Cerebrovascular activity is a major factor in the cerebrospinal fluid flow dynamics

Abstract

Cerebrospinal fluid (CSF) provides physical protection to the central nervous system as well as an essential homeostatic environment for the normal functioning of neurons. Additionally, it has been proposed that the pulsatile movement of CSF may assist in glymphatic clearance of brain metabolic waste products implicated in neurodegeneration. In awake humans, CSF flow dynamics are thought to be driven primarily by cerebral blood volume fluctuations resulting from a number of mechanisms, including a passive vascular response to blood pressure variations associated with cardiac and respiratory cycles. Recent research has shown that mechanisms that rely on the action of vascular smooth muscle cells (“cerebrovascular activity”) such as neuronal activity, changes in intravascular CO₂, and autonomic activation from the brainstem, may lead to CSF pulsations as well. Nevertheless, the relative contribution of these mechanisms to CSF flow remains unclear. To investigate this further, we developed an MRI approach capable of disentangling and quantifying CSF flow components of different time scales associated with these mechanisms. This approach was evaluated on human control subjects (n=12) performing intermittent voluntary deep inspirations by determining peak flow velocities and displaced volumes between these mechanisms in the fourth ventricle.

We found that peak flow velocities were similar between the different mechanisms, while displaced volumes per cycle were about a magnitude larger for deep inspirations. CSF flow velocity peaked at around 10.4 s (range 7.1-14.8 s, n=12) following deep inspiration, consistent with known cerebrovascular activation delays for this autonomic challenge. These findings point to an important role of cerebrovascular activity in the genesis of CSF pulsations. Other regulatory triggers for cerebral blood flow such as autonomic arousal and orthostatic challenges may create major CSF pulsatile movement as well. Future quantitative comparison of these and possibly additional types of CSF pulsations with the proposed approach may help clarify the conditions that affect CSF flow dynamics.

Introduction

Cerebrospinal fluid (CSF) assists normal functioning of the central nervous system by providing physical and immunological protection to the vulnerable neuronal tissue, as well as sustaining a suitable metabolic microenvironment (Spector et al., 2015). In particular, CSF movement has been demonstrated to aid clearance of metabolic waste products from the brain (Iliff et al., 2012; Nedergaard, 2013; Ringstad and Eide, 2020). Analogously, drug delivery via the intrathecal CSF route bypasses the Blood-Brain-Barrier and plays a unique role in development of novel therapies for cerebral diseases(Calias et al., 2014; Pardridge, 2020).

Waste clearance and drug delivery through the CSF may be mediated by both slow net flow and faster pulsatile flow (Hladky and Barrand, 2014). In adult humans, total CSF volume is about 150 mL, distributed over ventricular, perivascular, and sulcal spaces. Supported by slow net flow, CSF percolates through the brain interstitial space and glymphatic pathway at a rate of about 27 mL/h (Nilsson et al., 1992). The movement of CSF solutes through the glymphatic pathway may in part rely on fluid movement along the perivascular space (Iliff et al., 2012; Jessen et al., 2015). In addition to this net flow, there is pulsatile CSF flow with peak velocities orders of magnitude higher, thought to facilitate mixing of CSF and its constituents (Hladky and Barrand, 2014). Because of its potential role in waste clearance, knowledge about the pulsatile CSF flow and its internal drives may offer insight into their relationship to the pathophysiology of neurodegenerative diseases. Despite this great promise, CSF flow dynamics, especially in the human brain, are still poorly understood not only in the quantitative but also in the qualitative sense.

CSF flow associated with the cardiac cycle is well-established and has been characterized as a predominantly pulsatile movement of CSF by cardiac-gated phase-contrast MRI (Feinberg and Mark, 1987; Schroth and Klose, 1992). Simultaneous measurement of CSF and blood flow (both arterial and venous) has revealed that these CSF pulsations may be generated by total cerebral blood volume (CBV) changes secondary to the arterial pressure variations associated with the cardiac cycle (Alperin et al., 1996; Balédent et al., 2004; Martin et al., 2012). More recently, the importance of the respiratory cycle on CSF flow has been increasingly recognized (Yamada et al., 2013; Chen et al., 2015; Dreha-Kulaczewski et al., 2015; Yildiz et al., 2017; Takizawa et al., 2017; Dreha-Kulaczewski et al., 2017; Spijkerman et al., 2019). Thoracic pressure variations associated with respiration are thought to cause venous blood pressure changes, which like the cardiac pulsations are followed by total CBV and then CSF volume changes: inspiration has been found to lead to an influx of CSF into the cranium (cranial flow) and expiration to an outflow (caudal flow) (Yamada et al., 2013; Chen et al., 2015; Yildiz et al., 2017; Spijkerman et al., 2019). Since both cardiac- and respiratory- cycle mediated CSF pulsations originate from phasic blood pressure changes extrinsic to the cranium, we refer to them as barogenic pulsations.

