Biological validity, test–retest reliability, and behavioral relevance of the single-subject brain volumetric similarity network

Abstract

The T1-weighted brain magnetic resonance imaging (MRI)-based volumetric similarity network (VSN) offers an advantage in clinical settings due to its ease of acquisition and widespread availability. However, its validity, reliability, and behavioral relevance remain unclear. The present study aimed to assess the reproducibility and utility of the VSN as a foundation for future research and clinical applications. Here, we analyzed three datasets (total N = 354), with two datasets having repeated MR runs (Dataset 1: n = 86; Dataset 2: n = 49) and two having an attention measure (Datasets 1 and 3: n = 219). For each run and participant, the VSN was generated using interregional morphological similarity metrics. We examined whether the VSN reflects the brain’s cytoarchitecture and assessed its test–retest reliability by using connectome fingerprints in Datasets 1 and 2. We also examined the VSN’s behavioral relevance and further tested its predictive utility using connectome-based predictive modeling in Datasets 1 and 3. The VSN defined using the z-transformed interregional correlation showed significant spatial similarity with the cytoarchitectonic covariance network (rhos = 0.23 and 0.22 in Datasets 1 and 2, respectively; p < 0.01). The VSN also yielded high test–retest reliability, demonstrated by high identification accuracy (91% and 100% in Datasets 1 and 2, respectively). However, unlike the functional connectome (r > 0.31, p < 0.01), VSNs did not reliably predict individual differences in attention (r < 0.1, p > 0.3). This study demonstrates the biological validity and high reliability of the VSN to support brain fingerprinting of individual subjects, but not individual differences in attention.

1. Introduction

Human behavior and cognition are facilitated by and emerge from our brain, which has a complex and interconnected network (Wang and He, 2024). Although functional magnetic resonance imaging (fMRI) and diffusion tensor imaging are the main modalities for constructing the human brain networks, T1-weighted structural MRI has attracted attention recently in terms of delineating the whole-brain morphological connectivity patterns, owing to its unique advantages of widespread availability, high spatial resolution and signal-to-noise ratio, and robustness to artifacts. While traditional models were developed using population-level approaches, recent efforts have shifted toward individualized modeling. Individualized morphological brain networks can be defined by calculating the interregional morphological correlations within a participant, offering opportunities to study interindividual variances in the coordinated patterns of the brain morphology in both healthy and disease models (Wang and He, 2024).

Structural MRI allows the analysis of surface and volumetric features of the brain (Good et al., 2001; Sebenius et al., 2024). To date, most previous studies have focused on surface-based features to construct the morphological networks (Li et al., 2023; Sebenius et al., 2023; Seidlitz et al., 2018). However, a volume-based network (volumetric similarity network [VSN]), which was originally introduced in early studies on the morphological similarity network (Tijms et al., 2012), requires shorter processing time than surface-based methods and offers the ability to incorporate deep gray matter (GM) structures beyond the cortex (Gaser et al., 2024; Henschel et al., 2020). In the initial study by Tijms et al., the brain was parcellated into thousands of cubes (27 voxels per cube), and morphological connectivity was estimated by calculating the Pearson correlation of gray matter volume across voxels within predefined cubic regions (Tijms et al., 2012). However, as the cubes are defined in individual native space, the resulting networks differ in the number and spatial configuration of nodes across individuals. To standardize better across participants, the majority of subsequent studies generate VSNs at the regional level based on prior brain parcellations and uniform bins of regional gray matter volume (Li et al., 2021; Wang and He, 2024).

Despite its advantages, several research gaps remain in the development and application of the VSN as a brain network that represents an individual. First, while multiple approaches—including correlation-, distance-, and divergence-based methods—have been proposed for VSN construction, establishing methodological consensus is essential for improving reproducibility and facilitating the broader utility of VSN (Wang and He, 2024). Second, morphological similarity may reflect multiple mechanisms, including mutual trophic influences and shared experience-related plasticity (Mechelli et al., 2005). The homophily principle—i.e., architectonically similar areas are more likely to be connected—may represent one potential contributor to such similarity. Accordingly, we evaluated whether the VSN resembles cytoarchitectonic similarity (Sebenius et al., 2024). Third, to serve as a biomarker, the VSN must offer test–retest reliability and reflect an individual’s behavior or dysfunction (Ramduny and Kelly, 2025). Therefore, in the present study, we aimed to validate the cytoarchitectonic similarity, reliability, and behavioral relevance of the VSN.

