Human Brain Mapping

Total intracranial volume (TIV) is a major confounding factor in neuroimaging studies, particularly when studying sex differences in the brain. Different methods have been proposed to adjust for this effect, however, their impact has not been directly studied and compared. In this study, we sought to evaluate the impact of four most commonly used adjustment methods in the literature on the estimations of neuroanatomical sex differences. These methods included: the proportions method, the residuals method, the power corrected proportions method, and adding TIV as a covariate in a regression analysis. Leveraging data from the UK Biobank, we employed a matching approach to devise a gold standard as reference for comparing these methods. To achieve this, we matched the male and female participants based on age and TIV to remove the impact of TIV differences between sexes. We further modeled aging trajectories at the regional level, vertexwise, and voxelwise, using raw and adjusted values, and compared the obtained estimates against the gold standard. We found that across different metrics, adding TIV as a covariate was the best-performing method for removing the effect of TIV, in terms of the correlation between the estimates of the different subsamples and the gold standard as well as the degree of estimation bias. Furthermore, we showed that the commonly used smoothing of the morphometric measures can result in biased estimation of sex differences in these measures. Finally, we showed that while small in effect size, there still remains some neuroanatomically specific uncorrected effects for all adjustment methods.

Large-scale neuroimaging studies increasingly pool data across multiple cohorts, scanners, and acquisition protocols, introducing technical between-cohort variation that must be addressed before meaningful biological inference can be drawn. Existing harmonisation methods, particularly ComBat-based approaches, have been widely adopted for this purpose. However, they remain limited by Gaussian assumptions and by their focus on location or location-scale correction. In this study, we propose a unified hierarchical Generalised Additive Models for Location, Scale and Shape (GAMLSS) framework for multi-cohort harmonisation and normative modelling of structural neuroimaging data. The framework models cohort effects directly within all fitted distributional parameters, accommodates any parametric family for which exact inverse mapping is available, and returns harmonised values on the original measurement scale through centile-based quantile mapping. Normative deviation scores are obtained as a direct by-product of the same fitted model, enabling harmonisation and normative inference to be conducted jointly. The method was evaluated in a pooled longitudinal dataset comprising 88,126 observations across 237 structural neuroimaging features from six cohorts spanning childhood to late life: ABCD, IMAGEN, NCANDA, LIFE, UK Biobank, and MAS. Harmonisation performance was compared with ComBat, ComBat-GAM, and ComBat-LS using complementary criteria assessing data retention, residual batch effects, preservation of age-related and sex-related biological signal, and coherence of post-harmonisation lifespan trajectories. GAMLSS achieved near-complete removal of residual cohort effects, retained almost all valid observations post-harmonisation, and showed the strongest overall preservation of biological signal across validation metrics. In particular, it better preserved biologically plausible age trajectories for distributionally complex features such as white matter hypointensity volume, while simultaneously providing harmonised native-scale values and normative deviation scores within a single framework. These findings suggest that hierarchical GAMLSS offers a flexible and practical alternative to existing ComBat-based methods for large-scale neuroimaging harmonisation, particularly for features with non-Gaussian residual distributions and settings where cohort effects extend beyond differences in mean and variance.

Foundation models pre-trained on large neuroimaging datasets offer a promising approach to overcome the limited sample sizes typical of mental health imaging studies, yet their generalization across diverse clinical populations remains unclear. We present the first systematic benchmark of four publicly available structural MRI foundation models -- AnatCL, BrainIAC, 3D-Neuro-SimCLR, and SwinBrain -- on tasks relevant to mental health research. Using T1-weighted MRI from Parkin-sons Progression Markers Initiative (PPMI), Healthy Brain Network (HBN), and Nathan Kline Institute (NKI), we evaluate these models on sex classification, brain age prediction, and Parkinsons disease (PD) classification, benchmarking against models trained from FreeSurfer-derived cortical thickness and cortical surface area features, as well as an un-trained CNN baseline. Although some individual foundation models out-performed FreeSurfer on particular tasks and datasets, 3D-Neuro-SimCLR demonstrated the most consistent performance overall, with the notable exception of HBN sex classification, and all models failed to classify early-stage Parkinsons disease above chance. Notably, untrained CNNs achieved performance comparable to or exceeding FreeSurfer in multiple instances, establishing them as computationally efficient reference models. The cross-model feature correlation analysis reveals that foundation model representations correlate differently with traditional cortical measurements. These findings position structural MRI foundation models, particularly 3D-Neuro-SimCLR and AnatCL, as promising avenues to boost the performance of neuroimaging predictive models in mental health.

