Parasagittal Dural Space and Cerebrospinal Fluid (CSF) Changes Across the Lifespan in Healthy Adults: Implications for Glymphatic Flow

Background: Recent studies have suggested the importance of a glymphatic clearance pathway for brain parenchymal metabolic waste products. One fundamental but relatively under-explored component of this pathway is the anatomic region surrounding the superior sagittal sinus, which has been hypothesized to encompass lymphatic vessels. This so-called parasagittal dural (PSD) space likely plays a physiologically signicant role at the distal intracranial component of the human glymphatic circuit, yet owing to the relative novelty of this discovery, fundamental gaps persist in our knowledge of how this space changes with normal aging and intracranial bulk uid transport. Methods: We tested the hypotheses that volumetric magnetic resonance imaging (MRI) measures of the PSD space (i) are directly related to cerebrospinal uid (CSF) ow at the cerebral aqueduct, and (ii) increase with age. Healthy participants (n=62; age range = 20-83 years) provided informed, written consent and multi-modal 3 Tesla MRI was performed including phase contrast assessment of the CSF ow through the aqueduct of Sylvius, T 1 -weighted and T 2 -weighted MRI for tissue volume and PSD assessment. Standard anatomical and cognitive testing were applied to conrm inclusion criteria. PSD volume was extracted using a recently validated neural networks algorithm. Non-parametric regression models were applied to evaluate how PSD volume related to tissue volume and age cross-sectionally, and separately how PSD volume related to CSF ux (signicance criteria: two-sided p<0.05). Results: A signicant enlargement of PSD volume in relation to normal aging (p<0.001, Spearman’s-=0.6), CSF volume (p<0.001, Spearman’s- =0.6) and bulk CSF ux through the aqueduct of Sylvius (anterograde and retrograde, p<0.001) were observed. The elevation in PSD volume was not signicantly related to changes in tissue volume (p=0.11 and p=0.24 for gray and white matter, respectively). Findings are consistent with PSD volume increasing with age and bulk CSF ux. Conclusions: The ndings of this study are two folds, rst they highlight the feasibility of quantifying PSD volume non-invasively in vivo in humans using machine learning and non-contrast MRI. Second, that PSD volume increases with age, and relates to bulk CSF volume and ux. Values reported should provide useful normative ranges for how PSD volume adjusts with age, which will serve as a necessary pre-requisite for comparisons to persons with neurodegenerative disorders. In addition, the correlation analysis using Spearman’s correlation coecient indicated a strong correlation of PSD volume with a correlation coecient (p<0.001). Two linear models were applied to evaluate the relationship between PSD volume and CSF ux in the cerebral aqueduct. The results provide evidence for a signicant correlation of PSD volume and CSF ux for both anteretrograde and retrograde directionality. The model indicates that anteretrograde CSF ux correlates with PSD volume with a p-value equal to 0.01 (q-value equal to 0.04), and a Spearman’s correlation coecient of 0.36 (p-value=0.001). Retrograde CSF ux show stronger correlation, with a p-value=0.001 (q-value=0.004), and a Spearman’s correlation coecient of -0.50 (p-value inferior to 0.001).


Introduction
The proposed glial-lymphatic system, or glymphatic system, has prompted a reassessment of cerebrospinal uid (CSF) as an important mediator in neurologic waste product clearance. Increasing animal and human evidence supports the presence of this highly organized system whereby uid passes from the subarachnoid space to the periarterial space en route to the brain parenchymal interstitial space [1]- [4]. Net uid motion is also observed exiting the interstitial space via the perivenous space, en route back to the subarachnoid space. Impairment of this system has been suggested as a mechanism of waste product accumulation which may underly the development of multiple neurodegenerative diseases with unknown etiology, including but not limited to Alzheimer's disease, Parkinson's disease, and multiple sclerosis [5], [6].
