Automated 3D nnU-Net Segmentation of the Parasagittal Dura in Children with Autism Spectrum Disorder Using Clinical 3D T2-FLAIR

Automated 3D nnU-Net Segmentation of the Parasagittal Dura in Children with Autism Spectrum Disorder Using Clinical 3D T2-FLAIR | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Automated 3D nnU-Net Segmentation of the Parasagittal Dura in Children with Autism Spectrum Disorder Using Clinical 3D T2-FLAIR Francesca Castellotti, Tommaso Ciceri, Chiara Girardi, Elisa Mani, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9556003/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract Background The parasagittal dura (PSD) is a specialized meningeal structure that contains lymphatic vessels and contributes to cerebrospinal fluid (CSF) drainage and neuroimmune surveillance. In adults, aging and neurodegenerative diseases have been associated with increased PSD volume, suggesting that it may represent a potential biomarker of neurodegeneration. However, its functional role and the extent to which PSD volume may reflect pathological processes during early brain development remain poorly understood. Furthermore, no automated tools currently exist to segment the PSD in the very young developing brain. We present a pediatric automated tool for PSD segmentation and investigate PSD volume in children with autism spectrum disorder (ASD) and typically developing (TD) controls. Methods We developed a fully automated segmentation framework based on a 3D nnU-Net applied to clinical 3D T2-FLAIR MRI. The model was trained on manually annotated pediatric ASD scans (n = 144) and evaluated on an independent ASD test set (n = 56). To assess generalizability, the trained model was applied to independent typically developing (TD) cohorts, including an external dataset acquired with different imaging parameters. Total and regional PSD volumes (prefrontal, fronto-parietal, occipital) were quantified and compared between ASD and TD children, and correlations with age and brain volumes were examined. Results The segmentation framework achieved high performance in the ASD cohort (Dice–Sørensen coefficient (DSC) = 0.93 ± 0.02). Comparable performance was observed in the TD cohort (DSC = 0.90 ± 0.02), including the cohort acquired with different imaging parameters (DSC = 0.85 ± 0.04), indicating good generalizability. Performance remained consistent across PSD regions, with slightly higher accuracy in central segments. No significant differences in PSD volumes were found between ASD and TD groups. PSD volume is correlated with CSF (and ea-CSF), supporting a link with fluid compartment dynamics. Conclusions We present a publicly available reliable automated tool for whole-brain PSD segmentation in the pediatric population using clinical 3D T2-FLAIR sequences. Our results show no differences in PSD volumes between ASD and TD groups, while consistently demonstrating a positive correlation with CSF-related compartments. This tool provides a scalable approach for studying the role of PSD in neurofluid circulation during brain development. Parasagittal dura Autism spectrum disorder Pediatric Deep learning MRI segmentation neurofluids Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Background The parasagittal dura (PSD) represents a specialized mesenchymal stromal tissue containing critical anatomical structures that facilitate the efflux of cerebrospinal fluid (CSF) and solutes from the central nervous system (CNS) (1). Anatomically positioned between the dual layers of the cerebral dura mater and surrounding the superior sagittal sinus, the PSD is home to meningeal lymphatic vessels (2,3). These lymphatic channels are essential for neuro-immune surveillance and contribute to the preservation of CNS homeostasis by mediating the drainage of immune cells and antigens toward peripheral lymph nodes (3,4). PSD is readily identifiable on standard magnetic resonance imaging (MRI) sequences and may represent a useful in vivo target for investigating neurofluid dynamics (2). Volumetric quantification of the PSD has been performed by employing semi-automated and automated segmentation tools on MR images, primarily in adult populations (5–9). Empirical data show that PSD volume increases linearly with age in healthy adults, typically reaching a plateau around age 70 (5,9). Furthermore, a recent study applying automatic segmentation on Alzheimer’s disease cohorts has revealed a significant positive association between PSD volume and amyloid-β burden as quantified by positron emission tomography (7). The authors suggest that a hypertrophic PSD may constitute a compensatory response to the progressive impairment of lymphatic drainage associated with senescence, potentially serving as an MR surrogate marker for underlying neuroinflammatory processes and altered neurofluid kinetics (7,9). Autism spectrum disorder (ASD) represents a neurodevelopmental condition with a multifactorial etiology (10). Recent evidence indicates that elevated extra-axial cerebrospinal fluid (ea-CSF) volume may serve as a potential early-stage neuroimaging biomarker in neonates with high risk of developing ASD and young children with a diagnosis of ASD (11,12). Such volumetric expansions are hypothesized to reflect dysregulation of CSF kinetics (12,13). Consequently, impaired drainage, possibly involving the meningeal lymphatic system and PSD, may affect neuroimmunologic processes and contribute to the aberrant neurodevelopmental trajectories characteristic of ASD (14,15). To explore this hypothesis, in our previous investigation, a semi-automated 2D U-Net-based tool was employed to segment the PSD from 3D T2-FLAIR sequences and quantify its volume in a pediatric ASD cohort aged 2–8 years (16). We found a significant positive correlation between PSD, CSF, and ea-CSF volumes, confirming the previously established functional role of the PSD in CSF homeostasis (16). Additionally, the association identified between PSD volume and the degree of developmental delay suggests a role for this structure in the clinical manifestation of ASD (16). However, the use of a 2D architecture may not have fully captured the three-dimensional (3D) spatial continuity of this complex structure. Moreover, for practical purposes, only a central segment of the PSD was quantified, limiting anatomical completeness. Additionally, the segmentation required substantial manual post-processing to correct inaccuracies, introducing potential inter- and intra-rater variability and limiting its widespread usage among the scientific community. Currently, the only publicly accessible fully automated 3D U-Net–based tool for PSD segmentation uses high-resolution 3D T2-weighted imaging (6). However, since the model was trained and validated on a dataset comprising individuals aged 11–83 years, the same authors reported a tendency to overestimate the PSD volume when applied to younger pediatric cohorts (ages 5–10) (6). This highlights the need for age-specific segmentation tools that account for the unique neuroanatomical characteristics of the developing brain. Furthermore, reliance on high-resolution T2-weighted sequences may limit the method's broad clinical translational utility, as these acquisitions are not routinely included in standard pediatric neuroimaging protocols. To address these methodological limitations and the absence of automated tools for pediatric populations, we developed a fully automated 3D nnU-Net–based method for whole-PSD segmentation from standard pediatric 3D T2-FLAIR images, trained on a large manually annotated ASD cohort. We then evaluated its generalizability on independent datasets of typically developing children, including an external cohort acquired at Zurich Children’s Hospital with different imaging parameters. Using this approach, we quantified and compared total and regional PSD volume in both ASD and typically developing (TD) groups. A regional subdivision was adopted to account for the spatial heterogeneity of the PSD along the superior sagittal sinus, with prior studies showing non-uniform distribution and development across frontal, central, and posterior regions (6,17). This approach may enable a more anatomically grounded and developmentally sensitive characterization of PSD in pediatric population. Materials and methods This study aimed to develop, validate, and apply a fully automated deep learning–based framework for PSD segmentation in pediatric brain MRI. The overall workflow is illustrated in Fig. 1 . Briefly, 3D T2-FLAIR images were pre-processed and manually annotated to generate ground-truth PSD masks. A 3D nnU-Net model was trained on the ASD cohort and evaluated on independent datasets, including TD children and an external TD cohort acquired at Zurich Children’s Hospital. The resulting segmentations were subsequently used for regional PSD subdivision and volumetric analyses comparing ASD and TD groups. Datasets and preprocessing ASD Cohort Children diagnosed with ASD who underwent MRI between July 2021 and July 2025 were included in this study. Inclusion criteria were a confirmed diagnosis of ASD and age between 2 and 12 years. Exclusion criteria included MRI sequences affected by motion artefacts and positive for other pathologies or malformations. A total of 200 participants (mean age 4.62 ± 1.6; 33 females) were included in the study. Diagnosis of ASD was established by a multidisciplinary team at the Child Psychopathology Unit of IRCSS E. Medea (Bosisio Parini, Italy) according to DSM-V criteria (American Psychiatric Association, 2013). Diagnostic assessment included the Autism Diagnostic Interview–Revised (ADI-R) (18), administered to parents, and the Autism Diagnostic Observation Schedule–2nd edition (ADOS-2) (19) conducted with the child. TD Cohort The TD subjects were recruited from two institutes: IRCCS Eugenio Medea (TDM) and Zurich Children’s Hospital (TDZ). The initial TDM dataset included 35 subjects scanned between March 2022 and December 2025. Inclusion criteria were the absence of a diagnosis of ASD or developmental delay and age between 2 and 12 years. Exclusion criteria included motion-related artefacts affecting 3D T1-weighted and/or 3D T2-FLAIR images. The final TDM sample comprised 27 children (mean age 7.01 ± 2.8 years, 11 females). The initial TDZ dataset comprised 66 subjects. Inclusion criteria were the absence of ASD or developmental delay and age between 2 and 12 years. Exclusion criteria included motion artefacts affecting 3D T1-weighted and/or 3D T2-FLAIR images, absence of 3D T2-FLAIR sequences, and presence of neurological disorders or structural brain abnormalities. The final TDZ sample consisted of 33 children (mean age 6.60 ± 1.73 years, 19 females). Data preprocessing All ASD and TDM participants were scanned on a 3T MRI system (Achieva dStream; Philips Medical Systems) using a 32-channel head coil at the Diagnostic Imaging and Neuroradiology Unit of our institute. The MRI protocol for ASD and TDM participants included two sequences: (a) a 3D T2-FLAIR sequence with the following acquisition parameters: repetition time (TR) = 4800–6000 ms, echo time (TE) = 289–344 ms, inversion time (TI) = 1650–1860 ms, and acquisition matrix = 339 × 480 × 480; and (b) a 3D T1-weighted MPRAGE sequence with TR = 7–8.5 ms, TE = 3.2–3.96 ms, acquisition matrix = 256 × 256, and an isotropic voxel size of 1 × 1 × 1 mm³. The TDZ cohort was acquired using a 3D T2-FLAIR sequence with the following parameters: TR = 7000 ms, TE = 130 ms, TI = 1910–1960 ms, and acquisition matrix = 260 × 248 × 248. The 3D T1-weighted sequence was a SPGR BRAVO acquired sagittally with TR = 10 ms, TE = 4 ms, TI = 450 ms, acquisition matrix = 352 × 256, and slice thickness = 1 mm. 3D T2-FLAIR images were preprocessed using N4 bias field correction to reduce intensity non-uniformities (20). T1-weighted images were preprocessed using FreeSurfer’s automated recon-all pipeline, yielding regional brain volume estimates. CSF and ea-CSF masks were obtained using an in-house pipeline described previously (21). Automated PSD segmentation Annotation protocol of PSD Ground-truth PSD segmentations on ASD, TDM and TDZ were generated using a human-in-the-loop approach, similar to strategies previously described in neuroimaging segmentation study (22). Manual delineation was performed on 3D T2-FLAIR images. Initial segmentations were obtained from our previously developed 2D U-Net trained to segment the central portion of the PSD (16). These segmentations were systematically reviewed and refined under the supervision of a senior neuroradiologist (N.A.), ensuring anatomical accuracy and consistency across the dataset. Additional trained operators contributed to the refinement process following a standardized protocol, using FSLeyes to correct potential inaccuracies and extend the labels anteriorly and posteriorly to cover the full anatomical extent of the PSD. In earlier work, the PSD segmentation was limited to a central segment defined by a 60° arc along the cranial circumference, passing through the anterior and posterior commissures (16). In the present study, this constraint was removed, resulting in the full-length PSD segmentations (Fig. 2 ). Network architecture and training strategy Automated PSD segmentation was performed using the nnU-Net framework (23), a self-configuring deep learning approach based on a 3D U-Net architecture that automatically adapts network design, pre-processing, and training strategies to the input data. The network