2. Materials and methods

2.1. Participants and data acquisition

The present study analyzed data from the following three publicly available datasets: Yale dataset (Dataset 1) (86 participants aged 18–35 years, https://doi.org/10.15154/1520622) (Yoo et al., 2022); Beijing Normal University (BNU) dataset (Dataset 2) (50 participants aged 19–30 years, http://fcon_1000.projects.nitrc.org/indi/CoRR/html/bnu_1.html) (Lin et al., 2015); and ADHD-200 dataset from the Peking University site (Dataset 3) (219 participants aged 8–18 years, http://preprocessed-connectomes-project.org/adhd200/) (Bellec et al., 2017).

All participants included in Datasets 1 and 2 completed two MRI sessions within an interval of approximately 6 weeks. Both datasets included participants without a history of any neurological or psychiatric disorders. T1-weighted structural images in Dataset 1 were acquired for each participant using a magnetization-prepared rapid-gradient echo (MPRAGE) collected at the beginning of each session with the following parameters: TR = 1800 ms, TE = 2.26 ms, flip angle = 8°, acquisition matrix = 256 × 256, in-plane resolution = 1.0 mm², slice thickness = 1.0 mm, and 208 sagittal slices. The structural scans in Dataset 2 were acquired on a 3T Siemens Trio-Tim MRI scanner using the MPRAGE sequence with the following parameters: TR = 2530 ms; TE = 3.39 ms; TI = 1100 ms; slice thickness = 1.33 mm; number of slices = 144; no inter-slice gap; FOV = 256 × 256 mm²; and flip angle = 7° Dataset 1 was approved by the Yale University Institutional Review Board and Dataset 2 was approved by the Institutional Review Board of the State Key Laboratory of Cognitive Neuroscience and Learning at Beijing Normal University.

Dataset 3 included preprocessed MRI data from children with and without attention-deficit/hyperactivity disorder (ADHD). The current study included data from the Peking University site, which contains data obtained from individuals with a relatively low head motion and whose attention scores were successfully predicted by fMRI connectivity (Rosenberg et al., 2016; Yoo et al., 2022). The Research Ethics Review Board of the Institute of Mental Health, Peking University, approved the Dataset 3 collection. Moreover, informed consent was obtained from each participant’s parent, and all children agreed to participate in the study. Detailed descriptions of the inclusion criteria and imaging parameters and procedures are available online at http://fcon_1000.projects.nitrc.org/indi/adhd200/.

We also analyzed the fMRI connectomes from Datasets 1–3 for comparison. For Dataset 1, preprocessed fMRI data were obtained as described previously (Yoo et al., 2022). For Dataset 2, the fMRI connectome was derived from the available raw fMRI data, and the detailed preprocessing procedure is described in the Supplementary Methods. For Dataset 3, preprocessed fMRI data were obtained from the Athena pipeline (https://www.nitrc.org/plugins/mwiki/index.php/neurobureau:AthenaPipeline).

2.2. Structural image quality control and preprocessing

For Datasets 1 and 2, preprocessing was conducted using SPM12 and the CAT12 toolbox in MATLAB 2022b (MathWorks, Natick, MA, USA) environment with default settings to improve reproducibility (Gaser et al., 2024; Penny et al., 2011). Briefly, the input data were inhomogeneity-corrected; then, the skull and other non-brain elements were removed. After, the images were registered into the standard MNI space using the Diffeomorphic Anatomic Registration Through Exponentiated Lie Algebra algorithm (DARTEL) and segmented into GM, white matter (WM), and cerebrospinal fluid components. Quality checks of the images were performed visually and quantitatively with the options “Display slices” and “Check sample homogeneity” in the CAT12 toolbox (Gaser et al., 2024). One participant in Dataset 2 (0025,913, session 2) has an incomplete T1 brain image uploaded; thus, the participant was excluded, and 49 participants were included in the further analysis. Moreover, the image “grade” that summarizes the measures of image quality by the CAT12 toolbox was employed to identify the images with poor quality and incorrect preprocessing. All acquired images were graded “B (good, minimum score of 84.34)” or above.