Classically, the central nervous system (CNS) is recognized to be devoid of lymphatic vessels. However, there is emerging evidence that lymphatic vessels are indeed present in the region surrounding the superior sagittal sinus [7]. Following intrathecal gadolinium contrast administration in 18 human subjects, contrast was noted to concentrate in the CSF spaces near the vertex. With passage of time, contrast progressively accumulated within the tissue surrounding the superior sagittal sinus, or the parasagittal dural (PSD) space [8]. These ndings suggest that trans-arachnoid molecular passage occurs in this region and highlights the PSD space as a potential bridging link between the glymphatic circulation within the human brain and the dural lymphatic vessels. Despite this, there is limited knowledge of how variance in PSD space relates to aging, brain health, and intracranial CSF ow.
More speci cally, the primary source of CSF production is the choroid plexus (ChP), which is estimated to produce CSF at a rate of approximately 20 ml/h in adult humans [9]. While most of the ChP tissue resides within the atria of the lateral ventricles, there is choroidal tissue throughout the ventricular system in rough proportion to the overall size of the ventricular components. CSF produced in the lateral and third ventricles traverses the cerebral aqueduct (e.g., aqueduct of Sylvius) en route to the 4th ventricle and on to the more diffuse subarachnoid space [10], [11]. As the cerebral aqueduct represents the sole pathway for CSF e ux from the lateral and third ventricles, measurement of ow in this region offers an opportunity to quantify uid production by the third and lateral ventricles, which comprise the largest ChP complexes. To measure CSF e ux in the cerebral aqueduct, magnetic resonance (MR) phase contrast sequences, traditionally sensitive to arterial or venous ow, have been re-parameterized to enable quantitative CSF ow, primarily by pairing with cardiac phase and reducing the velocity encoding gradient to coincide with CSF ow (typically 10-20 cm/s). [12]. Aberrant phase contrast ow parameters have long been associated with idiopathic normal pressure hydrocephalus [13]. CSF ow through the cerebral aqueduct is also dependent on cardiac phase and has been shown to be highest in older adults and in males vs. females [14]. These ndings are consistent with previous studies that have shown age, and sex dependencies of CSF on tissue volume [15], [16], and suggest that CSF ow pro les may relate to total CSF volume.
The logical extension of this work is to understand how such CSF ow parameters relate to quantitative estimates of the PSD volume. Here, we test the hypotheses that increases in PSD volume parallel changes in CSF ow through the cerebral aqueduct, and additionally that PSD volume increases with advancing age. Findings are interpreted in the context of the growing literature on glymphatic physiology and bulk CSF ow.

Participants
All participants provided informed, written consent in accordance with the local institutional review board (IRB) and consistent with the Declaration of Helsinki and its amendments. All participants were scanned between January 2020 and September 2021 at Vanderbilt University Medical Center. Inclusion criteria for healthy control participants: age=20-83 years, no history of cerebrovascular disease, anemia, psychosis, or neurological disorder including but not limited to prior overt stroke, sickle cell anemia, schizophrenia, bipolar disorder, Alzheimer's disease, Parkinson's disease, or multiple sclerosis. Presence of non-speci c white matter lesions was not an exclusion criterion, as these lesions are extremely prevalent with aging, and we sought our cohort to be generalizable and representative. Clinical history was reviewed by a board-certi ed Neurologist (DOC; experience = 14 years) and anatomical imaging and angiography by a board-certi ed neuroradiologist (CDM; experience = 12 years) to ensure that inclusion criteria were met. Acquisition All participants were scanned at 3.0T (Philips Healthcare, Best, The Netherlands) using body coil radiofrequency transmission and SENSE array 32-channel reception. The scan protocol included standard anatomical imaging consisting of 3D T 1 -weighted MPRAGE (echo time=8.1ms, repetition time=3.7ms, ip angle=8°, resolution=1x1x1mm), 3D T 2 -weighted VISTA (echo time=0.31ms, repetition time=2.7ms, and spatial resolution=0.78x0.78x0.78mm), 2D T 2 -weighted FLAIR (echo time=120ms, repetition time=11,000ms, spatial resolution=1x1x4mm), 3D time-of-ight magnetic resonance angiography (echo time = 3.45ms, repetition time = 23ms, spatial resolution=0.39x0.39x1.4mm), and diffusion weighted imaging (DWI) (echo time=83 ms, repetition time=2923, b-value=1000s/mm 2 ; spatial resolution=1.8x1.8x4mm). These scans were primarily used for con rming healthy status and ensuring inclusion criteria; T 1 -weighted and T 2 -weighted scans were additionally used for brain parenchymal and PSD volume segmentation as described below.