Recent reports suggest that, in addition to these barogenic pulsations, other mechanisms may contribute to pulsatile CSF flow as well, specifically conditions that alter CBV by an active vascular response associated with electrocortical or sympathetic activity, or blood gas changes (Fultz et al., 2019; van Veluw et al., 2020; Yang et al., 2022; Picchioni et al., 2022). Since the episodic CBV changes are effectuated through action of cerebrovascular smooth muscle cells, we refer to these effects as “cerebrovascular activity” in general, and the associated CSF flow as myoactive pulsations. BOLD fMRI studies have demonstrated a tight relationship between widespread fMRI signal decreases and momentary CSF inflow to the brain via the fourth ventricle (Fultz et al., 2019; Yang et al., 2022; Picchioni et al., 2022). As with the barogenic pulsations, the myoactive CSF flow generated by intrinsic CBV changes can be explained by the Monro-Kellie doctrine of constant intracranial volume (Mokri, 2001): the myoactive CBV changes underlying the BOLD fMRI signal require compensatory intracranial CSF volume changes. Therefore, one expects a widespread in-/decrease in fMRI signal (and thus CBV) to be accompanied by an out-/inflow of CSF into the cranium, and movement of intracranial CSF.

As of yet, it remains unclear what the relative contribution of this myoactive effect is on overall CSF movement. While previous reports (Fultz et al., 2019; Yang et al., 2022; Picchioni et al., 2022) suggest that it may be large, quantitative assessment has not been performed due to the lack of a fast and robust CSF flow imaging technique capable of separating the different driving mechanisms of CSF movement. Existing MRI methods for the measurement of CSF flow include phase-contrast (Feinberg and Mark, 1987; Chen et al., 2015; Yildiz et al., 2017), inversion spin labeling (Yamada et al., 2013) and inflow-sensitized MRI (Dreha-Kulaczewski et al., 2015). Phase-contrast MRI (PC-MRI) has the advantage of whole brain coverage but has rather limited sensitivity to low flow velocities. Its application to CSF imaging requires a low velocity encoding value (VENC) of a few cm/s, leading to much longer flow encoding gradients and steeper gradient switching than in cardiac applications. Quantification of low-frequency CSF flow using real-time PC-MRI is also prone to confounding effects from spatial and temporal B₀ field changes (Peters et al., 2019). Inversion spin labeling and inflow-sensitized MRI have good sensitivity to low flow velocities but are lacking in the ability to provide accurate quantitative estimates.

Proper separation of different flow components requires a measurement method that covers the vastly different time scales at which the physiological mechanisms operate, ranging from fractions of a second to tens of seconds. To address this, we developed a previously reported qualitative MRI method based on balanced steady-state free precession (Oshio et al., 2019) into a quantitative approach to resolve the full range of time scales of CSF movement. We then studied human subjects taking isolated volitional deep inspirations, an effective way to generate myoaction through the effects of local CO₂ changes and sympathetic activation. The resulting myoactive CSF pulsations were then jointly quantified with common barogenic pulsations associated with cardiac and respiratory cycles, including flow velocities and displaced volume. The results indicate exceedingly large myoactive CSF displacement following deep inspirations, suggesting a major contribution of cerebrovascular activity to CSF dynamics.

Methods

Overview

Balanced steady-state free precession (SSFP, also known as True FISP/FIESTA) is a widely used MRI sequence especially in cardiac MRI, featured by its zero net (hence “balanced”) magnetic gradient moment within each TR (Carr, 1958; Scheffler and Lehnhardt, 2003). SSFP typically has a very short TR of a few ms (well below tissue T₂) and TE=TR/2, which in combination with the balanced gradients, make SSFP signal a superposition of a complex mix of signal generation pathways (Scheffler and Hennig, 2003; Weigel, 2015). The signal amplitude is frequency dependent with a repeating spectral pattern 1/TR Hz wide, each pattern alternating between “pass-band” and “transition-band”: When a moderate to large flip angle (e.g. 45 degrees) is used, SSFP manifests a unique T₂/T₁ contrast and insensitivity to frequency across the pass-band; Within the transition band, the signal amplitude rapidly drops to zero. The transition band is typically avoided in SSFP applications due to lack of tissue contrast, except for a few niche applications taking advantage of its strong frequency sensitivity (Scheffler et al., 2001; Miller, 2012; Wang et al., 2021). The alternating pattern of pass-bands and transition-bands can be observed in MRI by applying a small magnetic field gradient during SSFP imaging (Fig. 1A), essentially introducing a linear frequency offset along the applied gradient (Haacke et al., 1990; Scheffler et al., 2001; Markl and Pelc, 2004). Based on this imaging principle, Oshio et al. reported that with a sufficiently high temporal resolution, CSF movement with cardiac cycles can be qualitatively visualized near the dark transition-bands (Oshio et al., 2019). In this application, SSFP in combination with the background gradient offset is essentially a tagging method for motion detection (Axel and Dougherty, 1989; Yamada et al., 2008).

Figure 1.