has demonstrated competitive performance across a wide range of medical image segmentation tasks (23). Given that children with ASD represented the largest manually annotated cohort and constituted the primary focus of the present study, the model was trained on the ASD dataset. The ASD dataset with manual segmentations (n = 200) was split into a training set (n = 144,70%) and an independent test set (n = 56, 30%). Model development followed the standard nnU-Net pipeline, including five-fold cross-validation on the training set. The final predictions were obtained using the five-fold cross-validation ensemble, by averaging the outputs of the five trained models. The final trained model was evaluated exclusively on the independent test set, which was not used during training or validation. Evaluation of the segmentation network To assess whether the model generalized beyond the ASD training population and captured stable anatomical features of the PSD, the trained nnU-Net was applied to two independent cohorts of TD children (TDM and TDZ), which were not included in training. Notably, the TDZ cohort was acquired using different 3D T2-FLAIR imaging parameters, allowing assessment of the model’s performance across clinical imaging protocols. Model performance was evaluated on the independent ASD test set and on the TD cohorts using the Dice similarity coefficient (DSC), volumetric similarity coefficient (VS), and Hausdorff distance (HD). Segmentation performance was also evaluated separately for three anatomical regions — prefrontal, fronto-parietal and occipital, defined by subdividing the PSD post hoc along the cranial circumference, using a 60° arc defined by the anterior and posterior commissures (AC–PC) as the central reference, as described previously (16). Qualitative evaluation across cohorts was performed to assess whether the network captured the full extent of PSD anatomy. The training was conducted in parallel for each fold on two 48GB NVIDIA L40S. PSD assessment Total and regional PSD volumes (prefrontal, fronto-parietal, and occipital), defined using the regional subdivision described above, were extracted for each subject to perform volumetric analysis. Statistical analysis Statistical analyses were performed in Python (version 3.13.2) using the pandas, NumPy, SciPy, statsmodels, and neuroCombat libraries. All tests were two-tailed, and p < 0.05 was considered statistically significant. Analyses were conducted on the independent ASD test set and both TD cohorts (TDM and TDZ). To account for inter-site variability in volumetric estimates related to differences in MRI acquisition parameters between TDM and TDZ, ComBAT harmonization was applied to all volumetric measures (PSD, CSF, ea-CSF, and ICV) using the neuroCombat implementation (24). Age, sex, and diagnostic group were included as biological covariates to preserve during harmonization. Following harmonization, the two TD cohorts were merged into a single TD group (N = 60) for all subsequent statistical analyses. Welch’s t-tests were used to assess group differences between ASD and TD participants in age and cerebral volumes. Chi-square test was used to compare sex distribution between groups. Exploratory correlation analyses were conducted using Spearman’s rank correlation coefficient to examine the relations between age, PSD, and cerebral volumes, as well as between total and regional PSD volumes and cerebral volumes. To assess group differences in PSD measures while accounting for potential confounders, ordinary least squares (OLS) linear regression models were fitted with diagnostic group (ASD vs TD) as the main predictor, adjusting for age, sex, and CSF volume according to the following model: $$\:PSD\text{}\text{}={\beta\:}_{0}\text{}+{\beta\:}_{g}\text{}\cdot\:\text{G}\text{r}\text{o}\text{u}\text{p}+{\beta\:}_{a}\text{}\cdot\:\text{A}\text{g}\text{e}+{\beta\:}_{s}\text{}\cdot\:\text{S}\text{e}\text{x}+{\beta\:}_{c}\text{}\text{}\cdot\:\text{C}\text{S}\text{F}+\epsilon\text{}$$ where PSD is the dependent variable, β 0​ is the intercept, β i ​are the regression coefficients associated with each predictor, Group denotes diagnostic group (ASD vs TD), Age is age, Sex is biological sex, CSF is cerebrospinal fluid volume, and ϵ is the residual error. Because the TD group was not age-matched with the ASD group, an additional sensitivity analysis was performed to evaluate the potential confounding effect of age. Specifically, the regression analysis was repeated on a restricted subsample including only participants aged 2 to 9 years, using the same covariates as in the main models. Results Network Segmentation Performance In the ASD cohort, used for model development, the network achieved high segmentation accuracy, with a DSC of 0.93 ± 0.02, a VS of 0.98 ± 0.01, and a HD of 12.5 ± 9.0 (Table 1 ). To evaluate model generalizability, the trained network was subsequently applied to two independent cohorts ofTD children. In the TDM cohort, the model achieved a DSC of 0.90 ± 0.02, a VS of 0.98 ± 0.03, and a HD of 15.0 ± 5.9 (Table 2 ). In the TDZ cohort, acquired using different imaging parameters, segmentation performance remained robust, with a DSC of 0.85 ± 0.04, a VS of 0.96 ± 0.04, and a HD of 18.4 ± 8.2 (Table 2 ). A region-wise analysis was further performed to assess segmentation performance across the prefrontal, fronto-parietal, and occipital portions of the PSD (Table 1 , 2 ). The model achieved consistently high DSC and VS across regions, with slightly higher segmentation accuracy observed in the central segment. In the ASD test set, DSC ranged from 0.90 ± 0.05 in the prefrontal region to 0.94 ± 0.02 in the fronto-parietal region, with intermediate values in the occipital region (0.92 ± 0.03). Similar patterns were observed in the TD cohorts, with slightly higher segmentation accuracy in the fronto-parietal segment. Table 1 Segmentation performance of the proposed nnU-Net model on the ASD dataset. Metrics are reported for both the Total PSD and its regional subdivisions (prefrontal, fronto-parietal, and occipital). Results are expressed as Dice–Sørensen coefficient (DSC), volumetric similarity (VS), and Hausdorff distance (HD), reported as mean ± standard deviation. Total DSC VS HD (mm) 0.93 ± 0.02 0.98 ± 0.01 12.5 ± 9.0 Prefrontal 0.90 ± 0.05 0.95 ± 0.04 9.14 ± 9.79 Fronto-Parietal 0.94 ± 0.02 0.99 ± 0.01 6.36 ± 1.76 Occipital 0.92 ± 0.03 0.97 ± 0.04 9.78 ± 5.08 Table 2 Segmentation performance of the proposed nnU-Net model on the TDM and TDZ datasets . Metrics are reported for both the Total PSD and its regional subdivisions (prefrontal, fronto-parietal, and occipital). Results are expressed as Dice–Sørensen coefficient (DSC), volumetric similarity (VS), and Hausdorff distance (HD), reported as mean ± standard deviation. TDM TDZ DSC VS HD (mm) DSC VS HD (mm) Total 0.90 ± 0.02 0.98 ± 0.03 15.0 ± 5.9 0.85 ± 0.04 0.96 ± 0.04 18.4 ± 8.2 Prefrontal 0.88 ± 0.04 0.94 ± 0.05 7.9 ± 3.1 0.76 ± 0.10 0.87 ± 0.10 14.2 ± 11.0 Fronto-Parietal 0.92 ± 0.02 0.98 ± 0.02 7.8 ± 2.7 0.86 ± 0.05 0.96 ± 0.04 8.7 ± 3.7 Occipital 0.87 ± 0.06 0.93 ± 0.07 12.6 ± 6.1 0.78 ± 0.16 0.89 ± 0.17 18.1 ± 16.2 Qualitative inspection across cohorts confirmed that the network accurately captured the full anatomical extent of the PSD despite the considerable inter-subject variability characteristic of this structure. Figure 3 and Fig. 4 represent PSD segmentation for the entire PSD and its regional subdivisions respectively. PSD assessment Participant characteristics The statistical analyses included 56 children with ASD and 60 TD children (27 from IRCCS E. Medea and 33 from Zurich Children’s Hospital), following ComBAT harmonization of volumetric measures across sites. Demographic characteristics and global brain volumetric measures are summarized in Table 3 , including age, sex distribution, ICV, CSF, and ea-CSF. The groups were not age-matched, with TD participants being older on average (p 0.05). Total and regional PSD volumes (prefrontal, fronto-parietal, and occipital) for the ASD and TD groups are reported in Table 4 . Table 3 Demographic characteristics and global brain volumetric measures of the ASD and TD groups . Values are reported as mean ± standard deviation. Age (years) ASD (N = 56) TD (N = 60) p-value 5.1 ± 1.5 6.8 ± 2.3 < 0.0001 Sex (M/F) 45 / 11 30 / 30 0.001 eaCSF (cm³) 66.5 ± 17.7 67.0 ± 21.9 0.905 CSF (cm³) 128.4 ± 25.7 128.9 ± 35.4 0.927 ICV (cm³) 1378.3 ± 150.4 1353.9 ± 133.7 0.358 Table 4 Total and regional (frontal, frontoparietal, occipital) PSD volumes in the ASD and TD groups. Values are expressed as mean ± SD. PSD Total (cm³) ASD (N = 56) TD (N = 60) 4.9 ± 1.6 4.8 ± 1.2 PSD Frontal (cm³) 0.4 ± 0.4 0.5 ± 0.4 PSD Fronto-Parietal (cm³) 3.3 ± 1.0 3.3 ± 0.9 PSD Occipital (cm³) 1.1 ± 0.5 1.0 ± 0.5 Age-related effects on brain volumes In the ASD group, significant positive associations with age were observed for ICV (Spearman’s ρ = 0.39, p = 0.003). No significant correlations with age were found for CSF, ea-CSF, or PSD volumes (all p > 0.05). In the TD group, no significant age-related effects were observed in ICV, CSF, ea-CSF, or PSD volumes (all p > 0.05) (Tables in supplementary materials S1, S2). The absence of significant age-related effects on PSD volumes in both groups is further illustrated in Supplementary material (Figure S1 ), which shows scatter plots of total and regional PSD volumes against age, with fitted quadratic curves remaining essentially flat across all regions in both ASD and TD children. Associations between PSD and brain volumes Total PSD volumes showed significant positive correlations with CSF, ea-CSF, and ICV volumes in both ASD (p = 0.005, p = 0.002, and 0.03, respectively) and TD (p = 0.001, p = 0.004, and p = 0.03, respectively). Regional analyses demonstrated spatially heterogeneous patterns. In the ASD group, central and occipital PSD volumes exhibited the strongest positive correlations with CSF (p = 0.004, p = 0.01) and ea-CSF (p = 0.001, p = 0.004). In contrast, frontal PSD volume showed no significant associations with any brain volume measures. In the TD group, only the central PSD volume was significantly correlated with CSF (p = 0.002) and ea-CSF (p = 0.01). Notably, central PSD volume was also positively associated with ICV in both cohorts (p = 0.02 (ASD), p = 0.02 (TD)). These patterns are summarized in the correlation heat maps in Fig. 5 , which provide a visual overview of the relationships among PSD measures, cerebral volumes, and age in the ASD and TD groups. Linear Model The linear regression analysis performed on total PSD volume showed no significant differences between the ASD and TD groups (β = -0.002, p = 0.9). Similarly, sex and age were not significantly associated with PSD volume. In contrast, CSF volume showed a significant positive association with total PSD (β = 0.016, p < 0.001), indicating that higher CSF volumes were associated with larger PSD volumes. No significant group differences were observed in regional PSD volumes. The age-restricted sensitivity analysis (2–9 years) confirmed the main findings, showing no significant difference in PSD between ASD and TD groups (β = −0.061, p = 0.8). CSF volume continued to remain significantly associated with PSD (p < 0.001). Detailed results are provided in Supplementary tables (Table S3-S7). Discussion In this study, we designed and validated a fully automated 3D nnU-Net model to segment PSD in pediatric brain. This new model expands the analysis to cover the entire anatomical extent of the structure, whereas our previous approach was limited to a central portion of the PSD (16). The proposed model demonstrated reliable performance in both children with ASD and TD cohorts, including an independent external cohort acquired using different 3D T2-FLAIR imaging parameters. This high performance might be partially attributed to the human-in-the-loop framework used during ground-truth generation, in which an experienced neuroradiologist refined the initial U-Net predictions (22). While this expert-led framework introduces a certain methodological bias, literature indicates that such bias can be helpful when training models with limited data, potentially contributing to the model's reliability (25). The mean PSD volume observed in our pediatric cohort provides an initial reference for understanding the developmental trajectory of this structure. These values are consistent with previous lifespan studies reporting that the PSD space accounts for approximately 0.30% of total ICV during adolescence (6). The volumes measured in our cohort likely represent an early stage of the progressive expansion described by Hett et al. (2024), who reported an average annual increase of approximately 0.9 cm³. According to these normative data, PSD volume typically reaches 6–9 cm³ in healthy adults (6). The significance of using a 3D nnU-Net approach is further highlighted by comparisons with older studies that used different methodologies. For instance, Park et al. (2020) reported a mean volume of approximately 4.0 cm 3 in older adults, but they used a semiautomatic threshold-based method that is often susceptible to operator variability and image contrast variations (5). Similarly, Melin et al. (2023) reported a mean volume of 4.18 cm 3 using a manually corrected 2D U-Net, which may not fully capture the spatial continuity of the thin, complex PSD structure along the superior sagittal sinus (1). By modelling the structure's complete three-dimensional