CSF movement was recorded within the aqueduct of Sylvius. Here, four MR-compatible ECG electrodes were placed to enable retrospective cardiac phase correction. A single slice orthogonal to the aqueduct of Sylvius was placed above the location of the 4th ventricle, where the aqueduct is bound by the tectum posteriorly and mid-brain anteriorly. A velocity encoding gradient of 12 cm/s was applied along with 12 measurements over the cardiac cycle. Analysis CSF Flux: To estimate CSF ux in the aqueduct of Sylvius, we followed the acquisition protocol proposed in [12]. CSF ow parameters were quanti ed on the scanner console using the Q-Flow package provided by Philips Healthcare. The algorithm effectively subtracts phase contrast data acquired with the polarity of the bipolar phase contrast gradients reversed, which leads to computation of the following ow parameters: mean ux (ml/s), max anterograde ux (ml/s), max retrograde ux (ml/s).
Anatomical characteristics: All brain volumes were calculated using the T 1 -weighted acquisitions. First, intracranial volume (ICV) was estimated using the brain mask computed by the advanced normalization tools (ANTs) package [17], which utilized the MNI ICBM-152 version as a template [18]. CSF, gray matter (GM), and white matter (WM) volumes were calculated using the Atropos method [19].
Parasagittal dural space quanti cation: PSD volumes were computed using a semi-supervised segmentation method based on a combination of a fully connected neural network (F-CNN) and voxel clustering based on a gaussian mixture model (GMM) to label voxels as PSD or sagittal sinus based on their T 2 -weighted MRI signal intensities. The rst step of our method used an F-CNN to extract parasagittal space; this deep-learning model was trained using 20 T 2 -weighted MRI scans, which provides the best contrast to the adjacent subarachnoid space. This deep-learning model aims to estimate a binary mask of parasagittal space which includes both the superior sagittal sinus and contributing veins in the region of the PSD and the parasagittal dural space. Once the parasagittal space mask was computed, voxels belonging to the parasagittal mask were labeled as PSD, or superior sagittal sinus/contributing veins.
To achieve this, bias eld inhomogeneity was corrected using N4 inhomogeneity correction [20]. Next, in order to reduce anatomical variability, all T 2 -weighted MRI images were aligned to the MNI template [18] using non-linear registration computed with ANTs [17] with a control spacing point set to 2 mm. This value provides equipoise between the robustness of the registration (i.e., limitation of eventual registration artifacts) and increase of inter-subject similarity of the PSD. The PSD volumes are obtained using a semi-supervised machine learning method as described below.
First, a board-certi ed neuroradiologist (CDM; experience = 12 years) manually segmented the parasagittal space of 20 T 2 -weighted scans. This region of interest contains the superior sagittal sinus venous volume and PSD space volume along the sinus. Next, manual segmentation maps were used to train an automatic segmentation method based on an F-CNN using a U-Net architecture [21]. U-Net architecture was chosen for its good performance with medical image segmentation task and its ability to deal with limited training data size. In total, 180 overlapping patches were used to reconstruct the parasagittal space. Finally, a GMM was t, within the estimated parasagittal mask, using an expectationmaximization strategy on the T 2 -weighted MRI.