Contributing mechanisms of CSF flow manifest distinctly on continuously acquired SSFP. A An example image frame with tag positions of interest labelled: Reference (REF), fourth ventricle, dorsal and ventral sides of the cervical spine C2 level. The tags were equidistantly distributed except near the nasal cavity due to the strong nonlinear magnetic gradient resulted from the tissue-air interface. B,C Physiological recordings and SSFP signal intensity along the line segments (shown on A) plotted with time. The matrix resolution was 1.3 mm by 246 ms. Blue and green plots correspond to chest respiratory belt and PPG cardiac signals, respectively. At resting state (B), cardiac cycles (green arrows) and normal respiratory cycles (light blue arrows) were evident. During periodic deep breaths (C), in addition to the instantaneous CSF flow (dark blue arrows), there were also sustaining myoactive CSF surges (orange arrows) emerging at a few seconds delay from each deep breath. CSF, cerebrospinal fluid. SSFP, (balanced) steady state free precession. PPG, photoplethysmography.

We applied this method to a group of healthy volunteers in a supine resting state and also when they performed cued deep breaths to observe changes in CSF flow. In our case, the background magnetic field gradient was set to ~0.3 mT/m for qualitative observation, leading to equidistant tags separated by ~15 mm. The SSFP images were acquired at 246 ms/frame, sufficiently fast to sample the cardiac pulsations and other slower flow mechanisms (Supplementary Video). During normal respiration, periodic tag displacement and deformation (and by inference CSF flow) can be observed in the CSF spaces around the brainstem and spinal cord at both the cardiac and respiratory frequencies (Fig. 1B). During periodic deep inspirations, tag changes increased (Fig. 1C). These observations are consistent with previous reports on CSF flow induced by arterial and venous blood pressure changes with cardiac and respiratory cycles (Dreha-Kulaczewski et al., 2015; Yildiz et al., 2017; Aktas et al., 2019). Importantly, in addition to these immediate barogenic effects, a delayed myoactive effect was also observed that evolved over a much longer time scale. Specifically, deep inspirations were followed by extended periods of upwards tag displacement, indicative of cranial CSF flow surges, with their delay and duration consistent with a myoactive mechanism (Picchioni et al., 2022).

Having identified this novel, myoactive type of CSF pulsations in the SSFP data, we set out to determine its strength relative to barogenic pulsations. This required translating the observed changes in the SSFP tags into flow velocities. Because of the complex relationship between tag changes and temporal flow profile, including differential sensitivities for rapidly (i.e. cardiac modulated) and slowly (i.e. respiratory modulated and myoactive) changing flow, we opted to use a lookup table (“flow dictionary”) for this purpose. To this end, we generated a two-dimensional flow dictionary based on the simulated, combined effects of both rapid (with a sinusoidal velocity change associated with the cardiac cycle), and slow (slower processes assumed constant over the cardiac cycle) changes in flow on SSFP tags (Fig. 2). The rapid and slow changes are called “HF flow” and “LF flow” respectively, for “high frequency” and “low frequency” components. Phase shifts of the flow relative to the cardiac cycle and changes in cardiac frequency were accommodated by generating additional dictionary entries (Fig. 2C). These dictionary entries were then matched through correlation to measured tags for each cardiac cycle, which are small data patches with dimensions of cross-tag spatial profile by time (11 pixels by 1 cardiac cycle). The latter are referred to as “perturbation patterns”.

Figure 2.

Overview of dictionary-based CSF flow quantification. A-C CSF flow simulation and dictionary generation. A Simulation results for HF peak velocity of 10 mm/s with background LF flow of 2 mm/s. Yellow box contours the converged result within a single cycle used for further analysis. B Juxtaposition of a subset of such patches generated from the simulation for further processing. C Phase cycling and downsampling of a single patch. 8-step phase shifts were applied to account for the time delay between the cardiac cycle and flow patterns. D Simultaneous acquisition of the chest belt signal, PPG signal and SSFP MRI in the central sagittal plane. The PPG peaks are denoted by red circles and used for temporal binning of the MRI data, generating CSF flow “perturbation patterns” explained below. E CSF flow quantification by dictionary matching. At each CSF location of interest, dictionary entries were matched through correlation to perturbation patterns. Each cardiac cycle (corresponding to a single perturbation pattern) is denoted by a red dot, evenly distributed in time after zero-padding (blue) to a fixed width of 5 time points. The corresponding HF peak velocity and LF velocity values of the best match were recorded and plotted. Corr. Coef., correlation coefficient.

We performed flow quantification at the fourth ventricle and near the foramen of Magendie, two lower brain positions with minimal spatial variation of B₀ field, despite the capability of perturbation patterns to qualitatively capture CSF dynamics at multiple locations in the lower brain and the neck (Fig. 1). This is because the spatially cyclic B₀ variation originating from cervical vertebrae was comparable (~60 Hz over 20 mm) to the background gradient offset at 3 T, leading to warping of the perturbation patterns (Supplementary Fig. S1). Without subject-specific B₀ correction of the images or dictionary, this might bias the flow quantification results.