Combining multi-site MRI datasets increases statistical power and model generalisability but may be hindered by variability between sites. Harmonisation methods aim to remove potentially confounding variance while preserving biologically meaningful signals. However, this can be challenging, as each T1-weighted image reflects both scanner properties (e.g., field strength, sequence parameters) and individual biological characteristics (e.g., age, sex, ethno-cultural background, and pathology). Two image-based (HACA3, IGUANe) and two feature-based (neuroHarmonize, neuroCombat) harmonisation methods were assessed using T1-weighted brain imaging data from the Psy-ShareD database; 564 participants (295 schizophrenia, 269 controls) from seven studies acquired across 5 sites from the Psy-ShareD database. We trained several models to classify sites, schizophrenia diagnosis, age, and symptom levels. Site-classification accuracy was high for unharmonised data (90.1%) and for HACA3 (92.2%), slightly reduced with IGUANe (86.6%), and near chance for feature-based methods (4.2% neuroHarmonize; 1.8% neuroCombat), indicating effective bias removal. We fitted several models predicting biological signals including diagnosis, age, and symptom levels across different harmonisation methods. In most cases, classification with harmonised data performed at least as well as with unharmonised data. Generally, feature-based methods best remove site-related variance, but image-based approaches remain a promising avenue for preserving individual biological differences. This work provides practical guidance for selecting harmonisation strategies in multi-site psychiatric neuroimaging, depending on whether the priority is bias reduction or preservation of subject-level variability.

Prader-Willi Syndrome (PWS) is a rare genetic condition with multifaceted physical, behavioural and cognitive difficulties that is characterized by hyperphagia and low executive functioning. Food-seeking behaviours may be moderated by hormonal, cognitive, and psychological factors, and are thought to be mediated in part by functional brain abnormalities. Here, we used an experimental protocol integrating eyes opens resting state magnetoencephalography (MEG) - a high-resolution neurophysiological imaging technique - and neuropsychological profiling to understand the relationship between executive functioning, and intrinsic brain activity & functional connectivity in a prospective, cross-sectional cohort with PWS, and a sex-, age- and BMI-matched control group. We observed lower executive functioning in PWS as well as functional dysconnectivity across multiple channels of brain synchrony - in other words, across multiple frequency bands that mediate communication within and between brain networks - in the visual, attentional, and the default mode networks. Moreover, we found brain-wide changes in the topological structure of brain networks in those with PWS, with increased hubness of functional networks, but decreased centrality. However, none of these measures survived multiple comparison correction after correlating with neuropsychological outcomes, although there were moderate effect sizes (degree of association). This is the first study to combine neuropsychology and neurophysiological imaging to show that functional synchrony in multiple brain networks is dysregulated in PWS.

