Associations between anterior hypothalamic subunits and ADHD and autistic traits revealed by deep learning MRI segmentation

Abstract

The hypothalamus is a central regulator of neuroendocrine function and social behavior, yet its internal organization has remained difficult to examine in vivo in human neurodevelopmental research. Based on neuroendocrine models, we hypothesized that anterior hypothalamic subunits enriched in magnocellular neurosecretory nuclei—approximating the paraventricular (PVN) and supraoptic (SON) regions—form a structurally coherent axis linked to dimensional traits associated with autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD). To test this hypothesis, we applied a convolutional neural network–based automated segmentation framework to high-resolution T1-weighted MRI data from a non-clinical adult sample (n = 23), enabling volumetric quantification of ten hypothalamic subunits. After adjustment for biological sex, age, and total intracranial volume, PVN- and SON-associated volumes showed a significant positive structural covariance (r = 0.51, p = 0.022), consistent with a coherent anterior magnocellular-rich network. Reduced SON-associated volume was significantly associated with higher autistic traits (Autism-Spectrum Quotient: r =  − 0.45, p = 0.046) and greater ADHD symptom severity (CAARS: r =  − 0.59, p = 0.006). Bootstrapped path analyses (5,000 iterations) further supported this network organization, revealing a robust association between PVN- and SON-associated volumes (β = 0.51, p = 0.013) and a significant indirect association between PVN-associated volume and ADHD symptoms via SON-associated volume (β =  − 0.29, p = 0.006), with a marginal indirect association observed for autistic traits. Together, these preliminary findings provide in vivo evidence linking anterior hypothalamic structure to dimensional neurodevelopmental traits and highlight a potential neuroendocrine pathway relevant to psychiatric vulnerability.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s44192-026-00521-5.

Introduction

The hypothalamus comprises less than 1% of total brain volume, yet exerts disproportionate influence over survival-critical functions, including homeostasis, circadian regulation, and social behavior [1–3]. Within this compact and highly specialized structure, the paraventricular nucleus (PVN) and supraoptic nucleus (SON) form the core of the magnocellular neurosecretory system (MNS), a key neuroendocrine pathway responsible for the synthesis and systemic release of oxytocin (OXT) and arginine vasopressin (AVP) [4–6]. Although the PVN contains a heterogeneous neuronal population—including parvocellular neurons that regulate stress responses via the hypothalamic–pituitary–adrenal axis—the SON is composed almost exclusively of magnocellular neurons projecting to the posterior pituitary, underscoring its specialized neurosecretory role [7].

Accumulating evidence from human and animal research implicates the OXT and AVP systems in social bonding, empathy, emotional regulation, and stress responsivity, forming a central component of the so-called “social brain” [8, 9]. Disruption of these neuropeptide systems has therefore been proposed as a plausible pathophysiological mechanism underlying Autism Spectrum Disorder (ASD). Consistent with this framework, clinical trials of intranasal oxytocin administration have reported variable but promising effects on social functioning in subsets of individuals with ASD, suggesting that hypo-oxytocinergic states may characterize specific phenotypes within the autism spectrum [10–12]. However, a large multicenter randomized controlled trial in children and adolescents with ASD found no significant group-level benefit of intranasal oxytocin compared with placebo [13]. This null result has been interpreted as reflecting substantial phenotypic heterogeneity in oxytocinergic function, thereby motivating dimensional and individual-differences approaches such as that adopted in the present study. More recent work further suggests that responses to intranasal oxytocin may depend on condition-specific factors or biological subgroups, including dosage and oxytocin receptor genotype [14].

Beyond ASD, increasing evidence suggests that hypothalamic neuroendocrine mechanisms may also be relevant to Attention-Deficit/Hyperactivity Disorder (ADHD), a condition that frequently co-occurs with ASD [15]. While dopaminergic dysregulation remains central to prevailing models of ADHD, oxytocin has been shown to modulate reward processing and motivational salience through interactions with dopaminergic neurons in the nucleus accumbens [16]. These findings raise the possibility that hypothalamic integrity may play a previously underappreciated role in attentional regulation and impulse control, linking neuroendocrine function to dimensional neurodevelopmental traits that span diagnostic categories.