Dictionary generation

Bloch simulations were performed to create dictionary entries for SSFP perturbation patterns. Magnetizations (isochromats) along the z direction were modeled as vectors $\stackrel{\rightharpoonup }{M}=[M_x,M_y,M_z{]}^T$, and their temporal evolutions were calculated by applying rotation and relaxation matrices that account for RF excitation, off-resonance effects and relaxation effects (Brown et al., 2014). The spatial and temporal discretization steps of the Bloch simulations were 0.15 mm and 1 ms, respectively. The CSF velocity was assumed to be pseudo-periodic with a function of $v(t)=HFsin\left(\frac{2\pi }{T}t\right)+LF$, with a constant T of 1 s close to the normal cardiac period. The same parameters in the MRI experiments were used in the Bloch simulation, namely TR 6 ms, TE 3 ms, flip angle 45°. T₁ and T₂ values of CSF were assumed to be 4 s (Rooney et al., 2007) and 2 s (Qin, 2011), respectively. A background gradient along the flow direction was applied, leading to equidistant tags separated by 23 mm. HF (cardiac) peaks were assumed to range from 0-12 mm/s (step size 0.2 mm/s) and LF flow from −8-8 mm/s (step size 0.2 mm/s). The simulation was considered to have converged when the change in flow pattern (magnitude of transverse magnetization ∣M_x + iM_y∣ versus time in a cardiac cycle) was below 0.5% in the sense of Euclidean distance, i.e. second norm of the difference.

For each converged pattern (Fig. 2B), 8 equidistant phase shifts across the cardiac cycle T were taken to generate additional dictionary entries. The results were downsampled to 3, 4, and 5 time points to match cardiac periods of around 750, 1000 and 1250 ms, and downsampled in space to match the image resolution of 1.3 mm (Fig. 2C). The total number of dictionary entries was 61×81×8×3 = 118,584.

Monte-Carlo validation

We performed a Monte-Carlo simulation to validate the proposed approach and benchmark the accuracy considering potential error sources originating from the following modeling approximations: 1) that every dictionary entry was generated using a fixed set of HF and LF values repeated over several cardiac cycles while actual CSF flow has ever-changing multi-frequency components, 2) that a fixed cardiac period of 1 s was used in dictionary generation while actual cardiac periods are not constant, 3) that all parameters associated with the dictionary were discrete while actual parameters are continuous.

We numerically generated 500 CSF data sets of about 300 seconds long each using Bloch simulation. Three sinusoidal components were included with random periods of (mean±std) 1.0±0.2, 6.0±1.0, 20±5 s and random amplitudes of 6±3, 1±1, 3±1 mm/s, corresponding to the cardiac, respiratory, and myoactive oscillations. A normally distributed random number generator (randn in MATLAB) was used to generate 300 cardiac, 50 respiratory, and 15 myogenic periods and velocity amplitudes for each data set, using the specified mean and standard deviation values. Negative values of cardiac and respiratory magnitudes were set to 0. Timing of cardiac sinusoid peaks was used for data segmentation into perturbation patterns, which were subsequently correlated to the flow dictionary. The entry with the maximum correlation coefficient was considered the best match.

SSFP image acquisition

MRI experiments were conducted on a 3 T Prisma scanner (Siemens Healthineers, Erlangen, Germany) using a 64-channel head-and-neck receive array. Parameters of the single-slice SSFP were TR 6 ms, TE 3 ms, flip angle 45°, slice thickness 3 mm, field of view (FOV) 240 mm × 204 mm, image matrix size 180 × 123 (resolution 1.3 mm × 1.7 mm), readout bandwidth 250 kHz, SENSE parallel imaging rate 3, frame rate 246 ms. After B₀ shimming, a shim offset in the z-direction (head-foot) was applied to create equidistant dark tags. The main resonance frequency was fine-tuned to position a tag in the middle of the fourth ventricle.

Flow phantom validation

A flow phantom was constructed which consisted of a horizontal U-shaped PVC pipe (21 mm inner diameter) supplied by an elevated reservoir for constant flow, and a faucet on the other end to control the flow rate. Tap water with T₁ = 2.4 s and T₂ = 1.5 s was used to create flow.

Single-slice SSFP images were acquired through the central axis of the pipe. The z shim offset that gave 45 mm tag distance (instead of 23 mm in in vivo experiments) was used for improved delineation of line profiles across the tags. Five increasing flow velocities were tested ranging from 0 to ~13 mm/s. EPI based PC-MRI on an axial slice was taken as ground truth. Only flow in the center of the pipe was analyzed to avoid the radial velocity change due to the laminar velocity profile.

In vivo experiments

All procedures followed a human subject research protocol approved by the NIH Institutional Review Board (IRB). Written informed consent was obtained from all participants.