Thus, the nal maps provide labels for voxels belonging to the PSD space (hyperintense on the T 2weighted scans) and venous structures including the superior sagittal sinus and contributing veins (hypointense in the T 2 -weighted scans). Resulting segmentation maps were transformed to native space using the inverse transform. Therefore, all subsequent analyses are performed in the native space of the T 2 -weighted scans. All segmentation maps (PSD and tissue masks) were validated by a neuroradiologist (CDM; experience = 12 years).

Statistical analysis
To assess the validity of our hypotheses, we used a generalized linear model (GLM). We de ned three multivariate models to assess three different hypotheses.
1. Demographics and intracranial cavity volume (ICV): we de ned a model with PSD and max absolute ux (i.e., anteretrograde and retrograde) as separate dependent variables and age, sex, and intracranial volume as independent variables. 2. Demographics and tissue volumes (GM, WM, and CSF): we de ned a model with PSD and max CSF ux as separate dependent variables and age, sex, CSF, GM, and WM volumes as independent variables. 3. PSD space and bulk CSF Flux: we de ned a model with PSD volume as the dependent variable and max CSF ux, age, and sex as independent variables.
In addition of these linear models, we assessed the correlation of each pair of features using Spearman's rank correlation coe cient. This was tested using analysis of variance (ANOVA) and corrected for multiple comparison using false discovery rate (FDR) [22]. All reported p-values are reported as raw and corrected with FDR with signi cance xed to 0.05. All analyses were performed in Matlab (Mathworks) using the statistical toolbox.

Demographics
A summary of participant demographics is provided in Table 1. In total, 62 participants were included, with an age ranging from 20 to 83 years inclusive. All participants met neurological and radiological inclusion criteria as de ned in the Materials and Methods.  Table 1 and Supplementary  Table A1). Figure 1 shows an example of the PSD segmentation process and Figure 2 demonstrates the quanti cation process for CSF ux. Figures 3-4 summarize the relationships between tissue volumes, including PSD volume, and age, whereas Figure 5 summarizes the relationship between PSD volume and CSF ux. PSD volume was found to have a signi cant relationship with age: p<0.001 (q-values=0.009) and =0.59. Like PSD, bulk CSF ux (i.e., anteretrograde and retrograde) were also found to have a signi cant relationship with age: p=0.010, and p<0.001 (q-values=0.030, and 0.002) and =0.36, and -0.40 for PSD volume, anteretrograde, and retrograde ux, respectively. A signi cant sex relationship was observed with PSD volume (p-value<0.001). However, sex difference was not signi cant in relation to bulk CSF ux. These ndings are consistent with PSD volume being directly related to CSF ux through the aqueduct, as well as being elevated in males versus females.
The investigation of the relationship of PSD volume and CSF ux with ICV are detailed in Table A2 (see supplementary materials). We observed that ICV and PSD volume are not signi cantly related when sex is included as a covariate. Moreover, Spearman's correlation coe cient does not indicate any explanatory characteristic of ICV over the PSD volume ( =0.1 and 0.2, for male and female, respectively).
Relationship with brain tissue volumes After con rming the correlation of both proposed glymphatic markers (i.e., bulk CSF ux and PSD volume) with age and sex, we analyzed their relationships with brain tissue volumes. We observed no signi cant evidence of a relationship between brain tissue volume and either anteretrograde or retrograde CSF ux. Maximum retrograde CSF ux showed a signi cant relationship with CSF volume before controlling for multiple hypothesis testing with a p-value of 0.04, but this did not retain signi cance after FDR correction (q-value=0.08). Spearman's rank correlation method indicated a correlation of retrograde Using the same method of relationship, we assessed the relationship between PSD volume and brain tissue and CSF volumes (Figures 3-4). No evidence of a correlation between PSD enlargement and GM or WM volume was observed. PSD volume was observed to positively correlate with CSF volume with p-values=0.02 (q-value=0.04). In addition, the correlation analysis using Spearman's correlation coe cient indicated a strong correlation of PSD volume with a correlation coe cient =0.6 (p<0.001).