topology, the 3D nnU-Net used in this study may provide a more anatomically faithful volumetric representation. Segmentation was performed on 3D T2-FLAIR rather than 3D T2w images for three main reasons. First, 3D T2-FLAIR sequences are now routinely acquired in clinical protocols, enhancing the translational applicability of our method. In contrast, high-resolution 3D T2w images are not commonly acquired for standard clinical settings, limiting their practical use for widespread PSD segmentation. Second, 3D T2-FLAIR provides high contrast between PSD and the adjacent subarachnoid space, facilitating identification and manual segmentation of PSD structures. Third, prior studies establishing the role of PSD in neuroinflammation and CSF drainage have relied on 3D T2-FLAIR imaging to demonstrate meaningful results (1,2,8,26). A limitation of using 3D T2-FLAIR is the inability to clearly distinguish arachnoid granulations (AGs), which are better visualized on 3D T2w images (6). Consequently, PSD segmentation based on 3D T2-FLAIR may include AGs and potentially overestimate PSD volume. However, AGs are infrequent and small in early childhood: neonates and two-year-old children exhibit an average of 0.1 ± 0.3 granulations in the superior sagittal sinus, increasing to 1.2 ± 2.5 by 10 years of age, with 75% measuring approximately 4 mm (27). Therefore, in our pediatric cohort, any overestimation of PSD volume due to AG inclusion is likely negligible. No significant differences in PSD volumes were observed between ASD and TD children. This finding indicates that, within the present cohort, PSD volumes do not distinguish children with ASD from their TD peers. PSD volume showed consistent associations with CSF and ea-CSF, consistent with previous studies that identified the PSD as a potential sink for CSF (1,8,16). Together, these findings suggest that PSD volume may reflect general developmental properties of fluid compartments rather than disorder-specific alterations, although this interpretation requires further investigation. The biological significance of PSD volume in the developing brain is likely distinct from that proposed in healthy adults or in neurodegenerative conditions. The mesenchymal stroma that characterizes the PSD originates from meningeal tissue, which, including the meningeal lymphatic system, continues to mature during the postnatal period (28). Our previous study showed that PSD volume in children with ASD was negatively correlated with the severity of developmental delay, highlighting a possible multifactorial maturational process of these structures in the developing brain (16). The meninges harbor stem and progenitor cell populations involved in neuronal migration and glial development, underscoring the potential functional relevance of the meninges and adjacent PSD in brain maturation (29,30). In the study by Melin et al., PSD in adults is suggested to play a predominantly neuroimmunologic role rather than serving as a primary pathway for CSF drainage (1). However, in the child brain, given that AGs are still underdeveloped, the PSD may play a more important role, contributing to CSF drainage. Although no significant differences between ASD and TD groups were observed in our cohort, the strong association between PSD and CSF volumes supports the functional relevance of this structure in the developing brain. Our findings show that the prefrontal PSD appears relatively smaller compared to the fronto-parietal and occipital regions in children. This regional distribution is broadly consistent with reports in adults, in whom posterior PSD segments (parietal and occipital) are typically the most voluminous (6). However, from a developmental perspective, the relatively reduced frontal PSD observed in childhood may parallel the well-established posterior-to-anterior gradient of cortical maturation (31). While posterior regions mature earlier, the frontal lobes undergo more prolonged volumetric expansion and structural refinement (31). Thus, the regional PSD pattern observed here may reflect these broader maturational gradients of the developing brain rather than a static anatomical distribution. This study has limitations. First, there was an age difference between the ASD and TD groups. Although a dedicated analysis indicated that age did not significantly influence total or regional PSD volumes, residual developmental effects cannot be entirely excluded. Second, MRI acquisition in children with ASD often requires sedation for clinical reasons. While it remains uncertain, sedation may influence CSF distribution and related measures, as previously reported (16). Although this represents a potential confounder, it also reflects real-world clinical imaging conditions in pediatric ASD populations and should be considered when interpreting CSF-related findings. A limitation concerns the ComBAT harmonization approach: as ASD participants were exclusively acquired at IRCCS E. Medea, direct verification that the estimated site effect generalizes equivalently to the ASD population was not possible, and residual site effects cannot be entirely excluded. A further limitation relates to the development of the segmentation model. It was trained exclusively on a pediatric ASD cohort, which may limit its generalizability to other pediatric neurological disorders. At the same time, the stable performance observed on independent TD datasets suggests that ASD-specific anatomical features do not strongly influence the model. The availability of larger datasets including both ASD and non-ASD subjects, could enable mixed training strategies, potentially improving robustness and generalizability. Future work should test the current tool on pediatric cohorts with other neurological disorders and also explore multi-sequence approaches to further enhance flexibility and clinical applicability. Conclusion We developed a publicly available fully automated 3D nnU-Net framework for PSD segmentation in pediatric MRI using standard clinical 3D T2-FLAIR imaging. The model demonstrated high performance across independent cohorts and imaging protocols, supporting its reliability for large-scale studies. PSD volumes did not differ between children with ASD and TD controls and were consistently associated with CSF-related compartments. The proposed framework provides a scalable and reliable tool to investigate PSD and its role in neurofluid dynamics during brain development, and may enable broader applications in studying its contribution across diverse pediatric neurological disorders. Abbreviations AC PC–Anterior commissure–posterior commissure AGs Arachnoid granulations ASD Autism spectrum disorder cGM Cortical gray matter CSF Cerebrospinal fluid DSC Dice similarity coefficient ea CSF–Extra–axial cerebrospinal fluid HD Hausdorff distance ICV Intracranial volume MRI Magnetic resonance imaging nnU Net–Self–configuring deep learning framework based on the U–Net architecture OLS Ordinary least squares PSD Parasagittal dura TD Typically developing TDM Typically developing cohort from IRCCS Eugenio Medea TDZ Typically developing cohort from Zurich Children’s Hospital VS Volumetric similarity Declarations Ethics approval and consent to participate This study was approved by the Institutional Review Board of the IRCCS Eugenio Medea Scientific Institute (Decision No. 926/2022). Written informed consent was obtained from the legal guardians of all participants before inclusion in the study. All procedures were conducted in accordance with the Declaration of Helsinki. Data from typically developing children acquired at Zurich Children’s Hospital were shared under a formal data transfer agreement between Zurich Children’s Hospital and IRCCS Eugenio Medea. Consent for publication Not applicable Competing interests The authors declare that they have no competing interests. Funding This study was supported by the Italian Ministry of Health (Ricerca Corrente 2024, 2025, 2026) and 5×1000 funds for biomedical research. APC funded by Bibliosan. Author Contribution N.A., T.C., F.C., and C.G. contributed to the study conception and design. N.A., F.C., and C.G. performed the manual PSD annotations. T.C. and F.C. contributed to the development and design of the deep learning model and conducted the computational analyses. F.C. drafted a substantial portion of the manuscript. R.T., R.K., F.L., V.M. and E.M. were responsible for data curation. N.A., T.C., F.C.,R.T. participated in data analysis and interpretation. All authors contributed to manuscript revision and approved the final version. Acknowledgement The authors thank the children and their families for participating in this study. The authors further thank Gloria Rizzato for her valuable contribution to the previous work that laid the foundation for the present study. Data Availability The data generated and analysed in this article will be made available upon request. The automated PSD segmentation tool will be made publicly available on GitHub upon acceptance of the manuscript. References 1. Melin E, Ringstad G, Valnes LM. Human parasagittal dura is a potential neuroimmune interface. Commun Biol. 2023;6:260. 2. Absinta M, Ha SK, Nair G, Sati P, Luciano NJ, Palisoc M, et al. Human and nonhuman primate meninges harbor lymphatic vessels that can be visualized noninvasively by MRI. Elife. 2017;6:e29738. 3. Visanji NP, Lang AE, Munoz DG. Lymphatic vasculature in human dural superior sagittal sinus : Implications for neurodegenerative proteinopathies. Neurosci Lett. 2018;665:18–21. 4. Louveau A, Plog BA, Antila S, Alitalo K, Nedergaard M, Kipnis J. Understanding the functions and relationships of the glymphatic system and meningeal lymphatics. J Clin Invest. 2017;127(9):3210–9. 5. Park M, Kim JW, Ahn SJ, Cha YJ, Suh SH. Aging is positively associated with peri-sinus lymphatic space volume: Assessment using 3t black‐blood mri. J Clin Med. 2020;9:3353. 6. Hett K, McKnight CD, Leguizamon M, Lindsey JS, Eisma JJ, Elenberger J, et al. Deep learning segmentation of peri-sinus structures from structural magnetic resonance imaging: validation and normative ranges across the adult lifespan. Fluids Barriers CNS. 2024;21(1):15. 7. Song AK, Hett K, Eisma JJ, McKnight CD, Elenberger J, Stark AJ, et al. Parasagittal dural space hypertrophy and amyloid-β deposition in Alzheimer’s disease. Brain Commun. 2023;5(3):fcad128. 8. Ringstad G, Eide PK. Cerebrospinal fluid tracer efflux to parasagittal dura in humans. Nat Commun. 2020;11(1):354. 9. Hett K, McKnight CD, Eisma JJ, Elenberger J, Lindsey JS, Considine CM, et al. Parasagittal dural space and cerebrospinal fluid (CSF) flow across the lifespan in healthy adults. Fluids Barriers CNS. 2022;19(1):24. 10. Wang L, Wang B, Wu C, Wang J, Sun M. Autism Spectrum Disorder : Neurodevelopmental Risk Factors, Biological Mechanism, and Precision Therapy. Int J Mol Sci. 2023;24(3):1819. 11. Shen MD. Cerebrospinal fluid and the early brain development of autism. J Neurodev Disord. 2018;10(1):39. 12. Shen MD, Nordahl CW, Young GS, Wootton-Gorges SL, Lee A, Liston SE, et al. Early brain enlargement and elevated extra-axial fluid in infants who develop autism spectrum disorder. Brain. 2013;136(9):2825–2835. 13. Garic D, McKinstry RC, Rutsohn J, Slomowitz R, Wolff J, MacIntyre LC, et al. Enlarged Perivascular Spaces in Infancy and Autism Diagnosis, Cerebrospinal Fluid Volume, and Later Sleep Problems. JAMA Netw Open. 2023;6(12):e2348341. 14. Hughes HK, Moreno RJ, Ashwood P. Innate Immune Dysfunction and Neuroinflammation in Autism Spectrum Disorder ( ASD ). Brain Behav Immun. 2023;108:245–254. 15. Goines P, Van de Water J. The Immune System’s Role in the Biology of Autism. Curr Opin Neurol. 2010;23(2):111–7. 16. Agarwal N, Frigerio G, Rizzato G, Ciceri T, Mani E, Lanteri F, et al. Parasagittal dural volume correlates with cerebrospinal fluid volume and developmental delay in children with autism spectrum disorder. Commun Med. 2024;4(1):191. 17. Donahue MJ, Mcknight CD, Claassen DO, Hett K. The parasagittal dural space of the human brain. Brain. 2025;148(10):3481–3495. 18. Lord C, Rutter M, Le Couteur A. Autism Diagnostic Interview-Revised : A Revised Version of a Diagnostic Interview for Caregivers of Individuals with Possible Pervasive Developmental Disorders. J Autism Dev Disord. 1994;24(5):659–685. 19. Lord C, Rutter M, DiLavore PC, Risi S, Gotham K, Bishop SL. Autism Diagnostic Observation Schedule – Second Edition (ADOS-2). Los Angeles, CA: Western Psychological Corporation; 2012. 20. Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. 2010;29(6):1310–1320. 21. Frigerio G, Rizzato G, Peruzzo D, Ciceri T, Mani E, Lanteri F. Perivascular Space Burden in Children With Autism Spectrum Disorder Correlates With Neurodevelopmental Severity. J Magn Reson Imaging. 2025;62(5):1496–506. 22. Shang Z, Kaandorp M, Payette K, Fernandez Garcia M, Licandro R, Langs G, et al. Towards contrast- and pathology-agnostic clinical fetal brain MRI segmentation using SynthSeg. Neuroimage. 2026;327:121729. 23. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203–11. 24. Fortin JP, Cullen N, Sheline YI, Taylor WD, Aselcioglu I, Cook PA, et al. Harmonization of cortical thickness measurements across scanners and sites. Neuroimage. 2018;167(June 2017):104–20. 25. Budd S, Robinson EC, Kainz B. A survey on active learning and human-in-the-loop deep learning for medical image analysis. Med Image Anal. 2021;71:102062. 