Cerebrospinal Fluid Flow within Ventricles and Subarachnoid Space Evaluated by Velocity Selective Spin Labeling MRI

Participants

A total of 14 healthy volunteers (age: 39±17 years; 6 females, 8 males) participated, and the specific number of participants for each study is detailed below. Scanning procedures were conducted on a Philips MR Ingenia Elition 3.0T scanner (Philips Healthcare, Best, The Netherlands), utilizing a quadrature body transmit coil and a 32-channel receive head coil. Ethical approval was obtained from the Johns Hopkins Medicine Institutional Review Board (IRB), and all participants have provided their informed consent.

VSSL sequence

VSSL pulse train was used in this study, consisting of 9 excitation pulses (20° hard pulses), interleaved with pairs of refocusing pulses (180° composite pulses) and triangular gradient lobes of alternating polarity with 8 velocity-encoding steps (Figure 1A) (31). The control modules used a velocity-compensated gradient configuration for a more balanced diffusion-weighting effect ( b_label = 3.57 s ∕ mm², b_control = 1.36 s ∕ mm²)(53). Images were then acquired following a post-labeling delay (PLD) of 10 ms to minimize the possible exchange of labeled blood water with CSF.

• Download figure
• Open in new tab

Figure 1.

Illustration of the Fourier-transform-based velocity selective inversion pulse train, VSSL sequence, and the VSSL magnetization as a function of mean velocity. (A) Diagram illustrating the Fourier-transform-based velocity selective inversion pulse train, where velocity-selective labeling is employed to mark the flow within CSF. In each velocity-encoding step, a 20° excitation pulse is followed by a pair of refocusing pulses. Gradients with alternating polarity surround the refocusing pulses for the velocity-sensitive waveform (label), and uni-polar gradients are used for the velocity-compensated waveform (control) in the velocity-encoding steps. The segment time (T_seg) and maximum gradient (G_max) for the flow encoding are labeled in the sequence. (B) Diagram illustrating the VSSL sequence. After the velocity selective inversion pulse train, a post-labeling delay of 10 ms precedes the readout. A 3D-GRASE readout with a long TE is applied to image CSF while attenuating signals from parenchyma and blood. (C) Simulated Mz-velocity responses for the VSSL pulse train. Velocity-sensitive and velocity-compensated profiles are represented by solid and dashed blue lines, respectively. The cut-off velocity (V_cut) is delineated at the first intersection where ΔM = 1, highlighted by vertical black dashed lines at a velocity of 0.29 cm/s. Parameters used in the simulation, such as number of segments (8), T_seg (30 ms), and G_max (40 mT/m), are consistent with those employed in our study.

A 3D-gradient and spin echo (3D-GRASE) readout with linear order was used for image acquisition. The MRI signals from the parenchyma and blood are attenuated by a long TE, taking advantage of their much shorter T₂ relaxation times (<100 ms) than that of CSF (>1000 ms) (54). The imaging parameters were as follows: field of view = 220×165×160 mm³; slice number 80; acquisition matrix 112×82; reconstruction matrix size 224×224; acquisition resolution = 2×2×2 mm³; turbo spin echo (TSE) factor = 41 with linear ordering; echo-planar imaging (EPI) factor = 41; SENSE factor (S direction) = 2; echo spacing = 40 ms; TR/effective TE = 10 s/816 ms. An alternative 3D-GRASE readout changing to centric order and effective TE = 39 ms was also evaluated for comparison. We used two-sample t-tests to test whether there is a significant difference between the signal acquired from TE = 816 and 39 ms.

The control and label experiments were repeated six times for averaging, with a total scan duration of 5 minutes for each of the three orthogonal velocity-encoding directions (S-I: superior-inferior; A-P: anterior-posterior; L-R: left-right). M0 image was acquired without VSSL labeling and a TR of 20 s to ensure full recovery of CSF magnetization, other parameters are identical to the VSSL scan.

Simulation

Simulations were performed using MATLAB 2023b (MathWorks, Natick, MA, USA). VSSL module was evaluated through Bloch simulations, including 240-ms duration with eight 30-ms segments, 16 refocusing pulses (1.74 ms) with MLEV-16 phase cycling, 0.6 ms gradient lobe duration with 0.3 ms ramp time, and maximal gradient strength of 40 mT/m, as illustrated in Figure 1A. The cutoff velocity (V_cut) as defined in the VSASL guideline paper (55) was set at 0.29 cm/s. Mz values from label and control were simulated across velocities ranging from −2.0 cm/s to 2.0 cm/s at 0.01 cm/s intervals. In the simulation, a laminar flow model was assumed for simplification. T₁ and T₂ effects were not included in this simulation.