Dataset 3 was already preprocessed (Burner pipeline) and shared online (Pankaj). Briefly, T1 images were segmented into GM and WM probability maps, which determines additional tissue classes to reduce the confounding effects of non-brain (e.g. scalp) structural variability on the registration (Ashburner et al., 2014). Inter-subject normalization was performed using the DARTEL. The images were iteratively registered to the group average (population template), resulting in finer registration. The registration parameters were applied to each image to transform each image into the space of the final population template. Modulation was applied to conserve the global tissue volumes after normalization, and four participants were excluded during preprocessing due to quality issues according to the dataset description; hence, a total of 219 participants were finally enrolled.

2.3. Construction of individual VSNs

A flowchart of the data analytical pipeline used in the present study is shown in Fig. 1. Following the region-based morphological similarity framework described by Wang & He (2024), we constructed individual VSNs using predefined atlas parcels rather than cube-based voxel partitions. To construct the VSN at a single-subject level, we defined nodes using the Shen atlas and then constructed the edges by assessing the similarity of the modulated GM volumetric distribution, which reflects the absolute amount (volume) of GM (named “mwp1” in CAT12 output and “mwrc1” in VBM output) between nodes (Good et al., 2001; Raamana and Strother, 2017; Shen et al., 2013). We primarily used the Shen atlas, which parcellates the brain into 268 regions based on resting-state fMRI, to provide whole-brain coverage and maintain comparability with dozens of prior connectome fingerprinting and CPM-based attention prediction studies (Finn et al., 2015; Yoo et al., 2022).

Fig. 1.

A flowchart of the data analytical pipeline. Panel A shows each T1-weighted structural MR images in the Yale, BNU, and ADHD-200 datasets were processed to modulated normalized gray matter segmentations. The Shen atlas was used to divide the gray matter volumes into 268 regions of interest (ROI). Interregional morphological similarity was estimated between any pair of ROIs using histogram distance, resulting in a single-subject VSN. In panel B, the VSN was tested to determine whether it shows cytoarchitectonic correlation, reliability, and compared its reliability score against the whole brain voxels (whole brain), ROI values (Region), Morphometric INverse Divergence (MIND), and functional connectome (FC). Lastly, panel C shows the utility in predicting the individual attention was evaluated using connectome-based predictive modeling (CPM), and the prediction accuracy was compared to a whole-brain FC. Dataset 1 (Yale) included 92 participants in total. Of these, 86 participants had two scanning sessions and were used for test–retest reliability and identification analyses (denoted as ‘Dataset 1: Yale (2 sessions)’), whereas all 92 participants were used in all other analyses (denoted as ‘Dataset 1: Yale’). BNU, Beijing Normal University; VSN, volumetric similarity network; FC, functional connectome; MIND, Morphometric INverse Divergence; I_diff, the difference between the average within-subject similarity and the average between-subject similarity; CPM, connectome-based predictive modeling.

As mentioned in the Introduction, various methods are used to define pair-wise relations between the brain nodes to construct the individualized VSN, which are as follows: correlation-based (e.g. Pearson correlation), divergence-based (e.g. Kullback–Leibler divergence), and distance-based (e.g. Euclidean distance) methods (Wang and He, 2024). To comprehensively evaluate the reliability of the different VSN construction methods, we constructed 31 VSNs for each participant; of these, 30 were derived using methods available in the “graynet” toolbox (https://raamana.github.io/graynet/usage_cli.html#type-of-computation; the description of metrics is at https://loli.github.io/medpy/metric.html; full list is in the Appendix) and the remaining VSN was derived using the z-transformed correlation (hereafter referred to as the “z-transformed correlation”) approach. Finally, a total of 31 symmetric 268 × 268 adjacency matrices were generated per participant, each representing a VSN constructed using a different similarity measure.