26. Albayram MS, Smith G, Tufan F, Tuna IS, Bostancıklıoğlu M, Zile M, et al. Non-invasive MR imaging of human brain lymphatic networks with connections to cervical lymph nodes. Nat Commun. 2022;13(1):203. 27. Radoš M, Živko M, Periša A, Orešković D. No Arachnoid Granulations — No Problems : Number, Size, and Distribution of Arachnoid Granulations From Birth to 80 Years of Age. Front Aging Neurosci. 2021;13:698865. 28. Antila S, Karaman S, Nurmi H, Airavaara M, Voutilainen MH, Mathivet T, et al. Development and plasticity of meningeal lymphatic vessels. J Exp Med. 2017;214(12):3645–3667. 29. Decimo I, Dolci S, Panuccio G, Riva M, Fumagalli G, Bifari F. Meninges : A Widespread Niche of Neural Progenitors for the Brain. Neuroscientist. 2021;27(5):506–528. 30. Bifari F, Berton V, Pino A, Kusalo M, Malpeli G, Chio M Di, et al. Meninges harbor cells expressing neural precursor markers during development and adulthood. Front Cell Neurosci. 2015;9:383. 31. Semple BD, Blomgren K, Gimlin K, Ferriero DM, Noble-haeusslein LJ. Brain development in rodents and humans : Identifying benchmarks of maturation and vulnerability to injury across species. Prog Neurobiol. 2013;106–107:1–16. Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterialspsd.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 17 May, 2026 Reviews received at journal 14 May, 2026 Reviewers agreed at journal 07 May, 2026 Reviewers agreed at journal 05 May, 2026 Reviewers invited by journal 05 May, 2026 Editor assigned by journal 30 Apr, 2026 Submission checks completed at journal 30 Apr, 2026 First submitted to journal 28 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9556003","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":637970051,"identity":"f7527cea-f26b-4c72-8aef-3ce7be5d0299","order_by":0,"name":"Francesca Castellotti","email":"","orcid":"","institution":"Scientific Institute IRCCS Eugenio Medea, Bosisio Parini (LC)","correspondingAuthor":false,"prefix":"","firstName":"Francesca","middleName":"","lastName":"Castellotti","suffix":""},{"id":637970052,"identity":"fcf3677c-592f-436b-8705-6fafddff8b48","order_by":1,"name":"Tommaso Ciceri","email":"data:image/png;base64,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","orcid":"","institution":"Scientific Institute IRCCS Eugenio Medea, Bosisio Parini (LC)","correspondingAuthor":true,"prefix":"","firstName":"Tommaso","middleName":"","lastName":"Ciceri","suffix":""},{"id":637970053,"identity":"167c7ba3-959a-4d66-acf6-c764ad8e02f2","order_by":2,"name":"Chiara Girardi","email":"","orcid":"","institution":"Scientific Institute IRCCS Eugenio Medea, Bosisio Parini (LC)","correspondingAuthor":false,"prefix":"","firstName":"Chiara","middleName":"","lastName":"Girardi","suffix":""},{"id":637970054,"identity":"c149444e-bdd9-4d1e-b1d3-851d4b543009","order_by":3,"name":"Elisa Mani","email":"","orcid":"","institution":"Scientific Institute IRCCS Eugenio Medea, Bosisio Parini (LC)","correspondingAuthor":false,"prefix":"","firstName":"Elisa","middleName":"","lastName":"Mani","suffix":""},{"id":637970055,"identity":"88c8e3e7-6479-4740-9407-3eff135752cf","order_by":4,"name":"Fabiola Lanteri","email":"","orcid":"","institution":"Scientific Institute IRCCS Eugenio Medea, Bosisio Parini (LC)","correspondingAuthor":false,"prefix":"","firstName":"Fabiola","middleName":"","lastName":"Lanteri","suffix":""},{"id":637970056,"identity":"92f6bc0b-9b83-4d3f-a35e-5c9278ebc6d5","order_by":5,"name":"Valentina Mariani","email":"","orcid":"","institution":"Scientific Institute IRCCS Eugenio Medea, Bosisio Parini (LC)","correspondingAuthor":false,"prefix":"","firstName":"Valentina","middleName":"","lastName":"Mariani","suffix":""},{"id":637970058,"identity":"be04d653-a7bb-42da-a24c-a2f13d3d6a41","order_by":6,"name":"Ruth O'Gorman Tuura","email":"","orcid":"","institution":"University Children's Hospital Zurich","correspondingAuthor":false,"prefix":"","firstName":"Ruth","middleName":"O'Gorman","lastName":"Tuura","suffix":""},{"id":637970059,"identity":"2d537a1e-7f38-4f35-bf29-75f466f54dee","order_by":7,"name":"Raimund Kottke","email":"","orcid":"","institution":"University Children's Hospital Zurich","correspondingAuthor":false,"prefix":"","firstName":"Raimund","middleName":"","lastName":"Kottke","suffix":""},{"id":637970060,"identity":"f09cf65c-4c59-4aeb-b120-3e34629a7824","order_by":8,"name":"Nivedita Agarwal","email":"","orcid":"","institution":"Scientific Institute IRCCS Eugenio Medea, Bosisio Parini (LC)","correspondingAuthor":false,"prefix":"","firstName":"Nivedita","middleName":"","lastName":"Agarwal","suffix":""}],"badges":[],"createdAt":"2026-04-28 15:23:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9556003/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9556003/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":109222298,"identity":"4c82a512-19db-46dc-8a80-0cdf3bf5599d","added_by":"auto","created_at":"2026-05-13 21:07:02","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":193746,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy workflow for automated PSD segmentation and volumetric assessment.\u003c/strong\u003e 3D T2-FLAIR images acquired at IRCCS E. Medea (ASD and TDM cohorts) and Zurich Children’s Hospital (TDZ cohort) were used as input. After preprocessing, ground-truth PSD masks were generated through manual extension and refinement of prior segmentations. A 3D nnU-Net model was trained and internally validated on the ASD cohort. Model performance was evaluated on an independent ASD test set and further assessed in TD children, including the external Zurich cohort, to determine cross-cohort generalizability. The final automated 3D PSD masks were used to compute total PSD volume and to perform regional subdivision into prefrontal, central (fronto-parietal), and occipital segments. Volumetric measures were subsequently used for comparisons between ASD and TD participants.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-9556003/v1/7bfe9ca4b3201cb26425adcd.png"},{"id":109222269,"identity":"24e61190-dd0d-4154-aa25-0f23e5788dee","added_by":"auto","created_at":"2026-05-13 21:06:44","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":374455,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIllustration of the PSD label annotation.\u003c/strong\u003e The region of interest in the previous work was defined by tracing an arc along the cranial circumference, subtended by a 60° angle defined by the anterior and posterior commissure (AC–PC) landmarks (white). The anterior and posterior PSD portions highlighted in magenta were manually added using FSLeyes to achieve complete structural coverage.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-9556003/v1/c2fc52f1b763ca46b7fdc589.png"},{"id":109249169,"identity":"0995b280-3777-418c-b8f9-6fdf0a51a18a","added_by":"auto","created_at":"2026-05-14 08:43:02","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":472048,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eResult of PSD segmentation\u003c/strong\u003e superimposed on a 3D T2-FLAIR image\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-9556003/v1/5eb2fb8c2a2c1999d9e24a81.png"},{"id":109222301,"identity":"d46f424a-3758-4737-82cd-ef9efa0e94ad","added_by":"auto","created_at":"2026-05-13 21:07:11","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":306113,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e3D visualization of PSD subregions\u003c/strong\u003e: prefrontal (yellow), fronto-parietal (blue), occipital (magenta).\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-9556003/v1/3fc679a4bac2c2f44b1f7b78.png"},{"id":109249148,"identity":"b685723a-4f4a-45ac-acfb-8283f43079c8","added_by":"auto","created_at":"2026-05-14 08:42:52","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":167609,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eHeatmaps of Spearman correlation coefficients\u003c/strong\u003e\u003c/em\u003e\u003cem\u003e illustrating the relationships among PSD measures (total and regional) and cerebral volumes in ASD and TD participants. Statistically significant correlations are in bold.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-9556003/v1/0536459fde4096b7ae2b72d9.png"},{"id":109252166,"identity":"762d3255-4dd4-4111-ad80-8261249ff416","added_by":"auto","created_at":"2026-05-14 09:21:54","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1714154,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9556003/v1/3a2917e0-c1aa-4773-9493-f635eeae3fa1.pdf"},{"id":109216308,"identity":"6a83b9ec-fa1f-4516-9c37-c9558c778118","added_by":"auto","created_at":"2026-05-13 18:04:03","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":354344,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterialspsd.docx","url":"https://assets-eu.researchsquare.com/files/rs-9556003/v1/12d0a2bc3ba3e974f7ffbab7.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Automated 3D nnU-Net Segmentation of the Parasagittal Dura in Children with Autism Spectrum Disorder Using Clinical 3D T2-FLAIR","fulltext":[{"header":"Background","content":"\u003cp\u003eThe parasagittal dura (PSD) represents a specialized mesenchymal stromal tissue containing critical anatomical structures that facilitate the efflux of cerebrospinal fluid (CSF) and solutes from the central nervous system (CNS) (1). Anatomically positioned between the dual layers of the cerebral dura mater and surrounding the superior sagittal sinus, the PSD is home to meningeal lymphatic vessels (2,3). These lymphatic channels are essential for neuro-immune surveillance and contribute to the preservation of CNS homeostasis by mediating the drainage of immune cells and antigens toward peripheral lymph nodes (3,4). PSD is readily identifiable on standard magnetic resonance imaging (MRI) sequences and may represent a useful in vivo target for investigating neurofluid dynamics (2).\u003c/p\u003e \u003cp\u003eVolumetric quantification of the PSD has been performed by employing semi-automated and automated segmentation tools on MR images, primarily in adult populations (5\u0026ndash;9). Empirical data show that PSD volume increases linearly with age in healthy adults, typically reaching a plateau around age 70 (5,9). Furthermore, a recent study applying automatic segmentation on Alzheimer\u0026rsquo;s disease cohorts has revealed a significant positive association between PSD volume and amyloid-β burden as quantified by positron emission tomography (7). The authors suggest that a hypertrophic PSD may constitute a compensatory response to the progressive impairment of lymphatic drainage associated with senescence, potentially serving as an MR surrogate marker for underlying neuroinflammatory processes and altered neurofluid kinetics (7,9).\u003c/p\u003e \u003cp\u003eAutism spectrum disorder (ASD) represents a neurodevelopmental condition with a multifactorial etiology (10). Recent evidence indicates that elevated extra-axial cerebrospinal fluid (ea-CSF) volume may serve as a potential early-stage neuroimaging biomarker in neonates with high risk of developing ASD and young children with a diagnosis of ASD (11,12). Such volumetric expansions are hypothesized to reflect dysregulation of CSF kinetics (12,13). Consequently, impaired drainage, possibly involving the meningeal lymphatic system and PSD, may affect neuroimmunologic processes and contribute to the aberrant neurodevelopmental trajectories characteristic of ASD (14,15).\u003c/p\u003e \u003cp\u003eTo explore this hypothesis, in our previous investigation, a semi-automated 2D U-Net-based tool was employed to segment the PSD from 3D T2-FLAIR sequences and quantify its volume in a pediatric ASD cohort aged 2\u0026ndash;8 years (16). We found a significant positive correlation between PSD, CSF, and ea-CSF volumes, confirming the previously established functional role of the PSD in CSF homeostasis (16). Additionally, the association identified between PSD volume and the degree of developmental delay suggests a role for this structure in the clinical manifestation of ASD (16). However, the use of a 2D architecture may not have fully captured the three-dimensional (3D) spatial continuity of this complex structure. Moreover, for practical purposes, only a central segment of the PSD was quantified, limiting anatomical completeness. Additionally, the segmentation required substantial manual post-processing to correct inaccuracies, introducing potential inter- and intra-rater variability and limiting its widespread usage among the scientific community.\u003c/p\u003e \u003cp\u003eCurrently, the only publicly accessible fully automated 3D U-Net\u0026ndash;based tool for PSD segmentation uses high-resolution 3D T2-weighted imaging (6). However, since the model was trained and validated on a dataset comprising individuals aged 11\u0026ndash;83 years, the same authors reported a tendency to overestimate the PSD volume when applied to younger pediatric cohorts (ages 5\u0026ndash;10) (6). This highlights the need for age-specific segmentation tools that account for the unique neuroanatomical characteristics of the developing brain. Furthermore, reliance on high-resolution T2-weighted sequences may limit the method's broad clinical translational utility, as these acquisitions are not routinely included in standard pediatric neuroimaging protocols.