Frontotemporal Dementia (FTD) is a neurodegenerative disorder characterized by extensive atrophy in the frontal and temporal lobes of the brain as well as high cerebrovascular burden. While anatomical Magnetic Resonance Imaging (MRI) is well established for quantifying brain atrophy in FTD, the variability in (pre-)processing methods limits the generalizability and comparability of findings. This study systematically compared the robustness and sensitivity of multiple widely used neuroimaging metrics, namely Deformation-Based Morphometry (DBM), Voxel-Based Morphometry (VBM), Cortical Thickness (CT), and segmentation-based grey matter Volumes, in detecting atrophy across FTD subtypes. We processed 732 T1-weighted MRI scans from 156 participants with FTD and 139 healthy controls from the Frontotemporal Lobar Degeneration Neuroimaging Initiative using our in-house pipeline PELICAN (Dadar et al., 2025) for volumetric measures and FreeSurfer version 7 (Fischl, 2012) for CT and grey matter segmentations. Visual quality control using consistent quality control images at each step of the pipelines revealed significantly higher failure rates for CT (38.52%) and FreeSurfer segmentations (23.63%) relative to PELICANs volumetric measures (2.04% DBM, 3.05% VBM). Failure rates differed between FTD subtypes and were related to pathological burden. Particularly for FreeSurfer, errors occurred predominantly in regions with high prevalence of atrophy and White Matter Hyperintensities. In PELICAN, the addition of a FTD-specific template as an intermediate step during nonlinear registration decreased the failure rates in this step in the FTD population. We then applied linear regression models to assess each metrics sensitivity in detecting cross-sectional differences between FTD groups controls as well as linear mixed-effects models to determine which method is most sensitive to longitudinal anatomical changes. While CT yielded effect sizes comparable to VBM and DBM when analyzing the same subset of successfully processed scans, VBM and DBM demonstrated enhanced power to detect effects due to lower failure rates and higher participant retention in the full sample. Overall, we demonstrate that image processing methodology and pipeline selection profoundly influences effect sizes and statistical power to detect meaningful between-group differences or longitudinal changes. Volumetric measures (DBM and VBM) yielded sufficiently robust pipeline outcomes to maintain adequate statistical power for capturing atrophy patterns after quality control procedures.

Understanding the complex relationships among genetic variations, brain structural anatomy, and functional alterations is a fundamental yet challenging task in neuroimaging genetics. In this study, we employ a causal mediation framework under structural modeling to systematically investigate the mediating role of interand intra-network structural connectivity (SC) in linking whole-genome single nucleotide polymorphisms (SNPs) to brain functional connectivity (FC) across both resting-state and task-based conditions during neurodevelopment. Utilizing baseline and follow-up SC and FC network traits along with [~]500k SNPs along the genome from 11,666 unique subjects under the Adolescent Brain Cognitive Development (ABCD) study, we first conduct genome-wide association studies (GWAS) to identify candidate SNPs associated with structural and functional network traits. Subsequently, mediation analyses reveal key genetic exposures that directly influence brain functional networks and indirectly impact FC through SC network mediators. These results provide deeper insights into how genetic variations shape brain structural and functional network organizations, along with revealing the influence of brain anatomical topologies on functional fingerprints. This work enhances the understanding of the causal effect pathways among genetic factors and large-scale brain structural and functional networks, advancing our understanding of the genetic underpinnings of neurodevelopmental processes.

Preterm birth still represents a concrete emergency to be managed and addressed globally. Since cerebral white matter injury is the major form of brain impairment in survivors of premature birth, the identification of reliable, non-invasive markers of altered white matter development is of utmost importance in diagnostics. Diffusion MRI has recently emerged as a valuable tool to investigate these kinds of alterations. In this work, rather than focusing on a single MRI modality, we worked on a compound of beyond-DTI High Angular Resolution Diffusion Imaging (HARDI) techniques in a group of 46 preterm babies studied on a 3T scanner at term equivalent age and in 23 control neonates born at term. After extracting relevant derived parameters, we examined discriminative patterns of preterm birth through (i) a traditional voxel-wise statistical method such as the Tract-Based Spatial Statistics approach (TBSS); (ii) an advanced Machine Learning approach such as the Support Vector Machine (SVM) classification; (iii) establishing the degree of association between the two methods in voting white matter most discriminating areas. Finally, we applied a multi-set Canonical Correlation Analysis (CCA) in search for sources of linked alterations across modalities. TBSS analysis showed significant differences between preterm and term cohorts in several white matter areas for multiple HARDI features. SVM classification performed on skeletonized HARDI measures produced satisfactory accuracy rates, especially as for highly informative parameters about fibers directionality. Assessment of the degree of overlap between the relevant measures identified by the two methods exhibited a good, though parameter-dependent rate of agreement. Finally, CCA analysis identified joint changes precisely for those features exhibiting less correspondence between TBSS and SVM. Our results suggest that a data-driven intramodal imaging approach is crucial to extract deep and complementary information that cannot be extracted from a single modality.

