modulate large-scale integration across cortical and subcortical systems. Studies using dynamic functional connectivity have similarly reported transient increases in modular organization during affective states, supporting the adaptive reconfiguration hypothesis. By quantifying these effects through graph-theoretical and rich-club frameworks, the present study provides a complementary network-level interpretation of how emotional information is integrated and regulated across the human connectome. Limitations and future works There are several limitations to the current study that should be acknowledged. First, the sample size was relatively small, which may limit the generalizability of the findings. Future research should aim to replicate these results using larger and more diverse datasets to enhance statistical power and robustness. Second, the participant age range was restricted to young adults between 22 and 35 years old. Given that EP and brain network connectivity can vary with age [97] and gender [98], future studies should include a broader demographic spectrum to examine potential age and sex related effects on FC and network topology. Third, the emotional stimuli employed in this study were limited to negative emotions (fear and anger). Expanding the investigation to include a wider range of emotional states—such as positive emotions—will provide a more comprehensive understanding of how emotion perception influences brain networks. Finally, this study assumed static FC during facial EP. However, the brain operates as a dynamic system, and recent research highlights the importance of capturing temporal fluctuations in connectivity using dynamic FC approaches [99]. Incorporating dynamic analyses in future work could yield more nuanced insights into the temporal evolution of network interactions during emotional processing. Conclusions Our findings highlight that emotional face perception induces coordinated network reorganization, characterized by increased segregation among sensory networks, selective integration among cognitive hubs, and strengthened rich-club connectivity. These results underscore the central role of rich-club hubs in coordinating specialized and integrated processing across the brain, providing a mesoscale network framework for understanding the neural basis of facial emotion perception.