MRI experiments were performed on 17 healthy volunteers (age 22-31 years, 6 males) in the supine position. Padding was used on the sides of the head to keep a consistent head pose and to minimize head motion. Subjects were asked to wear a face mask during the scan in compliance with the institutional regulation during the COVID-19 pandemic. The effect of this on experimental results was estimated to be negligible (see Supplementary Information). Two pneumatic respiratory belts were fastened at the top and bottom of the rib cage. PPG (photoplethysmography) signal was taken on the tips of both index fingers. Physiological signals were amplified and sampled at 2 kHz (Biopac, Goleta, CA, USA). Scanner triggers were also recorded for data synchronization.

Subjects were instructed to take slow deep breaths, visually guided by a progress bar (4 s breathe in and 4 s breathe out) spaced by 50 s, skipping every third cycle. In total, 6 deep breaths were taken per run of 450 s. Five subjects were excluded from final analysis due to excessive hand motion that corrupted PPG measurements, and/or incompliance to breath modulation tasks. Two more subjects were excluded from the analysis of BOLD EPI due to the same reasons.

During the breathing task, single-slice SSFP images were continuously acquired in the central sagittal plane, which yielded the largest cross-sectional area of the fourth ventricle and the CSF canal. Two z-shim amplitudes were used to create tag-distances of approximately 15 mm and 23 mm respectively, both resulting in clear tags in the fourth ventricle, at the cranial-cervical junction, and C2-C3 cervical level. The data with 23 mm tag distance were used for quantitative analysis due to robustness to large displacements such as those associated with the myoactive flow. The data with 15 mm tag distance were used for the purpose of qualitative visualization (Fig. 1).

For BOLD fMRI, 25 axial slices parallel to the AC-PC line were acquired based on single-shot EPI. Imaging parameters: TR 2 s, TE 35 ms, FA 70°, slice thickness 2.5 mm, FOV 240 mm × 180 mm, isotropic in-plane resolution of 1.5 mm, SENSE rate 2, bandwidth 250 kHz. The same deep breathing paradigm was used as for SSFP.

3D T₁-MPRAGE was acquired at 0.8 × 0.8 × 1.0 mm³ resolution as an anatomical reference and to estimate the cross-sectional area of CSF spaces.

Flow estimation

Imaging data were reformatted based on the cardiac period as derived from the PPG peaks (3 image frames taken if below 875 ms, 4 frames if between 875-1125 ms, and 5 frames if above 1125 ms). HF flow peak velocity and instantaneous LF flow velocity were estimated for each cardiac cycle by dictionary matching, in which the dictionary entry with the maximum correlation coefficient was considered as the best match. Temporal LF peaks for the direct (barogenic) and delayed (myoactive) effects of deep inspiration were determined in intervals of [−3 s, −1 s] and [6 s, 16 s] around the peak inspiration, respectively. The 50 s resting state data sections without deep breaths were used for estimation of CSF flow corresponding to the cardiac (by averaging HF results) and respiratory (by regressing the LF results with the chest belt signal and its time derivative) cycle. Prior to the regression, respiratory signal and LF results were low-pass filtered up to 1 Hz with preserved phase.

To estimate average velocities considering the spatial profile of the flow, measurements from the central voxel were corrected by a factor of 0.5 for the fourth ventricle, assuming a laminar flow profile. Velocities from the foramen of Magendie were not corrected because of the small width compared to the voxel size, rendering spatially averaged velocities from the measurements directly.

For estimation of displaced volume in the fourth ventricle, the average velocity was multiplied by the cross-sectional area as determined from T₁-MPRAGE, and integrated assuming a sinusoidal temporal velocity profile with periods determined from the physiological measurements (0.95±0.03 s and 3.82±0.18 s for heartbeat and respiration respectively, mean ± SE). For sections with deep breaths, direct and delayed effects were assumed to have a halfsinusoid shape with durations of 4 and 10 s respectively. Displaced volume at the foramen of Magendie was not calculated due to the small size and the irregular shape that compromised the accuracy of the measurement.

BOLD EPI processing

Preprocessing of EPI fMRI data was performed with AFNI (analysis of functional neuroimages) (Cox, 1996), including removal of the initial 2 TRs, despiking, slice timing correction, motion correction by image alignment (6-parameters rigid-body), warping to the Talairach space, spatial blurring (3 mm FWHM, full width at half maximum), regressing out motion parameters and their first derivatives, demeaning and linear detrending, signal scaling. The result of the preprocessing was signal percentage change over the run. The averages of the signals in the gray matter and white matter were taken over the corresponding eroded masks generated by 3dSeg in AFNI. The average CSF inflow signal was extracted from a manually drawn ROI in the fourth ventricle on the bottom 4~6 slices of the unblurred data (Fultz et al., 2019).

Temporal interpolation of signals

Respiratory signals, HF and LF results from SSFP, as well as CSF/GM/WM results from EPI were cubically interpolated onto the same time axis with a 100 ms resolution prior to further analysis. The mean LF over each run was set to 0, assuming negligible net CSF flow on the time scale of a few minutes.