VSSL Optimization for CSF flow

Given that the T₁ and T₂ values of CSF significantly differ from those of brain tissue and blood, the parameters used in MRA and CBF studies are unsuitable for CSF flow mapping. Utilizing a longer T_seg reduces the cutoff velocity, which is beneficial for detecting slow CSF flow. However, a longer T_seg also prolongs the total duration of the VSSL module, potentially diminishing the overall signal due to T₂ decay. In regions with minimal flow motion, T₂ relaxation becomes the dominant factor, while the VSSL will increase in regions with a wide range of flow rates. Therefore, it is crucial to optimize T_seg to maximize the CSF signal, while G_max was set to the maximum value of 40 mT/m for the current scanner. Three volunteers (30∼56 years old; 2 females, 1 male) have undergone the study of optimizing the VSSL module for CSF imaging. To optimize the VSSL pulse train targeting the CSF flow around arteries, and flow in PVS and ventricles, we varied the duration of each velocity-encoding segment (T_seg). We fixed gradient strength (G_max) of 40 mT/m, ramp time of 0.3 ms, and a lobe of 0.6 ms, while adjusting T_seg as 10, 20, 30, 40, and 50 ms (V_cut = 0.86, 0.43, 0.29, 0.21, and 0.17 cm/s, VSSL module duration = 80, 160, 240, 320, 400 ms, respectively).

PVSAS and ventricle assignment

To identify the signal from PVS and ventricles or the CSF flow around the arteries in the SAS, images acquired with VSSL preparation were overlapped with velocity-selective MRA and T₁ weighted (T_1w) anatomical images from MPRAGE. MRA was obtained with a Fourier-transform-based velocity selective saturation module followed by a gradient-echo (GRE) readout (56). The VSSL pulse train applied 9 excitation pulses (10° hard pulses), T_seg = 10 ms, G_max = 30 mT/m, gradient lobes of 0.6 ms with ramp time of 0.2 ms. MRA scan used identical geometry as those VSSL scans including the field of view, resolution, and orientation. For the GRE readout, resolution = 1×1×1 mm³, TR/TE = 9.1/1.8 ms, flip angle = 8°, TFE factor = 90 with a centric ordering, and compressed sensing factor = 8 was used. The MPRAGE sequence parameters were as follows: field of view = 220×200×200 mm³; resolution = 1×1×1 mm³; TR/TE = 7.9/3.7 ms; TFE factor = 128 with a linear ordering; flip angle = 8°, SENSE factor (P direction) = 2. The total scan duration for MRA and MPRAGE were 1 min and 3.5 min, respectively.

Calibration of the CSF flow measured by VSSL

Eleven volunteers (22∼79 years old; 4 females, 7 males) were involved in the calibration study. As demonstrated in Figure 1C, the VSSL signal intensity can be approximated as a linear function of the flow velocity when the flow is slow. Consequently, we calibrated the VSSL-derived velocity maps to actual flow velocity using PC-MRI. The sagittal plane that covers the CA, 4V, and SC was localized using a sagittal MPRAGE scan. Retrospective cardiac-gated 2D phase contrast MRI scans were performed using a 2D single-slice GRE sequence with bipolar velocity-encoding gradients. Each experimental set comprised phase contrast data acquisition at flip angles of 20°, with the remaining MR parameters being a field of view of 220×160 mm², an in-plane resolution of 0.5×0.5 mm², a slice thickness of 4 mm, TR/TE of 30/20 ms. The velocity encoding (VENC) was set at 10 cm/s across 16 cardiac phases with the flow direction of S-I and at 3 cm/s with the flow direction of A-P and L-R. Physiological waveform monitoring was conducted using a Philips Peripheral Pulse Unit. To correct for background phase offset, all phase images from a cardiac cycle were first averaged to produce a mean phase image. The mean phase image was then processed using a median filter with a 50×50 kernel size. Finally, the filtered mean phase image was subtracted from the phase image of each cardiac phase. (57)