Withdrawal StatementThe authors have withdrawn this manuscript because during the peer-review process, they realised that their interpretation of the brain-age model presented in this paper was not fully accurate. While the analyses, statistics, and results remain valid, their interpretation of the internal test set performance metrics was inaccurate due to the non-linear shape of the distribution. In other words, although the overall R{superscript 2} is correctly reported as 0.80, this value does not capture the variability of the metrics across different age bins. For this reason, the authors are withdrawing the preprint. Therefore, the authors do not wish this work to be cited as reference for the project. The authors aim to re-run the analysis to provide a more robust version of the model and a new DOI will be linked on this page once the revised work is available. If you have any questions, please contact the corresponding author.

This study presents COGNET-STROKE, a novel meta-analytic brain functional decoding tool designed to predict cognitive deficits after stroke. The tool integrates Lesion Network Mapping with meta-analytic concept maps derived from Neurosynth, enabling the identification of the cognitive domains most affected by stroke-induced network disruptions. Validation analyses confirmed that identified patients with predicted motor and sensory impairment had significantly higher scores in their corresponding NIHSS-subscales, demonstrating predictability from lesion-induced dysconnectivity to behavioral impairment. COGNET-STROKE openly available at https://github.com/compneurobilbao/CogNet-stroke, offers a framework for individualized cognitive-deficit profiling, with implications for personalized rehabilitation strategies in stroke recovery.

Most psychiatric disorders are highly heritable and associated with altered brain structural and functional patterns. Data fusion analyses on brain imaging and genetics, one of which is parallel independent component analysis (pICA), enable the link of genomic factors to brain patterns. Due to the small to modest effect sizes of common genetic variants in psychiatric disorders, it is usually challenging to reliably separate disorder-related genetic factors from the rest of the genome with the typical size of clinical samples. To alleviate this problem, we propose sparse parallel independent component analysis (spICA) to leverage the sparsity of individual genomic sources. The sparsity is enforced by performing Hoyer projection on the estimated independent sources. Simulation results demonstrate that the proposed spICA yields improved detection of independent sources and imaging-genomic associations compared to pICA. We applied spICA to gray matter volume (GMV) and single nucleotide polymorphism (SNP) data of 341 unrelated adults, including 127 controls, 167 attention-deficit/hyperactivity disorder (ADHD) cases, and 47 unaffected siblings. We identified one SNP source significantly and positively associated with a GMV source in superior/middle frontal regions. This association was replicated with a smaller effect size in 317 adolescents from ADHD families, including 188 individuals with ADHD and 129 unaffected siblings. The association was found to be more significant in ADHD families than controls, and stronger in adults and older adolescents than younger ones. The identified GMV source in superior/middle frontal regions was not correlated with head motion parameters and its loadings (expression levels) were reduced in adolescent (but not adult) individuals with ADHD. This GMV source was associated with working memory deficits in both adult and adolescent individuals with ADHD. The identified SNP component highlights SNPs in genes encoding long non-coding RNAs and SNPs in genes MEF2C, CADM2, and CADPS2, which have known functions relevant for modulating neuronal substrates underlying high-level cognition in ADHD.

Analytical variability in neuroimaging pipelines contributes to concerns about reproducibility in the field. In structural MRI, different segmentation tools produce discrepant morphometric estimates that may influence downstream analyses, such as predictive modeling. We tested whether integrating several segmentation pipelines improves brain age prediction, and characterized the spatial and demographic structure of pipeline differences across several datasets. T1-weighted scans from five open-access datasets were processed with four widely-used structural segmentation pipelines. Brain age models were trained using single-pipeline features and compared with multi-pipeline aggregation strategies. Inter-pipeline variability was assessed across shared subcortical structures and examined in relation to age and sex. Integrating features across distinct segmentation frameworks improved predictive performance relative to individual pipelines, whereas aggregation within closely related software versions provided limited benefit. Variability was spatially structured and volumetric measures were often systematically associated with age and sex. These results suggest that segmentation differences reflect structured, demographically sensitive variation rather than random noise, and that multi-pipeline feature integration can enhance robustness in neuroimaging-based prediction.