2.4. Cytoarchitectonic correlates of the VSN

To explore whether the VSN reflects cytoarchitecture, we assessed the spatial similarity between the cytoarchitectonic covariance network based on the digital Von Economo–Koskinas atlas (Li et al., 2023; Scholtens et al., 2018) (http://www.dutchconnectomelab.nl/economo/) with group-averaged VSN in Datasets 1 and 2. The digital Von Economo–Koskinas atlas contained 35 layer-specific quantitative cytoarchitectonic features from 43 brain regions. Specifically, the 35 features were based on five measures: cell content, cell size, thickness at the dome (the outer curve of a gyrus), thickness at the wall (the side of the gyrus), and total thickness. These measures were taken from layers I to VI and the total of the layers (Triarhou, 2013). The regional values of each cytoarchitectonic feature were z-transformed and correlated between each pair of the 43 brain regions to derive a cytoarchitectonic covariance network (43 × 43). We also acquired 31 types of group-averaged VSNs using the same 43 region-based atlas normalized to the MNI space. Each type of group-averaged VSNs was correlated with the cytoarchitectonic covariance network (Spearman correlation). The significance levels of the correlations were estimated using permutation tests. The false discovery rate (FDR) procedure was used to correct for multiple comparisons across all 31 correlations at the p_FDR < 0.05. If a significant correlation was observed, we further quantified the contribution of each cytoarchitectonic feature by calculating the percentage change in the correlation before and after removing the feature to estimate the cytoarchitectonic covariance network. A higher percentage change indicates a greater contribution (Li et al., 2023). We further assessed interpretability by visualizing the VSN and cytoarchitectonic covariance network (43 × 43) and by summarizing edge-level correspondence using a rank-agreement approach. Briefly, we ranked all off-diagonal edges within each matrix and selected the 20 edges with the smallest absolute rank difference, and displayed as a binary map.

2.5. Reliability of the VSN

We used the following three metrics to examine the test–retest reliability of the VSN in Datasets 1 and 2: (1) Pearson correlation-based spatial similarity of the VSNs between two visits of the same participants; (2) connectome fingerprinting-based identification; and (3) differential identifiability (I_diff) (Amico and Goñi, 2018; Finn et al., 2015). In connectome fingerprinting, we conducted two complementary runs: (i) session 1 as the database and session 2 as the target, and (ii) session 2 as the database and session 1 as the target. The final results were obtained by averaging across these two directions. To obtain fingerprinting accuracy, we computed the correlations between a given participant in the target set and all the VSNs in the database and considered participant identification successful for the given participant if the correlation with the same participant is the highest. The identification accuracy of an entire dataset was defined as the proportion of correctly identified individuals out of the total number of participants. To test the significance of fingerprinting accuracy, we performed nonparametric permutation.

To evaluate the identifiability capability as a continuum, we adopted the I_diff (Amico and Goñi, 2018). I_diff is defined as I_self I_others, where I_self is the Pearson correlation values between the visits of the same participants and I _others is the Pearson correlation between visits of different participants. A higher I_diff implies a higher individual fingerprint among the population. We calculated the between-visit similarity, fingerprinting-based identification accuracy, and I_diff using the VSNs. Then, we compared these data to those acquired using the whole brain voxels, 268 ROIs from the Shen atlas, a recently suggested cortical similarity network named Morphometric INverse Divergence (MIND) (Sebenius et al., 2023), and the resting-state fMRI connectome. To test the robustness of our findings to the parcellation choice, we repeated the reliability analyses using the Von Economo–Koskinas atlas (43 × 43).