Data analysis

All analyses were processed using custom-written MATLAB scripts. The VSSL images were obtained by subtracting the control and labeled images in a pairwise fashion using a complex subtraction, and subsequently normalized by the M0 signal: S_control, S_label, and S_M0 represent the control, label, and M0 images in complex value, respectively. The Mag[⋅] operator takes the magnitude of a complex value. l_S–l, l_A–P, and l_L–R represent the normalized VSSL signal acquired with S-I, A-P, and L-R velocity-encoding direction, respectively. ROIs representing CSF within interpeduncular cistern (IPC), prepontine cistern (PPC), and spinal canal (SC) regions were manually delineated on the PC-MRI images to determine flow rates. The VSSL-derived values, extracted using the same ROIs, were correlated with flow velocities obtained from PC-MRI. As VSSL measured the average CSF velocity, the average absolute velocity measured by PC-MRI in one cardiac cycle was used for the calibration.

To access the CSF movement around the arteries in the SAS, we depicted the surroundings of the cerebral arteries shown on MRA. We chose three arteries for the SAS analysis, including M1 and M2 segments of the middle cerebral artery (MCA), A3 segment of the anterior cerebral artery (ACA), P1 and P2 segments of the posterior cerebral artery (PCA). The ventricular ROIs were manually delineated on individual participants’ M0 images, including the third ventricle (3V), fourth ventricle (4V), and frontal horn of the lateral ventricle (FLV), cerebral aqueduct (CA), and foramen of Monro (FMo). IPC and SC were also included in our analysis.

To determine whether there are statistically significant differences among the CSF velocities of three orthogonal directions (S-I, A-P, and L-R), we performed a one-way analysis of variance (ANOVA). Following the ANOVA, where a significant F-statistic indicated differences among group means, post-hoc pairwise comparisons were conducted using Tukey’s Honest Significant Difference (HSD) test. We considered a p-value of less than 0.05 to be statistically significant.

Simulation

Figure 1C illustrates the simulated Mz-velocity responses for the VSSL pulse train with laminar flow integration. At a flow velocity of 0 cm/s (stationary spins), the magnetization is fully inverted by the VSSL pulse train, whereas at flow velocities exceeding 0.29 cm/s, magnetization is partially attenuated (0 to 0.8 of Mz). Below 0.8 cm/s, the Mz displays a nearly linear relationship with the flow velocity. However, CSF exhibits complex flow patterns in the brain, characterized by a mixture of various orientations and flow velocities, making it challenging to simulate. The simulated Mz-velocity responses for the VSSL pulse train with varying segment times are shown in Supplementary Figure S1 for comparison.

Optimization of T_seg

Figure 2 illustrates the T_seg optimization of the VSSL pulse train for the CSF flow images. Figure 2A exhibits the VSSL images from a typical subject with varied T_seg ranging from 10 ms to 50 ms. The image quality shows an obvious improvement when T_seg is increased from 10 ms to 30 ms, as evidenced by the highly reduced background with longer T_seg. The background signal may partially originate from the PVS, but this requires further validation. Since the PVS signal is much weaker than the CSF signal in the ventricles and SAS, it is challenging to study and will be addressed in future research. The VSSL signal decays slowly as a function of T_seg in most regions. Figure 2B shows the normalized VSSL signal around the MCA, ACA, PCA, and ventricular regions including 4V, CA, and FLV. The signal around MCA and ACA remains relatively stable across T_seg. The signal around PCA, 4V, and CA decreases with T_seg. Notably, the signal in the FLV, considered as “free-water” with minimal motion (58), decreases from T_seg = 10 ms to 30 ms but stabilizes from T_seg = 30 ms to 50 ms. The signal in FLV can be used to assess the contribution of the diffusion component. With T_/seg = 30 ms, the normalized VSSL signal in FLV is only 0.006, which is far less than (<25%) the normalized VSSL signal around MCA (0.050), ACA (0.028), and PCA (0.070). Thus, T_seg = 30 ms was selected for further analysis due to better image quality, low diffusion-weighted effect, and relatively high VSSL signal.

• Download figure
• Open in new tab

Figure 2.