PurposeDiffusion MRI (dMRI) data typically suffer of significant cross-site variability, which prevents naively performing pooled analyses. To attenuate cross-site variability, harmonization methods such as the rotational invariant spherical harmonics (RISH) have been introduced to harmonize the dMRI data at the signal level. A common requirement of the RISH method, is the availability of healthy individuals who are matched at the group level, which may not always be readily available, particularly retrospectively. In this work, we propose a framework to harmonize dMRI without matched training groups. MethodsOur framework learns harmonization features while controlling for potential covariates using a voxel-based generalized linear model (RISH-GLM). RISH-GLM allows to simultaneously harmonize data from any number of sites while also accounting for covariates of interest, thus not requiring matched training subjects. Additionally, RISH-GLM can harmonize data from multiple sites in a single step, whereas RISH is performed for each site independently. ResultsWe considered data of training subjects from retrospective cohorts acquired with 3 different scanners and performed 3 harmonization experiments of increasing complexity. First, we demonstrate that RISH-GLM is equivalent to conventional RISH when trained with data of matched training subjects. Secondly, we demonstrate that RISH-GLM can effectively learn harmonization with two groups of highly unmatched subjects. Thirdly, we evaluate the ability of RISH-GLM to simultaneously harmonize data from 3 different sites. DiscussionRISH-GLM can learn cross-site harmonization both from matched and unmatched groups of training subjects, and can effectively be used to harmonize data of multiple sites in one single step.

Recent advances in neuroimaging modeling highlight the importance of accounting for subgroup heterogeneity in population-based neuroscience research through various investigations in large scale neuroimaging data collection. To integrate survey methodology with neuroscience research, we present an imaging data analysis aiming to achieve population generalizability with screened subsets of data. The Adolescent Brain Cognitive Development (ABCD) Study has enrolled a large cohort of participants to reflect the individual variation of the U.S. population in adolescent development. To ensure population representation, the ABCD Study has released the base weights. We estimated the associations between brain activities and cognitive performance using the functional Magnetic Resonance Imaging (fMRI) data from the ABCD Studys n-back working memory task. Notably, the imaging subsample exhibits differences from the baseline cohort in key child characteristics, and such discrepancies cannot be addressed simply by applying the ABCD base weights. We developed new population weights specific to the subsample and included the adjusted weights in the image-on-scalar regression model. We validated the approach through synthetic simulations and applications to fMRI data from the ABCD Study. Our findings indicate that population weighting adjustments influence association estimates between brain activities and cognition, emphasizing the importance of evaluating validity and generalizability in population neuroscience research.