\u003c/p\u003e \u003cp\u003eTo address these methodological limitations and the absence of automated tools for pediatric populations, we developed a fully automated 3D nnU-Net\u0026ndash;based method for whole-PSD segmentation from standard pediatric 3D T2-FLAIR images, trained on a large manually annotated ASD cohort. We then evaluated its generalizability on independent datasets of typically developing children, including an external cohort acquired at Zurich Children\u0026rsquo;s Hospital with different imaging parameters. Using this approach, we quantified and compared total and regional PSD volume in both ASD and typically developing (TD) groups. A regional subdivision was adopted to account for the spatial heterogeneity of the PSD along the superior sagittal sinus, with prior studies showing non-uniform distribution and development across frontal, central, and posterior regions (6,17). This approach may enable a more anatomically grounded and developmentally sensitive characterization of PSD in pediatric population.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cp\u003eThis study aimed to develop, validate, and apply a fully automated deep learning\u0026ndash;based framework for PSD segmentation in pediatric brain MRI. The overall workflow is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Briefly, 3D T2-FLAIR images were pre-processed and manually annotated to generate ground-truth PSD masks. A 3D nnU-Net model was trained on the ASD cohort and evaluated on independent datasets, including TD children and an external TD cohort acquired at Zurich Children\u0026rsquo;s Hospital. The resulting segmentations were subsequently used for regional PSD subdivision and volumetric analyses comparing ASD and TD groups.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eDatasets and preprocessing\u003c/b\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eASD Cohort\u003c/h2\u003e \u003cp\u003eChildren diagnosed with ASD who underwent MRI between July 2021 and July 2025 were included in this study. Inclusion criteria were a confirmed diagnosis of ASD and age between 2 and 12 years. Exclusion criteria included MRI sequences affected by motion artefacts and positive for other pathologies or malformations. A total of 200 participants (mean age 4.62\u0026thinsp;\u0026plusmn;\u0026thinsp;1.6; 33 females) were included in the study.\u003c/p\u003e \u003cp\u003eDiagnosis of ASD was established by a multidisciplinary team at the Child Psychopathology Unit of IRCSS E. Medea (Bosisio Parini, Italy) according to DSM-V criteria (American Psychiatric Association, 2013). Diagnostic assessment included the Autism Diagnostic Interview\u0026ndash;Revised (ADI-R) (18), administered to parents, and the Autism Diagnostic Observation Schedule\u0026ndash;2nd edition (ADOS-2) (19) conducted with the child.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eTD Cohort\u003c/h3\u003e\n\u003cp\u003eThe TD subjects were recruited from two institutes: IRCCS Eugenio Medea (TDM) and Zurich Children\u0026rsquo;s Hospital (TDZ).\u003c/p\u003e \u003cp\u003eThe initial TDM dataset included 35 subjects scanned between March 2022 and December 2025. Inclusion criteria were the absence of a diagnosis of ASD or developmental delay and age between 2 and 12 years. Exclusion criteria included motion-related artefacts affecting 3D T1-weighted and/or 3D T2-FLAIR images. The final TDM sample comprised 27 children (mean age 7.01\u0026thinsp;\u0026plusmn;\u0026thinsp;2.8 years, 11 females).\u003c/p\u003e \u003cp\u003eThe initial TDZ dataset comprised 66 subjects. Inclusion criteria were the absence of ASD or developmental delay and age between 2 and 12 years. Exclusion criteria included motion artefacts affecting 3D T1-weighted and/or 3D T2-FLAIR images, absence of 3D T2-FLAIR sequences, and presence of neurological disorders or structural brain abnormalities. The final TDZ sample consisted of 33 children (mean age 6.60\u0026thinsp;\u0026plusmn;\u0026thinsp;1.73 years, 19 females).\u003c/p\u003e\n\u003ch3\u003eData preprocessing\u003c/h3\u003e\n\u003cp\u003eAll ASD and TDM participants were scanned on a 3T MRI system (Achieva dStream; Philips Medical Systems) using a 32-channel head coil at the Diagnostic Imaging and Neuroradiology Unit of our institute.\u003c/p\u003e \u003cp\u003eThe MRI protocol for ASD and TDM participants included two sequences: (a) a 3D T2-FLAIR sequence with the following acquisition parameters: repetition time (TR)\u0026thinsp;=\u0026thinsp;4800\u0026ndash;6000 ms, echo time (TE)\u0026thinsp;=\u0026thinsp;289\u0026ndash;344 ms, inversion time (TI)\u0026thinsp;=\u0026thinsp;1650\u0026ndash;1860 ms, and acquisition matrix\u0026thinsp;=\u0026thinsp;339 \u0026times; 480 \u0026times; 480; and (b) a 3D T1-weighted MPRAGE sequence with TR\u0026thinsp;=\u0026thinsp;7\u0026ndash;8.5 ms, TE\u0026thinsp;=\u0026thinsp;3.2\u0026ndash;3.96 ms, acquisition matrix\u0026thinsp;=\u0026thinsp;256 \u0026times; 256, and an isotropic voxel size of 1 \u0026times; 1 \u0026times; 1 mm\u0026sup3;.\u003c/p\u003e \u003cp\u003eThe TDZ cohort was acquired using a 3D T2-FLAIR sequence with the following parameters: TR\u0026thinsp;=\u0026thinsp;7000 ms, TE\u0026thinsp;=\u0026thinsp;130 ms, TI\u0026thinsp;=\u0026thinsp;1910\u0026ndash;1960 ms, and acquisition matrix\u0026thinsp;=\u0026thinsp;260 \u0026times; 248 \u0026times; 248. The 3D T1-weighted sequence was a SPGR BRAVO acquired sagittally with TR\u0026thinsp;=\u0026thinsp;10 ms, TE\u0026thinsp;=\u0026thinsp;4 ms, TI\u0026thinsp;=\u0026thinsp;450 ms, acquisition matrix\u0026thinsp;=\u0026thinsp;352 \u0026times; 256, and slice thickness\u0026thinsp;=\u0026thinsp;1 mm.\u003c/p\u003e \u003cp\u003e3D T2-FLAIR images were preprocessed using N4 bias field correction to reduce intensity non-uniformities (20). T1-weighted images were preprocessed using FreeSurfer\u0026rsquo;s automated recon-all pipeline, yielding regional brain volume estimates. CSF and ea-CSF masks were obtained using an in-house pipeline described previously (21).\u003c/p\u003e\n\u003ch3\u003eAutomated PSD segmentation\u003c/h3\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eAnnotation protocol of PSD\u003c/h2\u003e \u003cp\u003eGround-truth PSD segmentations on ASD, TDM and TDZ were generated using a human-in-the-loop approach, similar to strategies previously described in neuroimaging segmentation study (22). Manual delineation was performed on 3D T2-FLAIR images. Initial segmentations were obtained from our previously developed 2D U-Net trained to segment the central portion of the PSD (16). These segmentations were systematically reviewed and refined under the supervision of a senior neuroradiologist (N.A.), ensuring anatomical accuracy and consistency across the dataset. Additional trained operators contributed to the refinement process following a standardized protocol, using FSLeyes to correct potential inaccuracies and extend the labels anteriorly and posteriorly to cover the full anatomical extent of the PSD. In earlier work, the PSD segmentation was limited to a central segment defined by a 60\u0026deg; arc along the cranial circumference, passing through the anterior and posterior commissures (16). In the present study, this constraint was removed, resulting in the full-length PSD segmentations (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eNetwork architecture and training strategy\u003c/h2\u003e \u003cp\u003eAutomated PSD segmentation was performed using the nnU-Net framework (23), a self-configuring deep learning approach based on a 3D U-Net architecture that automatically adapts network design, pre-processing, and training strategies to the input data. The network has demonstrated competitive performance across a wide range of medical image segmentation tasks (23).\u003c/p\u003e \u003cp\u003eGiven that children with ASD represented the largest manually annotated cohort and constituted the primary focus of the present study, the model was trained on the ASD dataset. The ASD dataset with manual segmentations (n\u0026thinsp;=\u0026thinsp;200) was split into a training set (n\u0026thinsp;=\u0026thinsp;144,70%) and an independent test set (n\u0026thinsp;=\u0026thinsp;56, 30%). Model development followed the standard nnU-Net pipeline, including five-fold cross-validation on the training set. The final predictions were obtained using the five-fold cross-validation ensemble, by averaging the outputs of the five trained models. The final trained model was evaluated exclusively on the independent test set, which was not used during training or validation.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eEvaluation of the segmentation network\u003c/h3\u003e\n\u003cp\u003eTo assess whether the model generalized beyond the ASD training population and captured stable anatomical features of the PSD, the trained nnU-Net was applied to two independent cohorts of TD children (TDM and TDZ), which were not included in training. Notably, the TDZ cohort was acquired using different 3D T2-FLAIR imaging parameters, allowing assessment of the model\u0026rsquo;s performance across clinical imaging protocols.\u003c/p\u003e \u003cp\u003eModel performance was evaluated on the independent ASD test set and on the TD cohorts using the Dice similarity coefficient (DSC), volumetric similarity coefficient (VS), and Hausdorff distance (HD). Segmentation performance was also evaluated separately for three anatomical regions \u0026mdash; prefrontal, fronto-parietal and occipital, defined by subdividing the PSD post hoc along the cranial circumference, using a 60\u0026deg; arc defined by the anterior and posterior commissures (AC\u0026ndash;PC) as the central reference, as described previously (16). Qualitative evaluation across cohorts was performed to assess whether the network captured the full extent of PSD anatomy. The training was conducted in parallel for each fold on two 48GB NVIDIA L40S.\u003c/p\u003e\n\u003ch3\u003ePSD assessment\u003c/h3\u003e\n\u003cp\u003eTotal and regional PSD volumes (prefrontal, fronto-parietal, and occipital), defined using the regional subdivision described above, were extracted for each subject to perform volumetric analysis.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eStatistical analyses were performed in Python (version 3.13.2) using the pandas, NumPy, SciPy, statsmodels, and neuroCombat libraries. All tests were two-tailed, and p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant. Analyses were conducted on the independent ASD test set and both TD cohorts (TDM and TDZ). To account for inter-site variability in volumetric estimates related to differences in MRI acquisition parameters between TDM and TDZ, ComBAT harmonization was applied to all volumetric measures (PSD, CSF, ea-CSF, and ICV) using the neuroCombat implementation (24). Age, sex, and diagnostic group were included as biological covariates to preserve during harmonization. Following harmonization, the two TD cohorts were merged into a single TD group (N\u0026thinsp;=\u0026thinsp;60) for all subsequent statistical analyses.\u003c/p\u003e \u003cp\u003eWelch\u0026rsquo;s t-tests were used to assess group differences between ASD and TD participants in age and cerebral volumes. Chi-square test was used to compare sex distribution between groups.\u003c/p\u003e \u003cp\u003eExploratory correlation analyses were conducted using Spearman\u0026rsquo;s rank correlation coefficient to examine the relations between age, PSD, and cerebral volumes, as well as between total and regional PSD volumes and cerebral volumes.\u003c/p\u003e \u003cp\u003eTo assess group differences in PSD measures while accounting for potential confounders, ordinary least squares (OLS) linear regression models were fitted with diagnostic group (ASD vs TD) as the main predictor, adjusting for age, sex, and CSF volume according to the following model:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:PSD\\text{}\\text{}={\\beta\\:}_{0}\\text{}+{\\beta\\:}_{g}\\text{}\\cdot\\:\\text{G}\\text{r}\\text{o}\\text{u}\\text{p}+{\\beta\\:}_{a}\\text{}\\cdot\\:\\text{A}\\text{g}\\text{e}+{\\beta\\:}_{s}\\text{}\\cdot\\:\\text{S}\\text{e}\\text{x}+{\\beta\\:}_{c}\\text{}\\text{}\\cdot\\:\\text{C}\\text{S}\\text{F}+\\epsilon\\text{}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere PSD is the dependent variable, β\u003csub\u003e0​\u003c/sub\u003e is the intercept, β\u003csub\u003ei\u003c/sub\u003e ​are the regression coefficients associated with each predictor, \u003cem\u003eGroup\u003c/em\u003e denotes diagnostic group (ASD vs TD), \u003cem\u003eAge\u003c/em\u003e is age, \u003cem\u003eSex\u003c/em\u003e is biological sex, \u003cem\u003eCSF\u003c/em\u003e is cerebrospinal fluid volume, and ϵ is the residual error.\u003c/p\u003e \u003cp\u003eBecause the TD group was not age-matched with the ASD group, an additional sensitivity analysis was performed to evaluate the potential confounding effect of age. Specifically, the regression analysis was repeated on a restricted subsample including only participants aged 2 to 9 years, using the same covariates as in the main models.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eNetwork Segmentation Performance\u003c/h2\u003e \u003cp\u003eIn the ASD cohort, used for model development, the network achieved high segmentation accuracy, with a DSC of 0.93\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02, a VS of 0.98\u0026thinsp;\u0026plusmn;\u0026thinsp;0.01, and a HD of 12.5\u0026thinsp;\u0026plusmn;\u0026thinsp;9.0 (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). To evaluate model generalizability, the trained network was subsequently applied to two independent cohorts ofTD children. In the TDM cohort, the model achieved a DSC of 0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02, a VS of 0.98\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03, and a HD of 15.0\u0026thinsp;\u0026plusmn;\u0026thinsp;5.9 (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). In the TDZ cohort, acquired using different imaging parameters, segmentation performance remained robust, with a DSC of 0.85\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04, a VS of 0.