2.6. Behavioral relevance of the VSN

To test whether the VSN reflects individual differences in behavior, we performed a brain-wide association test with attention-related behaviors in Datasets 1 and 3. In the participants from Dataset 1, the attention task performance was assessed with sensitivity (d′) for the gradual-onset continuous performance task (gradCPT) and the general attention factor, which was defined by a mean of z-scored behaviors across the attention-related tasks (Yoo et al., 2022). In Dataset 3, we used the ADHD Rating Scale-IV score for the attention measure (DuPaul et al., 1998).

To investigate the edge-wise association between the VSN and attention-related behaviors, we delineated the set of VSN edges showing significant correlation with the attention scores using the FDR with a threshold of p_FDR level of < 0.05 (that is, attention scores were correlated with each VSN edge, and significance was assessed using 5000 nonparametric permutations implemented in the Network-Based Statistic (NBS) toolbox, with FDR correction applied for multiple comparisons (Zalesky et al., 2010)). The “z-transformed correlation” VSN was used for this and all subsequent analyses, as it demonstrated the highest cytoarchitectural covariance and reliability in the previous analysis.

Connectome-based predictive modeling (CPM), a validated data-driven method (Shen et al., 2017; Yoo et al., 2022), was then used to test the predictive utility of VSN. CPM was trained to predict the individuals’ attention scores from the VSN and was validated via leave-one-out cross-validation within Datasets 1 and 3. CPM comprises the following three steps: feature selection, model building, and model validation. In each iteration, we held one participant out for model testing and used the remaining 91 (Dataset 1) or 218 (Dataset 3) participants for model training. In Dataset 1, 86 participants completed two scanning sessions. For these individuals, two VSNs were averaged across sessions, resulting in one representative VSN for each individual. The six participants who were excluded from the earlier reliability analysis had only one session and, therefore, were included only in the behavioral modeling. In model training, we first selected the edges showing significant correlation with the attention measures in the training set (Pearson’s r, p < 0.05). These features yielded both positive and negative edge masks depending on the signs of their correlation with behaviors. For each individual in the training set, the strengths of two networks (one from the positive and the other from the negative network) were measured by summing their connectivity strengths. Then, we fitted a general linear model between task performance (dependent variable) and strengths of the two networks (independent variables). Once the two network masks and a general linear model were constructed, we applied CPM to the held-out testing participant. CPM estimated the strengths of two networks for the test participant and predicted the participant’s task performance from their network strength measures. During the procedure, attention measures were z-scored within the training set. Then, the mean and standard deviation computed within the training set were used to normalize the behaviors of the left-out testing sample. The prediction performance of CPM was assessed by correlating the predicted and actual attention measures using Pearson’s correlation. Lastly, we also performed both edge-wise association and CPM analyses using the functional connectome for comparisons. To assess robustness to parcellation choice, we repeated the CPM analysis using the Von Economo–Koskinas atlas (43 × 43).

3. Results

3.1. Association of the VSN with the cytoarchitectonic covariance network

We observed that the group-averaged VSNs constructed using either “correlation” or “z-transformed correlation” showed the strongest and most significant correspondence with the cytoarchitectonic covariance network among the 31 VSN types. The two variants yielded nearly identical results after rounding (“correlation”: Dataset 1 ρ = 0.234, Dataset 2 ρ = 0.223; “z-transformed correlation”: Dataset 1 ρ = 0.225, Dataset 2 ρ = 0.218; df = 901, FDR-corrected p < 0.01; Fig. 2A and 2B). Three other VSN edge-defining methods showing significant correlations were the “exponential of the mean difference normalized by standard deviation” (Dataset 1; rho = 0.22, Dataset 2; rho = 0.21), “cosine” (Dataset 1; rho = 0.21, Dataset 2; rho = 0.21), and “histogram intersection” (Dataset 1; rho = 0.19, Dataset 2; rho = 0.19) (all FDR-corrected p < 0.01). The other 26 measures did not show significant association with the cytoarchitecture (all negative associations with rhos below 0). To identify the most influential cytoarchitectonic features, we assessed how the correlation changed when each feature was excluded from the covariance matrix. The total cell content in layer III was found to contribute the most (Dataset 1: 13.9 %, Dataset 2: 12.8 %; Fig. 2C and 2D, radar plot), followed by the thickness of dome in layer V (Dataset 1: 6.3 %, Dataset 2: 7.0 %). To facilitate interpretation of the VSN–cytoarchitectonic coupling, we provide binary map highlighting the rank-agreement driver edges (Figure S1). Based on the observed cytoarchitectonic correlation, we selected five VSN construction methods for the following analyses: “correlation,” “z-transformed correlation,” “exponential of the mean difference normalized by standard deviation,” “cosine,” and “histogram intersection.” Consistently, only these five VSN construction methods showed positive correlations with functional connectomes in both datasets (Fig. 2E).