Critically, the volumetric size of the PVN and SON may reflect the density and somatic size of their constituent magnocellular neurons, as nuclear volume and total cell number are closely coupled in these regions [17]. Animal studies further demonstrate that increased physiological demand for oxytocin—such as during lactation and parturition—is associated with reversible hypertrophy of magnocellular neurons in the SON, selectively affecting oxytocinergic cells [18, 19]. These findings suggest that in vivo volumetric measures of these nuclei may serve as indirect markers of individual differences in neuroendocrine capacity.

Despite strong theoretical and experimental support, direct investigation of hypothalamic structure in human neurodevelopmental research has long been constrained by a fundamental “resolution gap.” Standard clinical MRI lacks sufficient contrast to delineate individual hypothalamic nuclei, forcing many human studies to rely on peripheral neuropeptide measurements, which may not accurately reflect central neuroendocrine organization or function [20]. Recent advances in deep learning–based neuroimaging, however, have begun to overcome this limitation. In particular, convolutional neural network–based automated segmentation frameworks trained on histological atlases now permit reliable in vivo quantification of anatomically defined hypothalamic subunits using conventional T1-weighted MRI [20, 21].

Neuroanatomical studies have identified consistent, though modest, volumetric differences across multiple brain structures in ASD and ADHD. Large-scale meta-analyses from the ENIGMA consortium have reported alterations in subcortical volumes—including the caudate nucleus, amygdala, and hippocampus—in individuals with ADHD [22] and ASD [23], as well as shared structural abnormalities across neurodevelopmental conditions [24]. Despite this growing body of work, the hypothalamus has received comparatively little attention in human neurodevelopmental neuroimaging, largely because its small size and complex internal organization have historically exceeded the spatial resolution of conventional MRI. The recent development of automated hypothalamic segmentation methods therefore provides a new opportunity to investigate whether structural variation in hypothalamic subunits contributes to neurodevelopmental traits.

Leveraging these methodological advances, the present study tested the hypothesis that anterior hypothalamic subunits enriched in magnocellular neurosecretory nuclei—approximating the paraventricular (PVN) and supraoptic (SON) regions—form a structurally coherent anterior axis associated with dimensional traits of ADHD and ASD in a non-clinical adult population (n = 23). By integrating automated hypothalamic volumetry with behavioral measures, this work aims to provide an anatomically grounded framework for investigating hypothalamic contributions to neurodevelopmental variation in humans.

Materials and methods

Participants and ethical considerations

The study sample comprised twenty-three healthy adult volunteers (n = 23; 11 females, 12 males) recruited from the local community in Sendai, Japan. Participants had a mean age of 21.87 years (SD = 1.96). Recruitment was conducted through community advertisements designed to capture a broad range of behavioral phenotypes within a non-clinical population. Participants were screened through pre-study interviews to exclude individuals with a history of diagnosed neurological disorders, major psychiatric conditions requiring treatment (e.g., schizophrenia or bipolar disorder), or active substance abuse. Individuals with medical conditions contraindicating MRI (e.g., pregnancy, implanted metallic devices, or severe claustrophobia) were also excluded.

Participants were not excluded on the basis of self-reported neurodevelopmental traits, as the study was designed to capture the full continuum of subclinical variation within a neurotypical community sample. None of the participants reported a prior formal diagnosis of ASD or ADHD. Consistent with this dimensional approach, Autism-Spectrum Quotient (AQ) scores ranged from 11 to 33 (mean = 20.26), with a subset of participants scoring above the commonly used screening threshold (AQ ≥ 26), reflecting sampling of the upper tail of the autistic trait distribution.

Written informed consent was obtained from all participants, and all procedures were conducted in accordance with the Declaration of Helsinki. The study protocol was approved by the Institutional Review Board of Tohoku Fukushi University (approval No: RS190607).

No formal a priori power analysis was conducted because no published effect size estimates were available for hypothalamic subunit–neurodevelopmental trait associations at the time the study was designed. Post hoc power calculations based on the observed effect sizes (r = 0.45–0.59) indicate that the present sample provides approximately 59–89% power at α = 0.05 (two-tailed). Specifically, the SON–ADHD association (r =  − 0.589) was detected with 89% power, whereas the SON–AQ association (r =  − 0.451) was moderately powered (59%), warranting replication in larger samples.

Behavioral assessment: ADHD and ASD traits

To quantify dimensional neurodevelopmental traits of interest, participants completed two standardized Japanese self-report questionnaires assessing ADHD symptoms and autistic traits.