T_seg Optimization. (A) Example VSSL images from one subject acquired with different T_seg in sagittal and axial views. The left images, overlapped with MRA and MPRAGE images, show the ROIs for the VSSL signal presented below. (B) Normalized VSSL signal as a function of T_seg around MCA, ACA, and PCA, as well as at the 4V, CA, and FLV (N = 3, error bars represent standard derivation). The normalized VSSL signal around MCA and ACA are relatively stable across different T_seg, while the signal decreased significantly from T_seg = 10 to 30 ms in the CA, 4V, FLV, and around PCA. The arteries in SAS are indicated with yellow arrows, and other ROIs are labeled with green arrows.

The VSSL signal as a function of TE

Figure 3A presents representative normalized VSSL images acquired at TE of 816 ms and 39 ms for a subject. The velocity-encoding gradient was oriented along the S-I direction. The normalized VSSL signals around the arteries in the SAS and in the ventricles were plotted in Figure 3B and summarized in Table 1. The signals are substantially higher at TE = 39 ms compared to TE = 816 ms, with their signal ratios ranging from 1.14 (CA) to 2.09 (3V) (Table 1). For the signals around the arteries in SAS, the signal ratios ranged from 1.23 (A3) to 1.91 (P1). Although the normalized VSSL signals at TE = 39 ms are overall higher than those at TE = 816 ms with less T₂-decay effect, significant differences were only found in P1 and P2 segments of the PCA, IPC, and SC, which is expected for CSF with a long T₂ value. When using the VSSL module, a long Tseg (30 ms) was applied, resulting in an extended VSSL preparation time (270 ms), which effectively suppressed both parenchyma and blood signals. The M0 images acquired with a short TE were able to detect weak tissue signal as demonstrated in the Supplementary Figure S2, likely due to the exceptionally long acquisition time, which led to blurring and reduced SNR for the short T₂ components. Despite the enhanced signal at TE = 39 ms, the image quality at TE = 816 ms was superior. The centric ordering used for TE = 39 ms may result in contamination from tissue in the M0 images (Fig. S2).(59) Therefore, we utilized TE = 816 ms for the acquisition.

• Download figure
• Open in new tab

Figure 3.

The results for the normalized VSSL signal with different TEs. (A) The typical normalized VSSL images acquired at TE = 816 and 39 ms with S-I velocity-encoding gradient directions for a representative healthy subject. (B) The boxplot of normalized VSSL signals in each ROI with different TE. A two-sample t-test was performed for comparison (* p < 0.05, ** p < 0.01).

CSF flow patterns in SAS and ventricles

The CSF flow distribution revealed by the VSSL method is presented in Figure 4. The M0 image is shown in Figure 4A. The images, representing the sum signal of three orthogonal velocity-encoding directions , are depicted in Figures 4B-D. The average of VSSL images acquired from three velocity-encoding directions (S_average = ∑_{S–l,A–P,L–R}(S_control − S_label) /3) is overlaid with the MRA map and the MPRAGE image to identify the arteries in SAS and ventricles (Figures 4E-G). CSF flow around the arteries of M1 and M2 segments of MCA, A3 segment of ACA, and P1 and P2 segments of PCA was detected.

• Download figure
• Open in new tab

Figure 4.

The M0 map (A) and normalized VSSL images (l_sum) in sagittal (B), axial (C), and coronal (D) views. VSSL images were acquired with T_seg = 30 ms and G_max = 40 mT/m. VSSL image (S_average) was overlaid with the MRA map and MPRAGE image in sagittal (E), axial (F), and coronal (G) views. The MRA map is shown in red, and the unnormalized VSSL signal is displayed using a jet colormap. The arteries in SAS are indicated with yellow arrows, and other ROIs are labeled with green arrows.

To better visualize the velocity distribution both around the arteries in SAS and in ventricles, maximum-intensity projections (MIP) are shown in Figure 5 with three views. These MIP images were produced directly from the VSSL image S_average without any manual enhancement. The VSSL signal in the SAS, especially in the cortical sulci region as demonstrate in the Supplementary Figure S2, is not visible, indicating effective suppression of the diffusion component in the VSSL images (Figure S3). The CSF flow around the arteries of the PCA and MCA is much higher than in other arteries. Rapid CSF flow is mainly observed in the SC and IPC, as well as in ventricles, including the 3V, and 4V. The CSF flow in the CA and FMo is also measured, further validating the VSSL method.

• Download figure
• Open in new tab

Figure 5.