Parasagittal dura space and bulk CSF ux
Two linear models were applied to evaluate the relationship between PSD volume and CSF ux in the cerebral aqueduct. The results provide evidence for a signi cant correlation of PSD volume and CSF ux for both anteretrograde and retrograde directionality. The model indicates that anteretrograde CSF ux correlates with PSD volume with a p-value equal to 0.01 (q-value equal to 0.04), and a Spearman's correlation coe cient of 0.36 (p-value=0.001). Retrograde CSF ux show stronger correlation, with a p-value=0.001 (q-value=0.004), and a Spearman's correlation coe cient of -0.50 (p-value inferior to 0.001). Figure 6 shows representative cases of a young and older adult which demonstrate the observation of PSD volume increasing with age.

Discussion
Emerging evidence supportive of a human glymphatic system and its underlying role in cerebral waste clearance has prompted a reevaluation of neuro-uid circulation and relevance. It has recently been suggested that uid clearance may occur along lymphatic channels, which co-localize with the anatomic ρ ρ ρ region surrounding the superior sagittal sinus [7] and compelling evidence supports the presence of transarachnoid molecular passage in this region [23]. Our understanding of the anatomical and functional relevance of this region remains incomplete, partly due to a lack of robust methods for evaluating this space on neuroimaging, as well as how this space changes with age, sex, and standard measures of bulk CSF ow. Here, we provide evidence that the PSD volume increases with age, and also directly relates to CSF ux through the cerebral aqueduct.
As such, these ndings suggest that the PSD is an important component of the distal end of the glymphatic system whereby uid egresses from the intracranial compartment. We observed a signi cant enlargement of PSD space in relation to normal aging (p-value < 0.001, =0.6), CSF volume (p-value < 0.001, =0.6) and CSF ux in the cerebral aqueduct (retrograde and anteretrograde, p-values < 0.001, =0.32 and -50, respectively).
These ndings should also be considered in the context of the growing literature on PSD anatomy. The initial work demonstrating trans-arachnoid molecular passage in the PSD region was conducted by introducing gadolinium contrast into the subarachnoid space via lumbar puncture [23], with subsequent T 1 weighted imaging detailing contrast progression to the PSD. A subsequent study assessed the "perisinus lymphatic space volume" retrospectively assessing T 1 post contrast imaging in patients suspected of having brain metastasis [24]. In this study we use submillimeter 3D T 2 -weighted imaging of the brain to assess similar volumes. The technique we used allows for clear distinction of the PSD from both the adjacent subarachnoid space as well as the superior sagittal sinus and feeding cortical veins (see . Our delineation of this PSD volume very closely matches the 2D appearance and the 3D volumetric maps as detailed by Ringstad et al [23]. Importantly, this method obviates the need for contrast enhanced imaging to determine the PSD volume. Accordingly, this method can likely be applied to numerous noncontract imaging datasets of various patient cohorts in the public domain. We anticipate that the noncontrast nature of our method will accelerate future study of the PSD, which will in turn further de ne the relevance of this region in the context of glymphatic ow and CNS CSF clearance. Our results also con rm the previous ndings by Park et al that demonstrated increasing PSD volumes with age ( Figure 6). Our study expands upon this work to demonstrate that these increased PSD volumes are associated with increased CSF volume, though is not related to brain parenchymal volume. This observation is signi cant, as it implies that increased PSD cannot be explained simply by brain volume loss and therefore my provide insight into the pathophysiology of other neurodegenerative processes. Our study bene ts from a wider cohort age range compared to this previous work (mean age = 62.1, = 10.9 years, compared to our cohort with mean age = 50.4, = 18 years). We believe that the current cohort which ranged in age from 20 to 83 years enables us to detect evolution of PSD in a more comprehensive manner over the approximate adult human lifespan.