Attention-Deficit/Hyperactivity Disorder (ADHD) traits were assessed using the Conners’ Adult ADHD Rating Scale (CAARS) [25, 26]. The CAARS evaluates the frequency and severity of current ADHD symptoms in adults and provides a total score as well as subscale scores for inattention and hyperactivity/impulsivity. In the present study, the total CAARS score was used as the primary behavioral index of attentional regulation.

Autistic traits were assessed using the Autism-Spectrum Quotient (AQ), a 50-item self-report measure designed to quantify the degree to which adults exhibit traits associated with the autism spectrum [27, 28]. The AQ comprises five subdomains—social skill, attention switching, attention to detail, communication, and imagination—with total scores ranging from 0 to 50, where higher scores indicate greater autistic traits.

Statistical treatment of behavioral scores. Given the potential overlap between ASD and ADHD traits, we calculated a pairwise Variance Inflation Factor (VIF) between AQ and ADHD total scores (VIF = 1.66), indicating that while correlated, these variables were statistically distinct enough for separate analyses. Furthermore, to control for potential confounding effects of biological sex and age, behavioral scores were residualized using linear regression prior to correlation and path analyses with hypothalamic volumes.

MRI data acquisition and parameters

Structural neuroimaging data were acquired using a 3.0 Tesla MRI scanner (Siemens MAGNETOM Skyra fit, Siemens Healthineers, Erlangen, Germany) equipped with a 20-channel head coil. For all participants (n = 23), high-resolution T1-weighted images were obtained using a magnetization-prepared rapid gradient-echo (MPRAGE) sequence with the following parameters: repetition time (TR) = 2300 ms; echo time (TE) = 2.98 ms; inversion time (TI) = 900 ms; flip angle = 9°; voxel size = 1 × 1 × 1 mm³ isotropic; field of view (FOV) = 256 × 256 mm; 192 slices; and total acquisition time = 5 min 20 s. Images were acquired in the sagittal plane.

To ensure adequate image quality for automated segmentation of small hypothalamic subunits, all scans underwent strict quality control procedures. Structural images were visually inspected for motion artifacts, susceptibility-related distortions, and overall signal-to-noise ratio. No datasets were excluded due to insufficient image quality, resulting in a final sample of 23 high-quality structural scans for subsequent volumetric analyses.

Image preprocessing and the segmentation pipeline

Hypothalamic segmentation was performed using the pre-trained hypothalamus_seg tool described by Billot et al. [20]. The hypothalamus_seg algorithm is a convolutional neural network–based segmentation framework distributed within the FreeSurfer software package and trained on multi-contrast MRI together with histological atlas data. The tool can be applied either to images processed through the full FreeSurfer recon-all pipeline or directly to individual T1-weighted MRI scans with approximately 1 mm isotropic resolution, without requiring prior surface-based cortical reconstruction. In the present study, the pre-trained convolutional neural network model was applied to the T1-weighted images using the standalone segmentation procedure without modification.

T1-weighted images were processed using the tool’s default preprocessing pipeline, which includes resampling to approximately 1.0 mm isotropic resolution when necessary and intensity normalization by linearly mapping the 0.5th and 99.5th intensity percentiles to 0 and 1, respectively. Because the acquired images were already at 1.0 mm isotropic resolution (voxel size = 1 × 1 × 1 mm³), no effective resampling was performed.

The pipeline generated bilateral labels for anatomically defined hypothalamic subunits, which were subsequently used for volumetric analyses. Posterior hypothalamic subunits were included as a negative-control region in exploratory analyses to assess anatomical specificity.

Hypothalamic subunit definitions

The automated segmentation tool parcellates the hypothalamus into five anatomically defined subunits per hemisphere, yielding a total of ten subunits across the brain. These subunits are delineated based on established anatomical landmarks and are designed to approximate underlying functional hypothalamic nuclei. Specifically, the anterior–inferior subunit primarily encompasses the supraoptic nucleus (SON) and the suprachiasmatic nucleus; the anterior–superior subunit primarily encompasses the paraventricular nucleus (PVN) and the preoptic area; the tuberal–inferior subunit includes the infundibular (arcuate) and ventromedial nuclei; the tuberal–superior subunit includes the dorsomedial nucleus; and the posterior subunit contains the mammillary bodies and the posterior hypothalamic nucleus.

For all hypothalamic subunits, volumetric measures were derived using soft volumes, calculated as the sum of voxel-wise posterior probabilities across each subunit. This probabilistic approach accounts for uncertainty at subunit boundaries and provides increased sensitivity to subtle volumetric variation compared with binary segmentation masks, which is particularly advantageous when analyzing small hypothalamic structures. For the primary analyses, we focused on PVN- and SON-associated subunits.