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04, and a HD of 18.4\u0026thinsp;\u0026plusmn;\u0026thinsp;8.2 (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). A region-wise analysis was further performed to assess segmentation performance across the prefrontal, fronto-parietal, and occipital portions of the PSD (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e,\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The model achieved consistently high DSC and VS across regions, with slightly higher segmentation accuracy observed in the central segment. In the ASD test set, DSC ranged from 0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05 in the prefrontal region to 0.94\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02 in the fronto-parietal region, with intermediate values in the occipital region (0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03). Similar patterns were observed in the TD cohorts, with slightly higher segmentation accuracy in the fronto-parietal segment.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eSegmentation performance of the proposed nnU-Net model on the ASD dataset.\u003c/b\u003e Metrics are reported for both the Total PSD and its regional subdivisions (prefrontal, fronto-parietal, and occipital). Results are expressed as Dice\u0026ndash;S\u0026oslash;rensen coefficient (DSC), volumetric similarity (VS), and Hausdorff distance (HD), reported as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDSC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eVS\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHD (mm)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.93\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.98\u0026thinsp;\u0026plusmn;\u0026thinsp;0.01\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e12.5\u0026thinsp;\u0026plusmn;\u0026thinsp;9.0\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePrefrontal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.95\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e9.14\u0026thinsp;\u0026plusmn;\u0026thinsp;9.79\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFronto-Parietal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.94\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.99\u0026thinsp;\u0026plusmn;\u0026thinsp;0.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e6.36\u0026thinsp;\u0026plusmn;\u0026thinsp;1.76\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOccipital\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.97\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e9.78\u0026thinsp;\u0026plusmn;\u0026thinsp;5.08\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eSegmentation performance of the proposed nnU-Net model on the TDM and TDZ datasets\u003c/b\u003e. Metrics are reported for both the Total PSD and its regional subdivisions (prefrontal, fronto-parietal, and occipital). Results are expressed as Dice\u0026ndash;S\u0026oslash;rensen coefficient (DSC), volumetric similarity (VS), and Hausdorff distance (HD), reported as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTDM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTDZ\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eDSC\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eVS\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eHD (mm)\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003eDSC\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cem\u003eVS\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cem\u003eHD (mm)\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.98\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e15.0\u0026thinsp;\u0026plusmn;\u0026thinsp;5.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.85\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e18.4\u0026thinsp;\u0026plusmn;\u0026thinsp;8.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePrefrontal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.88\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.94\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e7.9\u0026thinsp;\u0026plusmn;\u0026thinsp;3.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.87\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e14.2\u0026thinsp;\u0026plusmn;\u0026thinsp;11.0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFronto-Parietal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.98\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e7.8\u0026thinsp;\u0026plusmn;\u0026thinsp;2.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.86\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e8.7\u0026thinsp;\u0026plusmn;\u0026thinsp;3.7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOccipital\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.87\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.93\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e12.6\u0026thinsp;\u0026plusmn;\u0026thinsp;6.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.89\u0026thinsp;\u0026plusmn;\u0026thinsp;0.17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e18.1\u0026thinsp;\u0026plusmn;\u0026thinsp;16.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eQualitative inspection across cohorts confirmed that the network accurately captured the full anatomical extent of the PSD despite the considerable inter-subject variability characteristic of this structure. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e represent PSD segmentation for the entire PSD and its regional subdivisions respectively.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003ePSD assessment\u003c/h2\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003eParticipant characteristics\u003c/h2\u003e \u003cp\u003eThe statistical analyses included 56 children with ASD and 60 TD children (27 from IRCCS E. Medea and 33 from Zurich Children\u0026rsquo;s Hospital), following ComBAT harmonization of volumetric measures across sites. Demographic characteristics and global brain volumetric measures are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, including age, sex distribution, ICV, CSF, and ea-CSF. The groups were not age-matched, with TD participants being older on average (p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001) and were not sex-matched (p\u0026thinsp;=\u0026thinsp;0.001). No significant between-group differences were observed in global brain volumetric measures (all p\u0026thinsp;\u0026gt;\u0026thinsp;0.05). Total and regional PSD volumes (prefrontal, fronto-parietal, and occipital) for the ASD and TD groups are reported in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eDemographic characteristics and global brain volumetric measures of the ASD and TD groups\u003c/b\u003e. Values are reported as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eAge (years)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eASD (N\u0026thinsp;=\u0026thinsp;56)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTD (N\u0026thinsp;=\u0026thinsp;60)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5.1\u0026thinsp;\u0026plusmn;\u0026thinsp;1.5\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6.8\u0026thinsp;\u0026plusmn;\u0026thinsp;2.3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.0001\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSex (M/F)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e45 / 11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e30 / 30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eeaCSF (cm\u0026sup3;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e66.5\u0026thinsp;\u0026plusmn;\u0026thinsp;17.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e67.0\u0026thinsp;\u0026plusmn;\u0026thinsp;21.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.905\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCSF (cm\u0026sup3;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e128.4\u0026thinsp;\u0026plusmn;\u0026thinsp;25.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e128.9\u0026thinsp;\u0026plusmn;\u0026thinsp;35.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.927\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eICV (cm\u0026sup3;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1378.3\u0026thinsp;\u0026plusmn;\u0026thinsp;150.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1353.9\u0026thinsp;\u0026plusmn;\u0026thinsp;133.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.358\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eTotal and regional (frontal, frontoparietal, occipital) PSD volumes\u003c/b\u003e in the ASD and TD groups. Values are expressed as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003ePSD Total (cm\u0026sup3;)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eASD (N\u0026thinsp;=\u0026thinsp;56)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTD (N\u0026thinsp;=\u0026thinsp;60)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.9\u0026thinsp;\u0026plusmn;\u0026thinsp;1.6\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.8\u0026thinsp;\u0026plusmn;\u0026thinsp;1.2\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePSD Frontal (cm\u0026sup3;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.4\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.5\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePSD Fronto-Parietal (cm\u0026sup3;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e3.3\u0026thinsp;\u0026plusmn;\u0026thinsp;1.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e3.3\u0026thinsp;\u0026plusmn;\u0026thinsp;0.9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePSD Occipital (cm\u0026sup3;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e1.1\u0026thinsp;\u0026plusmn;\u0026thinsp;0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e1.0\u0026thinsp;\u0026plusmn;\u0026thinsp;0.5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eAge-related effects on brain volumes\u003c/h2\u003e \u003cp\u003eIn the ASD group, significant positive associations with age were observed for ICV (Spearman\u0026rsquo;s ρ\u0026thinsp;=\u0026thinsp;0.39, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.003). No significant correlations with age were found for CSF, ea-CSF, or PSD volumes (all \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026gt;\u0026thinsp;0.05). In the TD group, no significant age-related effects were observed in ICV, CSF, ea-CSF, or PSD volumes (all \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026gt;\u0026thinsp;0.05) (Tables in supplementary materials S1, S2). The absence of significant age-related effects on PSD volumes in both groups is further illustrated in Supplementary material (Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e), which shows scatter plots of total and regional PSD volumes against age, with fitted quadratic curves remaining essentially flat across all regions in both ASD and TD children.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eAssociations between PSD and brain volumes\u003c/h2\u003e \u003cp\u003eTotal PSD volumes showed significant positive correlations with CSF, ea-CSF, and ICV volumes in both ASD (p\u0026thinsp;=\u0026thinsp;0.005, p\u0026thinsp;=\u0026thinsp;0.002, and 0.03, respectively) and TD (p\u0026thinsp;=\u0026thinsp;0.001, p\u0026thinsp;=\u0026thinsp;0.004, and p\u0026thinsp;=\u0026thinsp;0.03, respectively). Regional analyses demonstrated spatially heterogeneous patterns. In the ASD group, central and occipital PSD volumes exhibited the strongest positive correlations with CSF (p\u0026thinsp;=\u0026thinsp;0.004, p\u0026thinsp;=\u0026thinsp;0.01) and ea-CSF (p\u0026thinsp;=\u0026thinsp;0.001, p\u0026thinsp;=\u0026thinsp;0.004). In contrast, frontal PSD volume showed no significant associations with any brain volume measures. In the TD group, only the central PSD volume was significantly correlated with CSF (p\u0026thinsp;=\u0026thinsp;0.002) and ea-CSF (p\u0026thinsp;=\u0026thinsp;0.01). Notably, central PSD volume was also positively associated with ICV in both cohorts (p\u0026thinsp;=\u0026thinsp;0.02 (ASD), p\u0026thinsp;=\u0026thinsp;0.02 (TD)). These patterns are summarized in the correlation heat maps in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, which provide a visual overview of the relationships among PSD measures, cerebral volumes, and age in the ASD and TD groups.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eLinear Model\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe linear regression analysis performed on total PSD volume showed no significant differences between the ASD and TD groups (β = -0.002, p\u0026thinsp;=\u0026thinsp;0.9). Similarly, sex and age were not significantly associated with PSD volume. In contrast, CSF volume showed a significant positive association with total PSD (β\u0026thinsp;=\u0026thinsp;0.016, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001), indicating that higher CSF volumes were associated with larger PSD volumes. No significant group differences were observed in regional PSD volumes. The age-restricted sensitivity analysis (2\u0026ndash;9 years) confirmed the main findings, showing no significant difference in PSD between ASD and TD groups (β = \u0026minus;0.061, p\u0026thinsp;=\u0026thinsp;0.8). CSF volume continued to remain significantly associated with PSD (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Detailed results are provided in Supplementary tables (Table S3-S7).