IntroductionBrain-age is an estimate of the brains biological age derived from neuroimaging data, and has been proposed as a biomarker of brain health and disease risk. While brain-age estimation commonly uses high-field (HF) magnetic resonance imaging (MRI) (> 1.5 T) this is costly and inaccessible, limiting its applicability. Emerging ultra-low-field (ULF) MRI (< 0.1 T) technology is a cheaper and more accessible alternative, but its lower resolution raises questions about whether biomarkers like brain-age can be estimated reliably. MethodsWe assessed different brain-age pipelines in 23 adults scanned on one HF system (GE Signa Premier at 3 T) and two identical ULF systems (Hyperfine Swoop at 64 mT). 14 distinct acquisitions were used, defined by T1-or T2-weighting, resolution, and preprocessing: raw anisotropic orientations (axial, coronal, sagittal), isotropic scans, and super-resolution derivatives from multi-resolution registration (MRR) and SynthSR. These inputs (a total of n = 573 scans) were analysed with five brain-age software packages (BrainageR, SynthBA, MIDI, DeepBrainNet, Py-BrainAge). Performance evaluation entailed validity (brain-age vs. actual age), correspondence (ULF brain-age vs. HF brain-age), and test-retest reliability (ULF1 brain-age vs. ULF2 brain-age). ResultsOverall, performance was mixed across pipelines, though several ULF pipelines achieved performance comparable to HF. The four best-performing combinations were SynthBA on T2 scans without SynthSR, MIDI on T2 scans without SynthSR, PyBrainAge on T1 scans with SynthSR and using FreeSurfer recon-all-clinical, and BrainageR on T1 scans with SynthSR. These showed moderate-to- strong validity (r = 0.76-0.92, R2 = 0.54-0.64, MAE = 6.49-8.21 years), moderate- to-strong correspondence to HF (r = 0.84-0.93, ICC = 0.72-0.92), and excellent test-retest reliability (r = 0.97-0.99, ICC = 0.97-0.99). Moreover, some anisotropic acquisitions achieved comparable validity and reliability to MRR images when tested with the best-performing model, SynthBA (R2 = 0.57-0.62, ICC [CI] = 0.99 [0.97- 1.00], for coronal T2). ConclusionThis first systematic evaluation of brain-age at ULF demonstrates that accurate and reliable estimates can be achieved across multiple pipelines, with- out necessarily requiring image enhancement. Performance depended on the combination of model, scan type, and preprocessing. ULF brain-age estimation could be a practical and scalable tool for clinical decision-making, population research, and long-term patient monitoring, thereby helping to make advanced neuroimaging biomarkers more accessible worldwide.

3D surface-based computational mapping is more sensitive to localized brain alterations in neurological, developmental and psychiatric conditions than traditional gross volumetric analysis, providing fine-scale 3D maps of a wide range of surface-based features. Here we introduce a scalable toolkit for large-scale computational surface analysis, with efficient algorithms for multisite data integration, statistical harmonization, accelerated multivariate statistics, and visualization. We showcase the utility of the toolkit by mapping subcortical shape variations and factors that affect them across 21 international samples from the ENIGMA Bipolar Disorder Working Group (N=3,373).

Compared to traditional gross volumetrics, surface- based models provide greater spatial precision for understanding brain alterations related to developmental, neurological, and psychiatric disorders. Large-scale brain initiatives are combining data from around the world to discover and improve illness- related brain markers. Here, we present a toolkit for 3D brain geometry analysis aimed at addressing key challenges facing large- scale neuroimaging studies. Our framework incorporates scalable methods for multisite data integration, site-specific confound correction, accelerated statistical modeling, interpretable machine learning, and interactive results visualization. The toolkit was tested on data from 21 independently collected study samples participating in the ENIGMA Bipolar Disorder Working Group (N = 3,373). Compared to traditional volume features, we show how subcortical shape measures can be combined across study sites to capture spatially complex differences between diagnostic groups and associations with common treatments. Statistical modeling was accelerated using the Fast and Efficient Mixed- Effects Algorithm (FEMA) and achieved a 16-fold reduction in computation time compared to traditional approaches. Machine learning models showed shape features may provide greater predictive performance over traditional volumes for both diagnostic and treatment prediction tasks, with interpretable weight maps providing insights into the local features driving model performance.

Abstract / Key pointsO_LIThe influence of FreeSurfer version-dependent variability in reconstructed cortical features on brain age predictions is average small when varying training and test splits from the same data. C_LIO_LIFreeSurfer version differences can lead to some variability in brain age dependent on the choice of algorithm and individual differences in brain morphometry, highlighting the advantage of repeated random train-test splitting. C_LIO_LIShuffling of differently processed FreeSurfer data dependent on the FreeSurfer version increases performance and generalizability of the brain age prediction model. C_LI