Accordingly, the soft-volume metric can reflect not only overall size but also inter-individual differences in contour and local boundary configuration, because graded posterior probability information is retained rather than collapsed into a hard binary label [56].

To assess anatomical specificity, hypothalamic subunits were further grouped into an anterior, magnocellular-rich system (SON- and PVN-associated subunits) and a posterior hypothalamic region encompassing the mammillary bodies. The posterior subunit was included as a negative-control region in statistical analyses to evaluate whether observed brain–behavior associations were preferentially related to the anterior neurosecretory system rather than reflecting global hypothalamic volumetric variation.

Statistical analysis and bootstrapped path analysis

All statistical analyses were performed in Python (version 3.10), using the statsmodels library for regression-based residualization and scipy.stats for correlation analyses.

Covariate handling. To minimize potential confounding effects of demographic factors and global brain size, hypothalamic subunit volumes and behavioral scores were adjusted for biological sex, age, and total intracranial volume (TIV) using linear regression. The resulting residuals were used for all subsequent partial correlation and path analyses unless otherwise specified. Total intracranial volume (TIV; cm³) was estimated from intracranial masks generated using the deepbet [29] applied to T1-weighted images.

Partial correlation analyses. Associations between hypothalamic subunit volumes and behavioral measures were assessed using partial correlation analyses based on residualized variables. Statistical significance was set at p < 0.05 (two-tailed).

Primary and exploratory analyses. Guided by an a priori magnocellular hypothesis, primary analyses focused on PVN- and SON-associated subunits. Associations involving other hypothalamic subunits were examined to assess anatomical specificity. As no nominally significant associations were observed for these regions, they were not included in subsequent path analyses.

Multiple comparisons.

To address multiple comparison concerns, p-values for the three pre-specified primary analyses (PVN–SON structural covariance, SON–AQ, and SON–ADHD) were adjusted using the Benjamini–Hochberg false discovery rate (FDR) procedure [30]. All three primary associations remained significant after FDR correction (ranked p-values: 0.006, 0.022, and 0.046; corresponding Benjamini–Hochberg critical values: 0.017, 0.033, and 0.050, respectively).

When FDR correction was applied to the full set of exploratory comparisons (10 hypothalamic subunits × 2 behavioral measures), no association survived correction (Supplementary Table S1), supporting the anatomical specificity of the findings to the a priori defined anterior magnocellular subunits.

Total intracranial volume (TIV) was estimated from intracranial masks generated using deepbet [29], consistent with established neuroimaging practice [31]. Prior validation studies have reported high concordance (r > 0.95) between extraction-based and atlas-based TIV estimates.

Path models were specified a priori based on the established neuroanatomical hierarchy of the hypothalamic magnocellular system. To further examine model directionality, an alternative reverse mediation model (SON → PVN → outcome) was tested. This reverse model yielded non-significant indirect effects for ADHD (β =  + 0.056, p = 0.662, 95% CI [− 0.247, + 0.277]) and autistic traits (β =  − 0.047, p = 0.725, 95% CI [− 0.378, + 0.195]), results that are consistent with the hypothesized forward direction (PVN → SON → outcome).

Bootstrapped path analysis. To evaluate the hypothesized anterior magnocellular axis linking PVN volume to behavioral measures via SON volume, a path model was constructed. Given the modest sample size (n = 23), nonparametric bootstrap resampling with 5000 iterations was employed to estimate indirect effects. Statistical significance was determined based on bootstrap-derived 95% percentile confidence intervals.

Data quality and distributional assumptions.

Prior to analysis, distributions of all primary variables were assessed for normality using the Shapiro–Wilk test and for potential outliers using the 1.5 × IQR rule. None of the variables showed significant deviation from normality (AQ: W = 0.922, p = 0.074; ADHD: W = 0.960, p = 0.461; left SON volume: W = 0.972, p = 0.733; left PVN volume: W = 0.983, p = 0.952). No outliers were detected in any primary variable.

Observed score ranges were as follows: AQ, 11–33 (median = 20); ADHD total score, 30–81 (median = 49); left SON volume, 6–28 mm³; and left PVN volume, 13–34 mm³. Distribution plots for all primary variables are provided in Supplementary Fig. 1.