Fig. 2.

Associations of the VSN with the cytoarchitectonic covariance network.

Panels A and B show the scatter plots and correlation coefficients between the group-averaged VSNs and the cytoarchitectonic covariance network. The group-averaged VSN edges exhibited a significant positive correlation with the cytoarchitectonic covariance network in Datasets 1 (Panel A, first row, rho = 0.23) and 2 (Panel B, first row, rho = 0.22). Panels C and D show radar plots of the contribution of each cytoarchitectonic feature. In both datasets, the correlation was mainly driven by the cell contents in layer III followed by the dome thickness in layer V. Panel E shows the violin plot of the positive correlation between the VSNs and the functional connectome, and each dot represents a participant. Correlation_Z, “z-transformed correlation”; exp diff means norm, exponential of the mean difference normalized by standard deviation; hist intersection, histogram intersection

3.2. High test–retest reliability of the individual-level VSN

Fig. 3 shows the test–retest reliability, assessed by between-visit similarity, fingerprinting-based identification accuracy, and I_diff for the five selected VSNs. The reliability of the VSN was compared with that of whole-brain voxels, the 268 ROIs, MIND, and the whole-brain functional connectome. All five VSN construction methods achieved high between-visit within-individual similarity and an identification accuracy of > 90 % with permutation p < 0.001, comparable to those of the four comparison methods (Fig. 3A and 3B, first and second rows). All five tested VSNs yielded numerically higher between-visit similarity than the ROI values or functional connectome (Fig. 3A and 3B, first row). Among the five VSN methods, the VSNs defined by “correlation” and “z-transformed correlation” showed the highest discriminative ability (I_diff > 0.4) (Fig. 3A and 3B, third row). The robustness analysis with the Von Economo–Koskinas atlas gave same result; the VSNs defined by “correlation” and “z-transformed correlation” showed the highest discriminative ability (I_diff > 0.4). Notably, VSN yielded a higher I_diff than the whole brain voxel-based approach, suggesting its higher reliability in comparison to that of the original whole-brain structure. In summary, these results suggest that “correlation”- and “z-transformed correlation”- based VSNs provide a higher test–retest reliability over simple volumetrics, surface-based networks or functional connectome.

Fig. 3.

Brain-based identification result of the VSN, whole brain, and functional connectome. Most of the brain-based measures showed robust reliability evaluated by between-visit similarity and fingerprint identification accuracy (first and second rows), with the VSN constructed by “correlation” and “z-transformed correlation” showing enhanced I_diff (third row). This improvement was consistently observed in both Datasets 1 and 2. Fingerprint, the overall success rate obtained when an individual is identified based on their brain imaging derived between two imaging sessions; I_diff, the difference between the average within-subject similarity and the average between-subject similarity; Corr, Correlation; Corr Z, “z-transformed correlation”; EDMN, exponential of the mean difference normalized by standard deviation; Cosn, cosine; HistInt, histogram intersection; Voxel: whole-brain voxels; Region: 268 ROIs; MIND: Morphometric INverse Divergence; FC: functional connectome.