Results

The Monte-Carlo simulation results demonstrated reasonable matches between the continuous CSF flow and dictionary entries. The retrieved flow parameters were in general follow the ground truth values (Fig. 3). Over the simulated 145,247 cardiac cycles, the mean ± SE of the errors in HF were 0.120±0.006 mm/s; errors in LF were 0.048±0.004 mm/s; correlation coefficients were 0.9781±0.001. The errors were sufficiently small compared to the range of HF (12 mm/s) and that of LF (16 mm/s).

Figure 3.

Monte-Carlo simulation results from one of the 500 trials. Blue plots show the ground truth parameters corresponding to each cardiac cycle, i.e. numbers used to simulate the 3-frequency oscillating flow. Orange shows the dictionary matching results. The pattern width was either 3, 4 or 5 for cardiac periods closest to 0.75, 1.00, 1.25 s, respectively. The black plot is the correlation coefficient between the perturbation patterns and their best match. Data (perturbation patterns), their best match and the difference are show in a stacked manner. A section is zoomed in for improved visualization.

LF flow caused movement and shape change of the tags that scaled with velocity as shown in the flow phantom results (Fig. 4). Dictionary matching yielded accurate LF results in the range of 0-10 mm/s (Table 1). The best matches closely followed the profiles of the flow affected tags (Fig. 4B). The latter is confirmed by the high (>0.95) correlation coefficients reported in Table 1. These results suggest that the simulations faithfully capture the dynamics of tag changes with LF flow, and the dictionary approach is of sufficient accuracy for LF flow slower than 10 mm/s. However, LF flow faster than 10 mm/s resulted in substantial smearing of the tags that reduces the velocity sensitivity and compromises the quantification accuracy.

Figure 4.

Flow phantom validation result. A SSFP images of constant flow from five runs with increasing velocities. The tag distance was 45 mm. B SSFP signal profiles in the center of the tube (solid) and the best dictionary matches (dashed). Legend shows the flow velocity measured using PC-MRI, also in the center of the tube.

Table 1.

Comparison of velocity results in the flow phantom by PC-MRI and SSFP.

Test	Velocity from PC-MRI (mm/s)	Velocity from SSFP (mm/s)	Absolute Difference (mm/s)	Correlation coefficient of matching
1	0	0.0	0.0	0.97
2	1.72	2.0	0.3	0.99
3	3.39	3.2	0.2	0.99
4	6.00	6.0	0.0	0.99
5	12.83	9.6	3.2	0.99

An example of in vivo flow quantification in the fourth ventricle is shown in Fig. 5. HF and LF flows were successfully separated, with the LF flow fluctuating mainly at the respiratory and lower frequencies (see Fig. S2 for regression results), and the HF flow peaks fluctuating at higher frequencies. Myoactive flow can be observed on both perturbation patterns and the LF flow plot after deep inspiration that dwarfed the respiratory flow. The strong flow led to CSF movement reaching 10 mm, mixing CSF along the canal more efficiently.

Figure 5.

Flow quantification results associated with deep breaths at an example location of interest. A T₁ weighted anatomical reference and SSFP of the same sagittal slice. The voxel of interest was in the fourth ventricle. B Chest belt signal, perturbation patterns, and flow detection results. LF flow displacement is plotted in purple, calculated by integrating the LF flow velocity. Horizontal dashed lines denote zero on the LF plots. Vertical dashed lines denote the end of deep inspiration.

In group analysis results, myoactive effects peaked at 10.4 s (ranging 7.1-14.8 s across subjects) after deep inhalation in the fourth ventricle, and similarly near the foramen of Magendie (Fig. 6B). The timing is consistent with the BOLD fMRI data acquired as reference in this study (Fig. 6C), as well as our previous BOLD fMRI study of myoactive CSF inflow (Picchioni et al., 2022). In addition, the timing similarity between the myoactive CSF surge and the negative time derivative of BOLD in the brain (a surrogate for CBV reduction) (Fig. 6C) confirms the myoactive origin of the flow, consistent with previous reports (Fultz et al., 2019; Yang et al., 2022; Picchioni et al., 2022). In the fourth ventricle, peak velocity of the averaged myoactive CSF flow was 0.73 ± 0.17 mm/s by SSFP. Peak BOLD signal percentage changes in the gray matter, white matter and that due to CSF inflow were −0.97±0.09 % (delay 13.1 s, purple dashed line in Fig. 6C), −0.40±0.03 % (delay 14.9 s, green dashed line), 5.7±1.5 % (delay 11.0 s, maroon solid line), respectively.

Figure 6.

Barogenic and myoactive origins of CSF pulsations with deep inspiration. A T₁-weighted anatomical reference and an example SSFP frame labelled with CSF locations of interest. B CSF velocities in the fourth ventricle and near the foramen of Magendie measured by the proposed approach. Results are in mean ± standard error of the mean over subjects. C BOLD signal percent changes in gray matter (GM) and white matter (WM), their negative time derivatives, and CSF cranial inflow signal in the fourth ventricle, acquired using the same deep inspiration paradigm. CSF flow peaked at 11 s delay and 2-3 s later than the peak of negative time derivative of BOLD, a surrogate for cerebral blood volume reduction.