Results

Descriptive statistics and quality control

The final sample consisted of 23 healthy adults (12 males, 11 females) with a mean age of 21.87 years (SD = 1.96). Behavioral measures showed approximately symmetric distributions for both Autism-Spectrum Quotient (AQ) scores (mean = 20.26, SD = 6.65) and ADHD symptom scores (mean = 50.39, SD = 13.33).

Automated segmentation successfully delineated hypothalamic subunits in all participants. Visual inspection of segmentation outputs confirmed appropriate localization of the anterior–inferior (SON-associated) and anterior–superior (PVN-associated) subunits relative to the optic chiasm and the third ventricle (Fig. 1; Supplementary Fig. 2). Mean volumes of the target subunits were consistent with established anatomical ranges (left PVN: 23.26 ± 4.80 mm³; left SON: 16.70 ± 5.75 mm³; Fig. 2). No significant hemispheric differences in subunit volumes were observed at the group level (all p > 0.05).

Fig. 1.

Automated segmentation of hypothalamic subunits on a representative T1-weighted structural MRI image. a Sagittal view, b coronal view, c axial view, and d three-dimensional rendering depicting the anterior–inferior (supraoptic nucleus–associated; yellow) and anterior–superior (paraventricular nucleus–associated; blue) subunits in relation to the optic chiasm and third ventricle

Fig. 2.

Distributions of hypothalamic subunit volumes. Group-level mean volumes and inter-individual variability are presented for left and right paraventricular nucleus– and supraoptic nucleus–associated subunits, demonstrating anatomically plausible ranges and no evidence of lateralized volume differences (ns, not significant)

Structural covariance of the anterior magnocellular system

Consistent with the hypothesized anterior magnocellular axis, partial correlation analysis controlling for biological sex, age, and total intracranial volume (TIV) revealed a significant positive association between left PVN- and left SON-associated subunit volumes (r = 0.51, p = 0.022; Fig. 3). This association was not attributable to demographic factors or global brain size.

Fig. 3.

Structural covariance between anterior hypothalamic subunits and associations with neurodevelopmental traits. a Partial correlation between left paraventricular nucleus– and supraoptic nucleus–associated volumes, indicating significant positive structural coupling consistent with an anterior magnocellular axis. b Association between left supraoptic nucleus–associated volume and Autism-Spectrum Quotient scores. c Association between left supraoptic nucleus–associated volume and ADHD symptom scores. d Association between left paraventricular nucleus–associated volume and Autism-Spectrum Quotient scores (not statistically significant). e Association between left paraventricular nucleus–associated volume and ADHD symptom scores (not statistically significant). All correlations are adjusted for sex, age, and total intracranial volume. *p < 0.05, **p < 0.01

A comparable, though not statistically significant, association was observed in the right hemisphere (r = 0.43, p = 0.059), suggesting a trend toward bilateral structural coherence of the magnocellular-rich anterior hypothalamic system.

To assess potential hemispheric specificity, additional analyses were conducted for right-hemisphere subunits. These analyses revealed no significant associations between right-hemisphere subunit volumes and behavioral measures (right SON–AQ: r =  − 0.267, p = 0.255; right SON–ADHD: r =  − 0.163, p = 0.493; right PVN–AQ: r =  − 0.249, p = 0.290; right PVN–ADHD: r =  − 0.014, p = 0.952). Given the absence of significant behavioral associations in the right hemisphere, subsequent primary analyses focused on left-hemisphere subunit volumes. This approach is also broadly consistent with prior work highlighting left-hemisphere involvement in social-communicative processing [32].

Hypothalamic volume–behavior associations

Primary analyses examined associations between left PVN- and SON-associated subunit volumes and dimensional measures of autistic traits (AQ) and ADHD symptoms, guided by the magnocellular hypothesis. Partial correlations were computed using residuals adjusted for biological sex, age, and TIV. Right-hemisphere associations were examined exploratorily and did not reach statistical significance (right SON–AQ: r =  − 0.267, p = 0.255; right SON–ADHD: r =  − 0.163, p = 0.493; right PVN–AQ: r =  − 0.249, p = 0.290; right PVN–ADHD: r =  − 0.014, p = 0.952).

SON-associated subunits and behavioral traits

Left SON-associated subunit volume showed a significant inverse association with AQ scores (r =  − 0.451, p = 0.046; Fig. 3), indicating that smaller SON-associated volumes were related to higher levels of autistic-like traits. A stronger inverse association was observed between left SON-associated volume and ADHD symptom severity (r =  − 0.589, p = 0.006; Fig. 3).