3.3. Behavioral relevance of the VSN

Next, we examined the behavioral relevance of the VSN. We first identified the sets of edges showing a significant correlation with attention scores in Datasets 1 (sustained attention) and 3 (ADHD-RS-IV inattention scores) (Fig. 4A and 4B; Datasets 1 and 3 denoted with cyan and red, respectively). The VSN constructed by “z-transformed correlation” yielded only a small number of significant edges in each dataset, and no overlapping edges were observed across the two datasets (Fig. 4A). Contrarily, the whole-brain functional connectome revealed over 150 significant edges in both datasets (Fig. 4B), with the majority of edges linking visual attention and motor networks (Fig. 4C and D, red arrow).

Fig. 4.

Correlation of the network edges and attention scores in Datasets 1 (sustained attention) and 3 (ADHD-RS-IV inattention score). The false discovery rate was used to correct for multiple tests when identifying the significant edges (Dataset 1, cyan; Dataset 3, red). The VSN constructed with “z-transformed correlation” showed sparse edges (panel A), whereas many edges from the functional connectome showed a correlation with the attention score (panel B). Panels C and D show the number of functional edges detected in Datasets 1 and 3, respectively. The majority of edges connect the visual and motor networks in the two datasets (red arrow). CBL, cerebellum; SC, subcortical; SAL, salience; VAs, visual association; VI, visual A; VII, visual B; Mot, motor; DMN, default mode network; FP, frontoparietal; MF, medial frontal.

In CPM, the VSN showed no significant correlation between the predicted and true attention scores (sustained attention, r = 0.08, p = 0.43; general attention factor, r = 0.06, p = 0.55; ADHD-RS-IV inattention scores, r = 0.06, p = 0.30) (Fig. 5, column annotated “correlation_z”). The VSN with the Von Economo–Koskinas atlas gave similarly null results; sustained attention, r = 0.02; general attention factor, r = −0.39. In contrast, as consistently revealed in previous studies, the resting-state functional connectome successfully predicted attention (sustained attention: r = 0.39, p < 0.001; general attention factor: r = 0.38, p < 0.001; ADHD-RS-IV score: r = 0.32, p < 0.001; see Fig. 5, column annotated “FC”).

Fig. 5.

Attention prediction performance of the VSN and functional connectome. The resting-state functional connectome (“FC”) successfully predicted the attention-related scores, outperforming the “z-transformed correlation” VSN (“Correlation Z”).

4. Discussion

In the present study, we comprehensively validated the VSN in terms of its cytoarchitectonic similarity, test–retest reliability, and assessed its behavioral relevance. We discovered significant covariance between the VSN and cytoarchitecture in the all tested datasets (rhos > 0.22, “correlation” and “z-transformed correlation”). The effect sizes are approximately double those reported for surface-based morphological similarity networks (rho =0.111) (Li et al., 2023), and they are notable given the substantial cross-modal difference in scale between in vivo MRI (millimeter-scale voxels, atlas-level parcellation) and post-mortem cytoarchitectonic data (microscopic laminar-level measurements). Accordingly, VSN may more closely reflect the homophily principle, one rationale for VSNs, alongside other proposed mechanisms such as mutual trophic influences and shared experience-related plasticity (Mechelli et al., 2005; Sebenius et al., 2024). Among all the tested edge-defining metrics, we observed the strongest correlation when volumetric similarity was defined using “correlation” or “z-transformed correlation,” suggesting that it may be the most biologically relevant metric. The cytoarchitectonic feature that contributed most to the observed correlation was the total cell content in layer III, indicating a potential layer-specific relationship between the cytoarchitectonic properties and GM volume. As layer 3 has larger and more numerous pyramidal cells, which send excitatory outputs to other parts of the cortex, either nearby or in the opposite hemisphere, this finding aligns well with the homophily principle (Sebenius et al., 2024; Watson et al., 2010). The rank-agreement driver map (Figure S1) further suggests that the VSN–cytoarchitectonic correspondence is not dominated by a single lobe. Instead, concordant edges are broadly distributed across lobes, supporting a diffuse macro-scale alignment between MRI-derived volumetric similarity and cytoarchitectonic organization.