We summarize the average velocities (with correction for the laminar flow profile) and displaced volume for each type of pulsations in Table 1. Peak flow velocities for the delayed myoactive effect after deep inspiration fall short for those associated with cardiac pulsations but exceed those for the direct barogenic effects of deep inspirations. Notably, displaced volume with the myoactive effect far exceeds those for the other mechanisms, suggesting this mechanism may have particular relevance for mixing and possibly transport in the CSF compartment. Displaced volume was similar for cardiac and respiratory pulsations, consistent with a previous report on the large contribution of respiratory effects on CSF dynamics within the cranium (Dreha-Kulaczewski et al., 2015).

Discussion

Using a novel quantification method for the complex fluid flow in major CSF access canals to the human brain, we established a dominant myoactive component associated with cerebrovascular activity. The myoactive flow of CSF was elicited with a deep inspiration and the largest velocity effect was found to occur with a delay of 10.4 seconds, consistent with a vasoactive response. Displaced CSF volume with this novel, myoactive mechanism was found to be about an order of magnitude larger than that with barogenic pulsations associated with the cardiac and respiratory cycles, mechanisms that were up to now assumed to be the main drivers of CSF pulsations in the human brain (Chen et al., 2015; Dreha-Kulaczewski et al., 2015; Yildiz et al., 2017).

Deep breaths induce brief vasoconstriction (and an inflow of CSF into the cranium) predominantly through a cascade of events triggered by momentary reductions in intravascular CO₂, with the constrictive action effectuated in brain parenchyma locally by vascular smooth muscle cells (Brian, 1998; Willie et al., 2014). Additional vasoconstriction may be effectuated by sympathetic activity through autonomic nerves on extraparenchymal arteries (Hamel, 2006), possibly resulting from O₂ and CO₂ chemo-sensing involving the brainstem (Guyenet, 2014). Through these mechanisms, this delayed, myoactive vasoconstriction causes strong changes in CBV that is detectable by BOLD fMRI, with a peak effect at a delay time of 12~15 s (Birn et al., 2008; Bright et al., 2009; Lynch et al., 2020). While deep inspirations are less frequent than the cardiac and respiratory pulsations, the slowly varying nature of the vasoconstriction causes a sustained intracranial pressure gradient that allows momentum for CSF movement to build up, resulting in a large and lasting flow surge.

While the current study focuses on the generation of CSF pulsations by having subjects taking infrequent deep breaths, myoactive pulsations are not limited to the CBF regulatory response elicited by deep breaths. There are various other mechanisms that may generate vascular smooth muscle activity, including cortical and subcortical neural activity (through neurovascular control), and spontaneous fluctuations in sympathetic tone effectuated by the brainstem (Özbay et al., 2018, 2019). In fact, recent reports suggest a possible link between CSF pulsations and cerebral neural activity (Fultz et al., 2019; van Veluw et al., 2020), as well as sympathetic tone (Attarpour et al., 2021; Picchioni et al., 2022). A thorough understanding of the changes of these vascular regulatory origins across brain states (including working, resting awake, and sleeping), and their relative contributions to the myoactive CSF flow, remains to be established.

In human fMRI, spontaneous global fluctuations in BOLD are a ubiquitous phenomenon, and typically occur at a rate of 0.1 Hz or below (Daouk et al., 2017; Fultz et al., 2019; Whittaker et al., 2019; Attarpour et al., 2021). They are a major part of the so-called “global signal” (Liu et al., 2017), a phenomenon that is still poorly understood and to which contributions are incompletely charted. Recent fMRI studies show that the extent by which this global signal contributes to CSF dynamics is reduced in neurodegenerative pathologies such as Alzheimer’s Disease (Han et al., 2021b) and Parkinson’s Disease (Han et al., 2021a). The changed interplay may be driven by modified CSF pulsation patterns across the brain in the patient populations (Kiviniemi et al., 2016; Tuovinen et al., 2020; Rajna et al., 2021). While much mechanistic work is still needed, these initial reports provide early, circumstantial evidence that these pulsations may have relevance for brain health, possibly through supporting waste clearance.

Tracer clearance studies in mice (Xie et al., 2013; Hablitz et al., 2020) and man (Eide et al., 2021b) have shown that glymphatic clearance may preferentially occur during sleep and sleep-like conditions. Increased clearance under these conditions may in part relate to increases in interstitial spaces (Xie et al., 2013) and perivascular polarization of cell membrane water channel aquaporin-4 (Hablitz et al., 2020). Another possibility is that sleep facilitates CSF flow through myoactive mechanisms. BOLD fMRI global signal fluctuations have been found to increase during sleep, especially in the low-frequency range (<0.1 Hz) (Fukunaga et al., 2006; Horovitz et al., 2008; McAvoy et al., 2019; Soon et al., 2021), in part because of increases in cerebral regulatory variability (Özbay et al., 2019, 2018; Picchioni et al., 2022) and in part because of increased variation in neuro-electrical activity (Scholvinck et al., 2010). To understand the relevance of these global signal fluctuations to CSF movement, flow quantification with the methodology proposed here would be vital.