In contrast, left PVN-associated subunit volume was not significantly associated with either AQ (r =  − 0.300, p = 0.199) or ADHD symptom scores (r =  − 0.223, p = 0.346).

Exploratory analyses of other hypothalamic subunits did not reveal any nominally significant associations and were therefore not pursued further or included in subsequent modeling.

Anatomical specificity (negative control)

To assess anatomical specificity, posterior hypothalamic subunit volumes encompassing the mammillary bodies were examined as a negative-control region. No significant associations were observed between left mammillary body volume and AQ (r =  − 0.052, p = 0.826) or ADHD symptom scores (r =  − 0.233, p = 0.322; Supplementary Fig. 3).

Importantly, the associations between left SON-associated volume and behavioral measures remained significant when posterior hypothalamic volume was included as an additional covariate (AQ: r =  − 0.486, p = 0.035; ADHD: r =  − 0.559, p = 0.013), supporting the anatomical specificity of these effects to anterior magnocellular-rich subunits. In this analysis, partial correlations were computed using residualized variables adjusted simultaneously for biological sex, age, total intracranial volume (TIV), and left posterior hypothalamic (mammillary body) volume.

Bootstrapped path analysis

To further examine network-level relationships, bootstrapped path analyses were conducted to test the hypothesized anterior magnocellular axis linking PVN-associated volume to behavioral traits via SON-associated volume. All paths were estimated using residuals adjusted for biological sex, age, and TIV, with 5000 bootstrap iterations used to estimate indirect effects (Fig. 4).

Fig. 4.

Path analytic models linking anterior hypothalamic structure to ASD traits and ADHD symptoms. Bootstrapped structural equation models (5,000 iterations) depict positive coupling between paraventricular nucleus– and supraoptic nucleus–associated volumes and show an indirect association between paraventricular nucleus–associated volume and ASD/ADHD measures via supraoptic nucleus–associated volume, after adjustment for sex, age, and total intracranial volume. Non-significant paths are labeled as ns; *p < 0.05, **p < 0.01

Model A: autistic traits

The path from left PVN-associated volume to left SON-associated volume was significant (β = 0.51, p = 0.013, 95% CI [0.257, 0.725]). The path from left SON-associated volume to AQ scores was not significant (β =  − 0.36, p = 0.089, 95% CI [− 0.644, 0.004]). The indirect association between PVN-associated volume and AQ scores via SON-associated volume approached, but did not reach, statistical significance (β_ind =  − 0.19, p = 0.052, 95% CI [− 0.354, 0.003]). No direct association between PVN-associated volume and AQ scores was observed.

Model B: ADHD symptoms

In contrast, a significant indirect association between PVN-associated volume and ADHD symptoms was observed via SON-associated volume. The PVN–SON path was significant (β = 0.51, p = 0.013, 95% CI [0.257, 0.725]), as was the SON–ADHD path (β =  − 0.57, p = 0.005, 95% CI [− 0.786, − 0.202]). The resulting indirect effect was significant (β_ind =  − 0.29, p = 0.006, 95% CI [− 0.498, − 0.074]). No direct association between PVN-associated volume and ADHD symptom scores was observed.

Discussion

Structural coherence of the anterior magnocellular system

The present study provides in vivo evidence for structural coherence between anterior hypothalamic subunits approximating the paraventricular nucleus (PVN) and supraoptic nucleus (SON). The significant positive association between PVN- and SON-associated volumes is consistent with the known anatomical and developmental organization of the hypothalamo–neurohypophyseal system, in which these regions form a tightly coupled neuroendocrine network. This finding aligns with prior work emphasizing functional differentiation within this system, whereby PVN-associated regions integrate diverse afferent inputs, while SON-associated regions are specialized for neurosecretory output [33–36].

Importantly, our path analyses indicate that SON-associated volume occupies a proximal position linking anterior hypothalamic structure to behavioral variation, whereas PVN-associated volume contributes indirectly through its structural relationship with the SON. This pattern is compatible with a hierarchical network organization in which variability in downstream magnocellular-rich regions is more directly related to behavioral traits than upstream integrative regions [34]. Such an organization is consistent with models proposing that the functional capacity of neurosecretory output regions may represent a limiting factor in neuroendocrine regulation.