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we designed and validated a fully automated 3D nnU-Net model to segment PSD in pediatric brain. This new model expands the analysis to cover the entire anatomical extent of the structure, whereas our previous approach was limited to a central portion of the PSD (16). The proposed model demonstrated reliable performance in both children with ASD and TD cohorts, including an independent external cohort acquired using different 3D T2-FLAIR imaging parameters. This high performance might be partially attributed to the human-in-the-loop framework used during ground-truth generation, in which an experienced neuroradiologist refined the initial U-Net predictions (22). While this expert-led framework introduces a certain methodological bias, literature indicates that such bias can be helpful when training models with limited data, potentially contributing to the model's reliability (25).\u003c/p\u003e \u003cp\u003eThe mean PSD volume observed in our pediatric cohort provides an initial reference for understanding the developmental trajectory of this structure. These values are consistent with previous lifespan studies reporting that the PSD space accounts for approximately 0.30% of total ICV during adolescence (6). The volumes measured in our cohort likely represent an early stage of the progressive expansion described by Hett et al. (2024), who reported an average annual increase of approximately 0.9 cm\u0026sup3;. According to these normative data, PSD volume typically reaches 6\u0026ndash;9 cm\u0026sup3; in healthy adults (6). The significance of using a 3D nnU-Net approach is further highlighted by comparisons with older studies that used different methodologies. For instance, Park et al. (2020) reported a mean volume of approximately 4.0 cm\u003csup\u003e3\u003c/sup\u003e in older adults, but they used a semiautomatic threshold-based method that is often susceptible to operator variability and image contrast variations (5). Similarly, Melin et al. (2023) reported a mean volume of 4.18 cm\u003csup\u003e3\u003c/sup\u003e using a manually corrected 2D U-Net, which may not fully capture the spatial continuity of the thin, complex PSD structure along the superior sagittal sinus (1). By modelling the structure's complete three-dimensional topology, the 3D nnU-Net used in this study may provide a more anatomically faithful volumetric representation.\u003c/p\u003e \u003cp\u003eSegmentation was performed on 3D T2-FLAIR rather than 3D T2w images for three main reasons. First, 3D T2-FLAIR sequences are now routinely acquired in clinical protocols, enhancing the translational applicability of our method. In contrast, high-resolution 3D T2w images are not commonly acquired for standard clinical settings, limiting their practical use for widespread PSD segmentation. Second, 3D T2-FLAIR provides high contrast between PSD and the adjacent subarachnoid space, facilitating identification and manual segmentation of PSD structures. Third, prior studies establishing the role of PSD in neuroinflammation and CSF drainage have relied on 3D T2-FLAIR imaging to demonstrate meaningful results (1,2,8,26). A limitation of using 3D T2-FLAIR is the inability to clearly distinguish arachnoid granulations (AGs), which are better visualized on 3D T2w images (6). Consequently, PSD segmentation based on 3D T2-FLAIR may include AGs and potentially overestimate PSD volume. However, AGs are infrequent and small in early childhood: neonates and two-year-old children exhibit an average of 0.1\u0026thinsp;\u0026plusmn;\u0026thinsp;0.3 granulations in the superior sagittal sinus, increasing to 1.2\u0026thinsp;\u0026plusmn;\u0026thinsp;2.5 by 10 years of age, with 75% measuring approximately 4 mm (27). Therefore, in our pediatric cohort, any overestimation of PSD volume due to AG inclusion is likely negligible.\u003c/p\u003e \u003cp\u003eNo significant differences in PSD volumes were observed between ASD and TD children. This finding indicates that, within the present cohort, PSD volumes do not distinguish children with ASD from their TD peers. PSD volume showed consistent associations with CSF and ea-CSF, consistent with previous studies that identified the PSD as a potential sink for CSF (1,8,16). Together, these findings suggest that PSD volume may reflect general developmental properties of fluid compartments rather than disorder-specific alterations, although this interpretation requires further investigation. The biological significance of PSD volume in the developing brain is likely distinct from that proposed in healthy adults or in neurodegenerative conditions. The mesenchymal stroma that characterizes the PSD originates from meningeal tissue, which, including the meningeal lymphatic system, continues to mature during the postnatal period (28). Our previous study showed that PSD volume in children with ASD was negatively correlated with the severity of developmental delay, highlighting a possible multifactorial maturational process of these structures in the developing brain (16). The meninges harbor stem and progenitor cell populations involved in neuronal migration and glial development, underscoring the potential functional relevance of the meninges and adjacent PSD in brain maturation (29,30).\u003c/p\u003e \u003cp\u003eIn the study by Melin et al., PSD in adults is suggested to play a predominantly neuroimmunologic role rather than serving as a primary pathway for CSF drainage (1). However, in the child brain, given that AGs are still underdeveloped, the PSD may play a more important role, contributing to CSF drainage. Although no significant differences between ASD and TD groups were observed in our cohort, the strong association between PSD and CSF volumes supports the functional relevance of this structure in the developing brain.\u003c/p\u003e \u003cp\u003eOur findings show that the prefrontal PSD appears relatively smaller compared to the fronto-parietal and occipital regions in children. This regional distribution is broadly consistent with reports in adults, in whom posterior PSD segments (parietal and occipital) are typically the most voluminous (6). However, from a developmental perspective, the relatively reduced frontal PSD observed in childhood may parallel the well-established posterior-to-anterior gradient of cortical maturation (31). While posterior regions mature earlier, the frontal lobes undergo more prolonged volumetric expansion and structural refinement (31). Thus, the regional PSD pattern observed here may reflect these broader maturational gradients of the developing brain rather than a static anatomical distribution.\u003c/p\u003e \u003cp\u003eThis study has limitations. First, there was an age difference between the ASD and TD groups. Although a dedicated analysis indicated that age did not significantly influence total or regional PSD volumes, residual developmental effects cannot be entirely excluded. Second, MRI acquisition in children with ASD often requires sedation for clinical reasons. While it remains uncertain, sedation may influence CSF distribution and related measures, as previously reported (16). Although this represents a potential confounder, it also reflects real-world clinical imaging conditions in pediatric ASD populations and should be considered when interpreting CSF-related findings. A limitation concerns the ComBAT harmonization approach: as ASD participants were exclusively acquired at IRCCS E. Medea, direct verification that the estimated site effect generalizes equivalently to the ASD population was not possible, and residual site effects cannot be entirely excluded. A further limitation relates to the development of the segmentation model. It was trained exclusively on a pediatric ASD cohort, which may limit its generalizability to other pediatric neurological disorders. At the same time, the stable performance observed on independent TD datasets suggests that ASD-specific anatomical features do not strongly influence the model. The availability of larger datasets including both ASD and non-ASD subjects, could enable mixed training strategies, potentially improving robustness and generalizability. Future work should test the current tool on pediatric cohorts with other neurological disorders and also explore multi-sequence approaches to further enhance flexibility and clinical applicability.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eWe developed a publicly available fully automated 3D nnU-Net framework for PSD segmentation in pediatric MRI using standard clinical 3D T2-FLAIR imaging. The model demonstrated high performance across independent cohorts and imaging protocols, supporting its reliability for large-scale studies. PSD volumes did not differ between children with ASD and TD controls and were consistently associated with CSF-related compartments. The proposed framework provides a scalable and reliable tool to investigate PSD and its role in neurofluid dynamics during brain development, and may enable broader applications in studying its contribution across diverse pediatric neurological disorders.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ePC\u0026ndash;Anterior commissure\u0026ndash;posterior commissure\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAGs\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eArachnoid granulations\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eASD\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAutism spectrum disorder\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ecGM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCortical gray matter\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCSF\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCerebrospinal fluid\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDSC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDice similarity coefficient\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eea\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCSF\u0026ndash;Extra\u0026ndash;axial cerebrospinal fluid\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eHD\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eHausdorff distance\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eICV\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIntracranial volume\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMRI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMagnetic resonance imaging\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ennU\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eNet\u0026ndash;Self\u0026ndash;configuring deep learning framework based on the U\u0026ndash;Net architecture\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eOLS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eOrdinary least squares\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ePSD\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eParasagittal dura\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTD\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eTypically developing\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTDM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eTypically developing cohort from IRCCS Eugenio Medea\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTDZ\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eTypically developing cohort from Zurich Children\u0026rsquo;s Hospital\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eVS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eVolumetric similarity\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e \u003cp\u003eThis study was approved by the Institutional Review Board of the IRCCS Eugenio Medea Scientific Institute (Decision No. 926/2022). Written informed consent was obtained from the legal guardians of all participants before inclusion in the study. All procedures were conducted in accordance with the Declaration of Helsinki. Data from typically developing children acquired at Zurich Children\u0026rsquo;s Hospital were shared under a formal data transfer agreement between Zurich Children\u0026rsquo;s Hospital and IRCCS Eugenio Medea.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for publication\u003c/strong\u003e \u003cp\u003eNot applicable\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThis study was supported by the Italian Ministry of Health (Ricerca Corrente 2024, 2025, 2026) and 5\u0026times;1000 funds for biomedical research. APC funded by Bibliosan.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eN.A., T.C., F.C., and C.G. contributed to the study conception and design. N.A., F.C., and C.G. performed the manual PSD annotations. T.C. and F.C. contributed to the development and design of the deep learning model and conducted the computational analyses. F.C. drafted a substantial portion of the manuscript. R.T., R.K., F.L., V.M. and E.M. were responsible for data curation. N.A., T.C., F.C.,R.T. participated in data analysis and interpretation. All authors contributed to manuscript revision and approved the final version.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe authors thank the children and their families for participating in this study. The authors further thank Gloria Rizzato for her valuable contribution to the previous work that laid the foundation for the present study.