The VSN exhibited high test–retest reliability across two independent datasets. We consistently found that the VSNs successfully distinguished individuals across both Datasets 1 and 2 with high between-visit within-subject similarity and an identification accuracy of > 90 % (Fig. 3). Unlike previous studies that focused on cortical thickness–derived networks, we demonstrated that the VSN yields high fingerprinting accuracy, highlighting the unique contribution of GM volume to individual brain identification. (Li et al., 2023; Wang et al., 2016). Moreover, we further evaluated the identifiability metrics of I_diff, and compared them to the values achieved using whole-brain voxels, ROIs, MIND (a surface-based network), and functional connectome (Fig. 3, last row). Across all reliability metrics, VSNs constructed with “correlation” and “z-transformed correlation” yielded the greatest discriminability (I_diff > 0.4), even using substantially fewer features (35,778) than the whole-brain voxels (size of 113×137×113). Despite the availability of various VSN construction methods, there is currently no consensus on a standardized approach. Our results support the notion that the VSN is a robust and biologically grounded measure of network organization of individual brains. Lastly, the results from the cytoarchitectonic similarity and reliability analyses suggest that the VSN edges defined by “correlation” and “z-transformed correlation” would be the most promising candidate for a standardized VSN construction measure. It is worth noting that we identified similar covariance results between Datasets 1 and 2, despite their differences in terms of the data collection sites (i.e. Fig. 2A and B are similar). This cross-dataset replication supports the biological validity, reliability, and generalizability of the VSN.

Meanwhile, the behavioral relevance of the VSN was lower than anticipated. As attention is a well-known property that can be reliably predicted using functional connectomes (Rosenberg et al., 2016; Yoo et al., 2022), we expected that the VSN could replicate the prediction performance. However, we observed that the VSN could not accurately predict individual differences in attention, irrespective of the demographics and processing pipeline (Datasets 1 and 3, Fig. 5). Although prior studies have employed individual morphological brain networks to predict cognition or disease status (Li et al., 2023; Zhao et al., 2022), our results indicate that the attentional association of the VSN is less reliable and that functional connections have better utility compared to the VSN, at least for attention prediction using linear models. Together, these findings suggest that VSNs are highly reliable but comparatively less behaviorally sensitive with respect to attention. In contrast, the functional connectome achieves stronger behavioral prediction but is known to be more state-dependent and less reliable across sessions. This pattern is consistent with the reliability–sensitivity trade-off increasingly recognized in connectome research (Finn and Rosenberg, 2021), and highlights that the two modalities may capture complementary dimensions of brain organization. Lastly, our behavioral evaluation was performed with attention-related measures and a linear prediction model. These null findings should be cautiously interpreted and should not be taken as evidence that VSNs lack behavioral relevance in general. Other behavioral domains and non-linear approaches remain to be tested. Future work should examine broader behavioral phenotypes and prediction frameworks to further characterize the potential behavioral relevance of VSN.

In conclusion, our data revealed that the VSN defined by using the “z-transformed correlation” method demonstrated the most plausible network structure in terms of cytoarchitectonic covariance and enhanced test–retest reliability. These findings support the potential of “z-transformed correlation”-based VSN as a standardized volumetric morphological network at an individual level. However, the VSN did not predict attention-related behaviors, whereas the functional connectome showed stronger behavioral relevance. Thus, future research should further explore whether VSNs are relevant for other behaviors.

Supplementary Material

Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.neuroimage.2026.121862.

Appendix. List of VSN construction methods

• histogram-based: manhattan, chebyshev, chebyshev_neg, chi_square, correlate, correlate_1, cosine, cosine_1, cosine_2, cosine_alt, Euclidean, fidelity_based, histogram_intersection, histogram_intersection_1, jensen_shannon, minowski, noelle_1, noelle_2, noelle_3, noelle_4, noelle_5, relative_bin_deviation, relative_deviation, sorensen_dist, correlate_Z

• weights on original features: diff_medians, diff_medians_abs, diff_-means, diff_means_abs, exp_diff_means_norm_std, ranksum_statistic

Data and code availability statement

All data that support the findings of this study are from publicly available datasets.

The codes are available on request from the corresponding author.