The anatomical specificity of this network was supported by negative-control analyses, which revealed no associations between posterior hypothalamic (mammillary) volumes and behavioral measures. This finding suggests that the observed associations are not attributable to generalized hypothalamic or global brain volume effects, but rather reflect regionally specific variation within anterior magnocellular-rich subunits. Furthermore, by employing sex-adjusted residuals, we minimized potential confounding from baseline differences in brain volume, indicating that the observed structural–behavioral associations reflect region-specific volumetric variation independent of global brain scaling related to biological sex or intracranial volume.

Anterior hypothalamic structure and autistic traits

We observed a significant inverse association between SON-associated volume and autistic traits, as measured by the Autism-Spectrum Quotient. This finding is consistent with theoretical frameworks implicating hypothalamic neuropeptide systems in social cognition and stress regulation [37–39]. While oxytocin and vasopressin have been proposed as key mediators of social behavior, the present study extends this literature by providing structural evidence that variation in anterior hypothalamic subunits is related to subclinical autistic traits in humans. Recent research in animal models further suggests that oxytocin may influence social deficits, potentially through mechanisms such as oligodendrocyte differentiation [40].

However, the mediation pathway linking PVN-associated volume to autistic traits via SON-associated volume did not reach conventional statistical significance. This suggests that, although the direction of effects is consistent with the hypothesized anterior magnocellular network, the association with autistic traits may be weaker or more heterogeneous than that observed for ADHD-related symptoms. These findings underscore the importance of interpreting structural–behavioral associations within a dimensional and transdiagnostic framework, particularly given the known overlap between autistic and attentional traits.

ADHD symptoms and neuroendocrine network organization

The most robust behavioral association observed in this study was the inverse relationship between SON-associated volume and ADHD symptom severity. This finding expands prevailing neurobiological models of ADHD, which have traditionally emphasized dopaminergic fronto-striatal circuits [41–43], by implicating anterior hypothalamic structure as an additional contributor to attentional and impulse-regulation processes. Recent evidence also highlights that oxytocin levels may correlate negatively with impulsivity and inattention, suggesting a hypo-oxytocinergic component in ADHD pathophysiology [44–46].

Animal and human studies have demonstrated extensive interactions between hypothalamic neuropeptide systems and mesolimbic dopamine pathways, suggesting a plausible mechanism by which variation in SON-associated structure could influence attentional regulation [47–49]. Within this framework, reduced SON-associated volume may reflect diminished capacity of neuroendocrine modulation of reward and salience processing, contributing to increased impulsivity and attentional instability. We propose that a robust SON volume reflects a high-capacity system that may exerts a “stabilizing” modulatory influence on dopaminergic firing [50]. Importantly, this association was specific to anterior hypothalamic subunits and was not observed in posterior regions, supporting its anatomical specificity.

Regarding the measurement instruments used, the Autism-Spectrum Quotient (AQ) was specifically developed to quantify autistic traits as a continuous distribution in community samples [27], whereas the Conners’ Adult ADHD Rating Scale (CAARS) was originally developed for clinical assessment of ADHD symptoms. Both instruments have nevertheless demonstrated utility as dimensional measures in adult community samples and have published normative data [25, 51]. We acknowledge that these instruments differ in their theoretical foundations, and that future studies employing measures designed explicitly for dimensional assessment in community populations would further strengthen cross-trait comparisons.

Structural covariance as an intermediate phenotype

Increasingly, neuroimaging research conceptualizes coordinated structural variation across brain regions as an intermediate phenotype linking neurodevelopmental processes to behavioral outcomes [52, 53]. In this context, the present findings identify a specific anterior hypothalamic structural network associated with dimensional ADHD and ASD traits, extending prior large-scale studies that have reported subcortical anatomical variation across neurodevelopmental conditions [24, 54].

Limitations and future directions

The predominance of left-sided associations observed in this study is consistent with prior work on hemispheric specialization related to social and communicative processing [32, 55]. However, given the modest sample size and the presence of trend-level associations in the right hemisphere, conclusions regarding lateralization should be considered preliminary. Future studies employing larger samples and higher-resolution imaging will be essential to clarify hemispheric asymmetries within hypothalamic networks.