Cerebral aqueductal ux and parasagittal dural space volume ρ ρ ρ σ σ PSD volumes were signi cantly correlated to maximum anteretrograde and retrograde CSF ux in the cerebral aqueduct. The cerebral aqueduct and PSD are at the proximal and distal ends of the glymphatic circuit respectively. This correlation presents further evidence of an organized system of CSF circulation and metabolism and suggests of complex physiologic interplay between various anatomic structures.
Further study is needed across the human lifespan and in various disease cohorts to determine the sequence of dysfunction across the various regions of the glymphatic circuit. Investigations which characterize these ndings across differing neurodegenerative disease cohorts may sheds light on how aberrant CSF ow contributes to various neurodegenerative conditions.
Finally, we also assessed for a relationship between CSF ux in the cerebral aqueduct with white and gray matter volumes, CSF volume, gender, and age. Our experiments indicate that CSF volumes show signi cant correlation with maximum retrograde CSF ux in the cerebral aqueduct. This again supports that these ndings cannot be fully explained by brain volume loss.
In one prior study investigating the relationship of different CSF dynamics at the level of the cerebral aqueduct, CSF movement was shown to be dependent on age and gender [14]. Only part of these readouts can be con rmed by our experiments, even though we see correlation with age, sex difference was inconclusive in our analysis. However, it is noteworthy that maximum anteretrograde and retrograde CSF ux have not been directly investigated in this previous study, this could explain why we did not observe gender dependencies with CSF ux in our data.
The study ndings should also be considered in light of several limitations. First, we evaluated the PSD volume across the life-span cross-sectionally as is common in neuroimaging studies, rather than longitudinally. Second, while the largest cohort of PSD volume data presented to date, the sample size of 62 was moderate and presented multiple co-variates from being included in analysis. However, we characterized the health of each participant both radiologically and neurologically as described in the inclusion criteria, and all participants met rigorous healthy volunteer criteria. Given the moderate sample size, we also focused hypotheses on those that could be tested responsibly with the sample size.

Conclusions
Findings highlight the feasibility of quantifying PSD volume non-invasively in vivo in humans using machine learning and MRI, that PSD volume increases with age, and that PSD volume relates to bulk CSF volume and ux. Values reported should provide useful normative ranges for how PSD volume adjusts with age, which will serve as a necessary pre-requisite for comparisons to persons with neurodegenerative disorders. Findings also motivate the evolving hypothesis that aging compromises the e ciency of glymphatic circulation, leading to PSD hypertrophy. These processes may contribute to agerelated neurological changes and possible vulnerability to further pathological disruption of CSF clearance processes.
Declarations Figure 1 Parasagittal dura space (PSD) de nition using T2-weighted MRI. Three coronal slices are taken from the posterior aspect of the frontal lobe to the medial aspect of the parietal lobe. In this gure the PSD appears in green, the rest of the parasagittal space appears in red (i.e., sagittal sinus, and afferent veins).

Figure 2
Illustration of cerebral aqueduct ux acquisition. On the left, two curves representing the recorded CSF ux through the aqueduct of Sylvius over one arterial pulsatile cycle for a young adult (22 years old) and an elderly subject (71 years old). On the right, oblique axial and sagittal slices of T2-weighted MRI indicating the localization of the aqueduct of Sylvius. Due to high CSF ow through the aqueduct of Sylvius, the MR signal is dephased and appears hypointense on 3D T2-weighted MRI. Relationship of age with different anatomical characteristics for cerebrospinal uid (CSF), gray matter (GM), white matter (WM), and parasagittal dural (PSD) volumes. As expected, data show a signi cant correlation of GM and CSF volume with age. In addition, this study shows a novel signi cant relationship of normal aging with PSD volume, and CSF ux in the cerebral aqueduct. Non-signi cant trends appear in light gray shade; signi cant trends appear in dark gray shade.