Maximum-Intensity Projections (MIP) of the VSSL image (S_average) in sagittal (A, D), coronal (B, E), and axial (C, F) views. The top row images (A-C) use a jet colormap to show the VSSL signal in the SAS and ventricles. The bottom row images (D-F) employ a zoomed scale to highlight the small arteries, as shown by blue arrows in the inset figure. The arteries in SAS are indicated with yellow arrows, and other ROIs are labeled with green arrows.

The VSSL method effectively detects CSF flow around the major intracranial arteries (MCA, ACA, and PCA). However, detecting the flow around the basilar artery (BA) is challenging due to the extremely high CSF flow in the PPC, which obscures the CSF flow signal around the BA. The MIP images demonstrate that rapid CSF flow is confined around the MCA, ACA, and PCA, as the signals in the adjacent SAS are relatively low.

Flow velocity calibration with PC-MRI

The CSF movement in the IPC, PPC, and SC were successfully detected using 2D high-resolution PC-MRI in all subjects. Figure 6A illustrates the imaging position of PC-MRI and the detected flow map with VENC of 10 cm/s along the S-I direction from one exemplary subject. The normalized VSSL signal acquired with S-I velocity-encoding direction is plotted in Figure 6B from the same subject for comparison. Figure 6C displays the PC velocity and PC absolute velocity waveform of the ROI shown in Figure 6A for illustration. The pulsatile flow pattern can be clearly visualized. The velocity waveforms were also extracted from all detected ventricles, which showed consistent pulsatile flow patterns across a cardiac cycle. Figure 6D shows the correlation between the normalized VSSL signal and average PC velocity of the whole cardiac cycle values measured by PC-MRI in all three directions, which shows a clear linear correlation. We applied a linear regression model with no intercept term to evaluate the correlation of normalized VSSL signal intensity and PC velocity, and then obtained the velocity maps.

V_S–l, V_A–P, and V_L–R represent the obtained velocity along S-I, A-P, and L-R direction, respectively. The absolute velocity is calculated by .

• Download figure
• Open in new tab

Figure 6.

Calibration of VSSL with PC-MRI. (A) Velocity map obtained with PC-MRI for a representative healthy subject, showing pulsatile CSF flow throughout the SC, IPC, and blood flow in vessels. (B) Typical normalized VSSL images acquired with the S-I velocity-encoding direction on the same subject. (C) Representative PC flow curves with the S-I velocity-encoding direction for one cardiac cycle measured with pulse gating from the ROI shown in Figure A (yellow box). The procedure for calculating the average flow rate is illustrated in the image below. This involves taking the absolute values of the dynamic flow rates over cardiac cycle and then computing their average to obtain the final result. (D) Linear regression model between normalized VSSL signal and average velocity of the whole cardiac cycle in all three directions for the manually selected ROIs in the IPC, PPC, and SC. R² = 0.95.

The CSF flow along different directions

The typical velocity maps derived from Equation [2] with velocity-encoding gradients along S-I, L-R, and A-P directions and the corresponding boxplots are shown in Figure 7. The VSSL-derived velocities from different ROIs are listed in Table 2. CSF flow around M1, P1, and P2 exhibit high flow velocities of 0.461 ± 0.143 cm/s. The CSF velocities around M2 and A3 are much slower (0.155 ± 0.062 cm/s), but the flow velocity is still about 3 times higher than that of the FLV (0.041 ± 0.013 cm/s). Significant orientation dependence was observed around M2 and A3 (p < 0.05), with the highest velocity along the vessel walls. While CSF flow around other arteries also show a consistent trend of the highest velocity along the vessel walls, it does not reach statistical significance. The average CSF flow velocity in ventricles is 0.309 ± 0.116 cm/s. A significantly higher flow velocities along the S-I direction are found in most ventricular regions compared to those of A-P and L-R directions. The CSF flow velocities are 0.525 ± 0.105 cm/s in IPC and 0.819 ± 0.157 cm/s in SC.

• Download figure
• Open in new tab

Figure 7.

The VSSL-derived velocity maps for three different velocity-encoding directions.

(A) The VSSL-derived velocity maps of S-I, A-P, and L-R directions for a representative healthy subject. (B) The boxplots of VSSL-derived velocity in each ROI across three orthogonal velocity-encoding directions. One-way ANOVA with post-hoc Tukey’s honestly significant difference tests for pairwise comparisons was used for comparison (* p < 0.05, ** p < 0.01, *** p < 0.001).