Several directions for future research emerge from these findings. High-field (7T) MRI could provide improved delineation of magnocellular and parvocellular subdivisions within anterior hypothalamic regions. Combining structural imaging with dynamic neuroendocrine assays during behavioral tasks would further elucidate structure–function relationships. In addition, while the present study relied on probabilistic volumetric measures, dedicated shape analyses—such as spherical harmonics point distribution modeling and related surface-based approaches—could complement soft-volume estimates by localizing regional morphological deformations in future work [57]. Finally, longitudinal studies are needed to determine whether variation in SON-associated volume represents a stable developmental trait or reflects experience-dependent neuroplasticity.

Additional limitations warrant acknowledgment. First, the modest sample size (n = 23) limits statistical power and generalizability; accordingly, the present findings should be regarded as preliminary and require replication in larger independent cohorts.

Second, cognitive ability was not included as a covariate in the primary analyses. To address this point, we conducted a sensitivity analysis incorporating WAIS indices (Processing Speed Index: mean = 115.1, SD = 8.8; Verbal Comprehension Index: mean = 113.1, SD = 9.1; Perceptual Organization Index: mean = 111.1, SD = 11.2) as an additional covariate. These indices were not significantly correlated with left SON volume (r = 0.125, p = 0.570), AQ scores (r =  − 0.086, p = 0.697), or ADHD scores (r =  − 0.135, p = 0.539). Importantly, the SON–ADHD association remained significant after adjustment for WAIS indices (r =  − 0.588, p = 0.008), and the SON–AQ effect size was essentially unchanged (Δr = 0.001; adjusted p = 0.053), with the slight shift in significance likely reflecting reduced degrees of freedom. Full WAIS-adjusted results are reported in Supplementary Table S2.

Third, neurodevelopmental traits were assessed using self-report measures without independent clinical evaluation. Fourth, the cross-sectional design precludes causal inference.

Conclusion

This study provides novel in vivo evidence linking anterior hypothalamic subunit structure to dimensional traits of ADHD and ASD in a non-clinical population. Using automated deep learning–based segmentation of high-resolution MRI, we identified a coherent structural relationship between PVN- and SON-associated subunits and demonstrated that variation in SON-associated volume is systematically related to social and attentional traits. The specificity of these associations to anterior magnocellular-rich regions, and their absence in posterior hypothalamic subunits, underscores the relevance of hypothalamic network organization in neurodevelopmental variation. Together, these findings highlight the potential of hypothalamic subunit volumetry as an anatomically grounded approach for advancing precision psychiatry.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1 ^{(1,016.1KB, docx)}

Author contributions

U.C.: Conceptualization, Data curation, Formal Analysis, Methodology, Resources, Writing—original draft, Writing—review & editing. S.K.: Conceptualization, Acquiring behavioral data, Formal Analysis, Writing—review & editing. Y.S: Conceptualization, Data curation, Formal Analysis, Investigation, Methodology, Software, Project administration, Validation, Visualization, Writing—original draft, Writing—review & editing. M.F.: Formal Analysis, Investigation, Writing-review & editing. S.O.: Conceptualization, Data curation, Formal Analysis, Funding, Investigation, Project administration, Resources, Supervision, Writing—review & editing.

Funding

This study was supported by JSPS KAKENHI (Grant Number 25K03427, 24K00059 and 23K25200), Cooperative Study Program (25NIPS607&26NIPS608) of National Institute for Physiological Sciences, the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant (RS-2026-25516696) funded by the Ministry of Science and ICT (MSIT) of the Republic of Korea, Brain/MINDS2.0 from AMED JP25wm0625001, and MEXT/CURE JPMXP1323015488 (Spin-L program No, spin25XN026).

Data availability

The datasets generated and/or analysed during the current study are not publicly available due to ethical and privacy restrictions related to participant confidentiality but are available from the corresponding author on reasonable request.

Code availability

The Python analyses in this study were implemented using adapted scripts based on existing open-source libraries and are not available as an integrated, shareable code package. The hypothalamic segmentation tool (hypothalamus_seg) is publicly available at https://github.com/BBillot/hypothalamus_seg.

Declarations

Ethics approval

This study was approved by the Institutional Review Board of Tohoku Fukushi University (Approval No. RS190607) and was conducted in accordance with the principles of the Declaration of Helsinki.

Consent to participate

Written informed consent was obtained from all participants prior to their participation in the study. Participants were informed about the purpose of the study, the procedures involved, potential risks, and their right to withdraw from the study at any time without any negative consequences.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Clinical trial number

Not applicable.

Dual publication

The submitted work has not been published elsewhere in any form or language (partially or in full).