The inter- and intra- subject variation in the estimate of flow velocities by respiration and cerebrovascular activity may be partly attributed to differences in ventilation volume or respiratory effort. To control for this variability, a flow gauge would be necessary, in place of the respiration belt used in the current study as a semi-quantitative measure of respiratory effort. The subjects enrolled in the study were healthy young adults aging 22-31, and all respiration signals were visually inspected to ensure the quality of deep breaths. Therefore, the average tidal volume and inspiratory reserve volume among the healthy population should apply. The former is 500 mL for both men and women, and the latter is 3300 mL for men and 1900 mL for women (Tortora and Derrickson, 2016).

In addition to revealing the myoactive flow, we were also able to quantify cardiorespiratory effects on flow at the fourth ventricle, yielding similar results to a previous study using fast PC-MRI (Chen et al., 2015). We confirmed that cardiac modulated velocity was higher than respiratory modulated velocity in the fourth ventricle, regardless of the respiration depth. Nevertheless, displacement volume per cycle was similar (−0.05 mL/cycle) for both mechanisms.

We demonstrate that SSFP in combination with a proper background magnetic field gradient is capable of qualitatively detecting CSF dynamics of various time scales associated with different physiological mechanisms. The proposed SSFP quantification method yielded consistent results within the cranium but would need B₀ field maps to account for tag scaling effects for CSF quantification in the neck. When fully developed, the proposed approach may become a useful tool to image CSF flow abnormalities, and ultimately to understand if and how these abnormalities may play a role in a myriad of neurodegenerative diseases. Structurally abnormal CSF cavities in pathology, such as hydrocephalus (Balédent et al., 2004; Bradley, 2015; Eide et al., 2021a) and Chiari malformation (Bunck et al., 2012), are typically accompanied by changes in flow magnitude and direction. Previous research focused on the changes in the cardiac related flow, which may be insufficient to fully characterize the CSF dynamics. Moreover, with the discovery and characterization of the glymphatic system in mice, there has been increasing efforts to identify its relevance in neurodegenerative diseases in humans. Imaging of the changed CSF flow patterns in patients, such as Alzheimer’s population (Tuovinen et al., 2020; Han et al., 2021b; Rajna et al., 2021), may shed light on the disease mechanism and facilitate its treatment or prevention. Finally, knowledge about the spatial and circadian changes of CSF dynamics along the entire CSF canal may lead to improvement in therapeutical outcome of intrathecal drug delivery (Dreha-Kulaczewski et al., 2018).

Conclusions

Combination of SSFP with a background magnetic field gradient can be used to create and monitor spatial tags in CSF compartments. The movement and deformation of these tags can be used to quantify pulsatile flow resulting from various mechanisms and occurring over a range of time scales. The mechanisms include previously described pulsations generated by cardiac and respiratory cycles, as well as a recently discovered slower mechanism resulting from cerebrovascular activity. The spatial-temporal “perturbation patterns” of the tags can be explained by simulations and translated to CSF velocities via a dictionary approach. Using this approach, we found CSF displaced volume by cerebrovascular activity (“myogenic”) effects to be significantly larger than the well-established cardiorespiratory pressure (“barogenic”) effects in healthy human subjects performing intermittent deep breaths. The myogenic CSF flow peaked at a time delay of 10.4 s from the end of deep inspiration, consistent with previous observations based on BOLD fMRI, suggesting its origin of CBV changes. This finding suggests that cerebrovascular activity may have an outsized role in the generation of CSF pulsations.

Table 2.

Average peak velocities and displaced volume corrected for the laminar flow profile (mean ± standard error; n = 12).

Location	Cross Section Area (mm2)	Average Peak Velocities (mm/s)				Displaced Volume (mL)
		Barogenic			Myoactive	Barogenic			Myoactive
		Cardiac	Normal Respiration	Deep Breath (direct)	Deep Breath (delayed)	Cardiac	Normal Respiration	Deep Breath (direct)	Deep Breath (delayed)
Fourth Ventricle	100±8	1.41±0.18	0.08±0.03	0.24±0.08	0.57±0.07	0.04±0.01	0.01±0.01	0.06±0.02	0.36±0.05
Foramen of Magendie	-	5.03±0.76	0.32±0.08	0.66±0.10	1.22±0.25	-	-	-	-

Highlights.

• SSFP tagging allowed visualization of CSF flow of various amplitude and time scales

• CSF flow by cardiac, respiratory and cerebrovascular activities was quantified using a dictionary method

• Cerebrovascular activity is a major contributor to pulsatile CSF flow

• The vasoactive CSF flow peaked at a 10.4 s delay from the end of deep inspiration

• BOLD fMRI corroborated the vasoactive nature of the delayed flow