An open discussion in studies of intrinsic brain functional connectivity is the mitigation of head motion-related artifacts, particularly in the presence of peculiar symptomatology such as in Parkinsons disease (PD). Previous studies show that Independent Component Analysis (ICA) denoising improves the reproducibility of functional connectivity findings by detecting sources of non-neural signals. However, there is still no consensus about which pre-processing pipeline should be applied in natural high motion populations such as PD, particularly in relation to novel functional network descriptions derived from dynamic connectivity analyses. In this study, we investigated how different pre-processing pipelines affect intrinsic brain connectivity metrics, both static and dynamic, derived from a group of young healthy controls (HC) and a group of PD participants. A total of 20 HC and 20 PD subjects participated in this 3 T MRI study. Resting-state functional MRI images were used to test the effects of the pre-processing pipeline of static (sFC) and temporal-varying functional connectivity (dFC) estimations. Both MRI datasets were pre-processed using three different workflows differing in the motion correction approach: (i) standard motion realignment (mc); (ii) motion outlier detection and deweighting based on image intensity change estimations (DVARS) and (iii) ICA-based noise removal using reference noise features (AROMA). Furthermore, the PD dataset was also processed with a fourth method by applying an ICA-based denoising (FIX), previously trained on the HC group. sFC analysis was performed using Group ICA, by temporally concatenating different pre-processing types in pairs of different runs. Two types of dFC analyses were considered: innovation-driven co-activation patterns (iCAPs) and co-activation patterns (CAPs). CAPs allow dFC estimations that do not require the deconvolution of the hemodynamic response function and its derivative, thus potentially being less sensitive to head-motion related noise. We found that regardless of substantial head motion differences in the two groups, sFC results were consistent across denoising strategies. Conversely, dFC was extremely sensitive to denoising strategies, particularly for the PD group with the transient-based dFC analyses. Indeed, the use of the peak-based dFC framework enables the detection of time-varying networks but in a way that is highly dependent on the motion correction pipeline. In conclusion, we show that dynamic functional network representations are highly sensitive to both head motion and to fMRI denoising methods. These findings stress the importance of considering and reporting these experimental aspects to help with the reproducibility and interpretation of different studies. Future work is needed to further investigate transient-based dFC strategies that are more robust to head motion.

IntroductionEnlarged perivascular spaces (PVS) are imaging markers of cerebral small vessel disease (CSVD) that are associated with age, disease phenotypes, and overall health. Quantification of PVS is challenging but necessary to expand an understanding of their role in cerebrovascular pathology. Accurate and automated segmentation of PVS on T1-weighted images would be valuable given the widespread use of T1-weighted imaging protocols in multisite clinical and research datasets. MethodsWe introduce segcsvdPVS, a convolutional neural network (CNN)-based tool for automated PVS segmentation on T1-weighted images. segcsvdPVS was developed using a novel hierarchical approach that builds on existing tools and incorporates robust training strategies to enhance the accuracy and consistency of PVS segmentation. Performance was evaluated using a comprehensive evaluation strategy that included comparison to existing benchmark methods, ablation-based validation, accuracy validation against manual ground truth annotations, correlation with age-related PVS burden as a biological benchmark, and extensive robustness testing. ResultssegcsvdPVS achieved strong object-level performance for basal ganglia PVS (DSC = 0.78), exhibiting both high sensitivity (SNS = 0.80) and precision (PRC = 0.78). Although voxel-level precision was lower (PRC = 0.57), manual correction improved this by only ~3%, indicating that the additional voxels reflected primary boundary- or extent-related differences rather than correctable false positive error. For non-basal ganglia PVS, segcsvdPVS outperformed benchmark methods, exhibiting higher voxel-level performance across several metrics (DSC = 0.60, SNS = 0.67, PRC = 0.57, NSD = 0.77), despite overall lower performance relative to basal ganglia PVS. Additionally, the association between age and segmentation-derived measures of PVS burden were consistently stronger and more reliable for segcsvdPVS compared to benchmark methods across three cohorts (test6, ADNI, CAHHM), providing further evidence of the accuracy and consistency of its segmentation output. ConclusionssegcsvdPVS demonstrates robust performance across diverse imaging conditions and improved sensitivity to biologically meaningful associations, supporting its utility as a T1-based PVS segmentation tool.