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data generated and analysed in this article will be made available upon request. The automated PSD segmentation tool will be made publicly available on GitHub upon acceptance of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e1. Melin E, Ringstad G, Valnes LM. Human parasagittal dura is a potential neuroimmune interface. Commun Biol. 2023;6:260.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e2. Absinta M, Ha SK, Nair G, Sati P, Luciano NJ, Palisoc M, et al. Human and nonhuman primate meninges harbor lymphatic vessels that can be visualized noninvasively by MRI. Elife. 2017;6:e29738.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e3. Visanji NP, Lang AE, Munoz DG. Lymphatic vasculature in human dural superior sagittal sinus : Implications for neurodegenerative proteinopathies. Neurosci Lett. 2018;665:18\u0026ndash;21.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e4. Louveau A, Plog BA, Antila S, Alitalo K, Nedergaard M, Kipnis J. Understanding the functions and relationships of the glymphatic system and meningeal lymphatics. J Clin Invest. 2017;127(9):3210\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e5. Park M, Kim JW, Ahn SJ, Cha YJ, Suh SH. Aging is positively associated with peri-sinus lymphatic space volume: Assessment using 3t black‐blood mri. J Clin Med. 2020;9:3353.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e6. Hett K, McKnight CD, Leguizamon M, Lindsey JS, Eisma JJ, Elenberger J, et al. Deep learning segmentation of peri-sinus structures from structural magnetic resonance imaging: validation and normative ranges across the adult lifespan. Fluids Barriers CNS. 2024;21(1):15.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e7. Song AK, Hett K, Eisma JJ, McKnight CD, Elenberger J, Stark AJ, et al. Parasagittal dural space hypertrophy and amyloid-β deposition in Alzheimer\u0026rsquo;s disease. Brain Commun. 2023;5(3):fcad128.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e8. Ringstad G, Eide PK. Cerebrospinal fluid tracer efflux to parasagittal dura in humans. Nat Commun. 2020;11(1):354.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e9. Hett K, McKnight CD, Eisma JJ, Elenberger J, Lindsey JS, Considine CM, et al. Parasagittal dural space and cerebrospinal fluid (CSF) flow across the lifespan in healthy adults. Fluids Barriers CNS. 2022;19(1):24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e10. Wang L, Wang B, Wu C, Wang J, Sun M. Autism Spectrum Disorder : Neurodevelopmental Risk Factors, Biological Mechanism, and Precision Therapy. Int J Mol Sci. 2023;24(3):1819.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e11. Shen MD. Cerebrospinal fluid and the early brain development of autism. J Neurodev Disord. 2018;10(1):39.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e12. Shen MD, Nordahl CW, Young GS, Wootton-Gorges SL, Lee A, Liston SE, et al. Early brain enlargement and elevated extra-axial fluid in infants who develop autism spectrum disorder. Brain. 2013;136(9):2825\u0026ndash;2835.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e13. Garic D, McKinstry RC, Rutsohn J, Slomowitz R, Wolff J, MacIntyre LC, et al. Enlarged Perivascular Spaces in Infancy and Autism Diagnosis, Cerebrospinal Fluid Volume, and Later Sleep Problems. JAMA Netw Open. 2023;6(12):e2348341.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e14. Hughes HK, Moreno RJ, Ashwood P. Innate Immune Dysfunction and Neuroinflammation in Autism Spectrum Disorder ( ASD ). Brain Behav Immun. 2023;108:245\u0026ndash;254.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e15. Goines P, Van de Water J. The Immune System\u0026rsquo;s Role in the Biology of Autism. Curr Opin Neurol. 2010;23(2):111\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e16. Agarwal N, Frigerio G, Rizzato G, Ciceri T, Mani E, Lanteri F, et al. Parasagittal dural volume correlates with cerebrospinal fluid volume and developmental delay in children with autism spectrum disorder. Commun Med. 2024;4(1):191.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e17. Donahue MJ, Mcknight CD, Claassen DO, Hett K. The parasagittal dural space of the human brain. Brain. 2025;148(10):3481\u0026ndash;3495.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e18. Lord C, Rutter M, Le Couteur A. Autism Diagnostic Interview-Revised : A Revised Version of a Diagnostic Interview for Caregivers of Individuals with Possible Pervasive Developmental Disorders. J Autism Dev Disord. 1994;24(5):659\u0026ndash;685.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e19. Lord C, Rutter M, DiLavore PC, Risi S, Gotham K, Bishop SL. Autism Diagnostic Observation Schedule \u0026ndash; Second Edition (ADOS-2). Los Angeles, CA: Western Psychological Corporation; 2012.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e20. Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. 2010;29(6):1310\u0026ndash;1320.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e21. Frigerio G, Rizzato G, Peruzzo D, Ciceri T, Mani E, Lanteri F. Perivascular Space Burden in Children With Autism Spectrum Disorder Correlates With Neurodevelopmental Severity. J Magn Reson Imaging. 2025;62(5):1496\u0026ndash;506.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e22. Shang Z, Kaandorp M, Payette K, Fernandez Garcia M, Licandro R, Langs G, et al. Towards contrast- and pathology-agnostic clinical fetal brain MRI segmentation using SynthSeg. Neuroimage. 2026;327:121729.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e23. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e24. Fortin JP, Cullen N, Sheline YI, Taylor WD, Aselcioglu I, Cook PA, et al. Harmonization of cortical thickness measurements across scanners and sites. Neuroimage. 2018;167(June 2017):104\u0026ndash;20.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e25. Budd S, Robinson EC, Kainz B. A survey on active learning and human-in-the-loop deep learning for medical image analysis. Med Image Anal. 2021;71:102062.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e26. Albayram MS, Smith G, Tufan F, Tuna IS, Bostancıklıoğlu M, Zile M, et al. Non-invasive MR imaging of human brain lymphatic networks with connections to cervical lymph nodes. Nat Commun. 2022;13(1):203.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e27. Radoš M, Živko M, Periša A, Orešković D. No Arachnoid Granulations \u0026mdash; No Problems : Number, Size, and Distribution of Arachnoid Granulations From Birth to 80 Years of Age. Front Aging Neurosci. 2021;13:698865.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e28. Antila S, Karaman S, Nurmi H, Airavaara M, Voutilainen MH, Mathivet T, et al. Development and plasticity of meningeal lymphatic vessels. J Exp Med. 2017;214(12):3645\u0026ndash;3667.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e29. Decimo I, Dolci S, Panuccio G, Riva M, Fumagalli G, Bifari F. Meninges : A Widespread Niche of Neural Progenitors for the Brain. Neuroscientist. 2021;27(5):506\u0026ndash;528.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e30. Bifari F, Berton V, Pino A, Kusalo M, Malpeli G, Chio M Di, et al. Meninges harbor cells expressing neural precursor markers during development and adulthood. Front Cell Neurosci. 2015;9:383.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e31. Semple BD, Blomgren K, Gimlin K, Ferriero DM, Noble-haeusslein LJ. Brain development in rodents and humans : Identifying benchmarks of maturation and vulnerability to injury across species. Prog Neurobiol. 2013;106\u0026ndash;107:1\u0026ndash;16.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"fluids-and-barriers-of-the-cns","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"fbcn","sideBox":"Learn more about [Fluids and Barriers of the CNS](http://fluidsbarrierscns.biomedcentral.com/)","snPcode":"12987","submissionUrl":"https://submission.nature.com/new-submission/12987/3","title":"Fluids and Barriers of the CNS","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Parasagittal dura, Autism spectrum disorder, Pediatric, Deep learning, MRI, segmentation, neurofluids","lastPublishedDoi":"10.21203/rs.3.rs-9556003/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9556003/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe parasagittal dura (PSD) is a specialized meningeal structure that contains lymphatic vessels and contributes to cerebrospinal fluid (CSF) drainage and neuroimmune surveillance. In adults, aging and neurodegenerative diseases have been associated with increased PSD volume, suggesting that it may represent a potential biomarker of neurodegeneration. However, its functional role and the extent to which PSD volume may reflect pathological processes during early brain development remain poorly understood. Furthermore, no automated tools currently exist to segment the PSD in the very young developing brain. We present a pediatric automated tool for PSD segmentation and investigate PSD volume in children with autism spectrum disorder (ASD) and typically developing (TD) controls.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe developed a fully automated segmentation framework based on a 3D nnU-Net applied to clinical 3D T2-FLAIR MRI. The model was trained on manually annotated pediatric ASD scans (n = 144) and evaluated on an independent ASD test set (n = 56). To assess generalizability, the trained model was applied to independent typically developing (TD) cohorts, including an external dataset acquired with different imaging parameters. Total and regional PSD volumes (prefrontal, fronto-parietal, occipital) were quantified and compared between ASD and TD children, and correlations with age and brain volumes were examined.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe segmentation framework achieved high performance in the ASD cohort (Dice–Sørensen coefficient (DSC) = 0.93 ± 0.02). Comparable performance was observed in the TD cohort (DSC = 0.90 ± 0.02), including the cohort acquired with different imaging parameters (DSC = 0.85 ± 0.04), indicating good generalizability. Performance remained consistent across PSD regions, with slightly higher accuracy in central segments.\u003c/p\u003e\n\u003cp\u003eNo significant differences in PSD volumes were found between ASD and TD groups. PSD volume is correlated with CSF (and ea-CSF), supporting a link with fluid compartment dynamics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003cbr\u003e\nWe present a publicly available reliable automated tool for whole-brain PSD segmentation in the pediatric population using clinical 3D T2-FLAIR sequences. Our results show no differences in PSD volumes between ASD and TD groups, while consistently demonstrating a positive correlation with CSF-related compartments. This tool provides a scalable approach for studying the role of PSD in neurofluid circulation during brain development.\u003c/p\u003e","manuscriptTitle":"Automated 3D nnU-Net Segmentation of the Parasagittal Dura in Children with Autism Spectrum Disorder Using Clinical 3D T2-FLAIR","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-13 18:03:58","doi":"10.21203/rs.3.rs-9556003/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-05-18T00:44:33+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-14T17:12:08+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"96631800435378104922800438495585815116","date":"2026-05-07T09:33:55+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"232761438525352981687948366491367409857","date":"2026-05-05T16:00:09+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-05-05T08:55:30+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-30T16:38:55+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-30T11:58:07+00:00","index":"","fulltext":""},{"type":"submitted","content":"Fluids and Barriers of the CNS","date":"2026-04-28T15:05:51+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"fluids-and-barriers-of-the-cns","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"fbcn","sideBox":"Learn more about [Fluids and Barriers of the CNS](http://fluidsbarrierscns.biomedcentral.com/)","snPcode":"12987","submissionUrl":"https://submission.nature.com/new-submission/12987/3","title":"Fluids and Barriers of the CNS","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"07458c3f-77ac-4ffd-a4c8-e72be9536cec","owner":[],"postedDate":"May 13th, 2026","published":true,"recentEditorialEvents":[{"type":"editorInvitedReview","content":"","date":"2026-05-18T00:44:33+00:00","index":15,"fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-14T17:12:08+00:00","index":14,"fulltext":""},{"type":"reviewerAgreed","content":"96631800435378104922800438495585815116","date":"2026-05-07T09:33:55+00:00","index":11,"fulltext":""},{"type":"reviewerAgreed","content":"232761438525352981687948366491367409857","date":"2026-05-05T16:00:09+00:00","index":10,"fulltext":""},{"type":"reviewersInvited","content":"5","date":"2026-05-05T08:55:30+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-30T16:38:55+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-30T11:58:07+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-13T18:03:59+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-13 18:03:58","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9556003","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9556003","identity":"rs-9556003","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}