Temporal fluctuation in lateral ventricle volume and its coupling with CSF inflow and global BOLD signal

Abstract

Recent studies have highlighted the intricate relationship between cerebrospinal fluid (CSF) dynamics and global brain activity, suggesting a role in neurovascular coupling and brain waste clearance. The lateral ventricles are believed to play a key role in linking global BOLD (gBOLD) signals to CSF inflow (CSF_in) to the fourth ventricle. In this study, we developed a method to reliably quantify lateral ventricle volume (LVV) in fMRI data. Using three independent datasets, including resting-state and task-based fMRI, we assessed dynamic changes in LVV and their associations with gBOLD and CSF_in. Our findings reveal a strong anti-correlation between LVV and gBOLD across all datasets, with an average gBOLD lag of approximately 1 s. The derivative of the LVV time series were positively correlated with CSF_in, with CSF_in lagging LVV changes by 1.4–2.4 s. A moderate negative correlation was also observed between CSF_in and gBOLD, consistent with prior research. These results support the hypothesis that LVV fluctuations, driven by global cerebral blood volume oscillations, regulate CSF movement into and out of the fourth ventricle. Our findings provide a foundation for further investigations into the role of LVV dynamics in aging and neurological disorders.

Introduction

Cerebrospinal fluid (CSF) plays an important role in clearing metabolic waste from the brain and is an essential component of the glymphatic system. The flow of CSF in the brain is very complex, involving multiple pathways and exhibiting pulsations driven by cardiac and respiratory cycles^{1, 2, 3–4}. Recent research has identified slow-wave dynamics of CSF inflow (CSF_in) during non-rapid eye movement (NREP) sleep⁵. These oscillations, predominantly measured below or within the fourth ventricle, are closely coupled with electrophysiological and hemodynamic activities. Subsequent studies showed that such coupling also exists during wakefulness^6,7. Notably, reduced coupling between CSF_in to the fourth ventricle and global blood oxygen level-dependent (BOLD) activity has been linked with Alzheimer’s disease (AD)–related pathology, and further underscores the critical role of physiological modulation in the clearance of AD-related brain waste⁷.

The global BOLD (gBOLD) signal, measured as the mean MRI signal in the gray matter, and CSF_in, quantified as signal enhancement in CSF voxels outlined in the 1–2 lowest (most inferior) slices, are independent measurements belonging to distinct physiological systems: the vascular and glymphatic systems, respectively. Understanding the dynamics of CSF flow and its relationship to the vascular system requires identifying the driving forces and boundary conditions that regulate CSF movement. Converging evidence suggests that cerebral blood volume (CBV) serves as the primary driving force behind CSF flow^3,8,9. For instance, particle tracking velocimetry has revealed a coupling between pulsatile CSF flow and the velocity of the arterial wall motion³. More recently, direct coupling between arterial dilation/constriction and periarterial CSF flow velocity has been observed using particle tracking techniques⁸. In addition, a poroelastic model predicts that arterial vasodilation generates convective CSF flow, and low-frequency oscillations in vessel dilation promotes CSF transport through the paravascular space¹⁰.

Although the coupling between BOLD signal and CBV is well established^11,12, a significant gap remains in linking gBOLD signals to CSF_in to the fourth ventricle, a region distant from localized CBV changes. Figure 1 illustrates the relationship between CBV, CSF flow, gBOLD, and CSF_in to the fourth ventricle. The BOLD effect is modulated by CBV both regionally and globally, while the regional CSF (rCSF) flow is modulated by regional CBV (rCBV). However, there is no direct interaction between global CBV (gCBV) changes and CSF_in to explain observed coupling of oscillations between the gBOLD signal and CSF_in.

Fig. 1 [Images not available. See PDF.]

Illustration of the relationship between cerebral blood volume (CBV), BOLD, lateral ventricle volume (LVV), regional CSF (rCSF) flow, and CSF inflow (CSF_in) to the fourth ventricle. Rectangles represent regional quantities, while circles denote global or bulk quantities. Multiple arrows (from BOLD and regional CBV [rCBV]) illustrate how regional effects combine to generate global measures. Solid thicker arrows indicate direct modulation effects. It is well established that regional CBV changes induce BOLD signal change and drive regional CSF flow. A coupling between gBOLD and global CBV (gCBV) can be inferred, as CBV and BOLD signals are both scalar quantities. However, the relationship between gCBV and CSF_in remains undetermined due to the directional and complex nature of CSF flow in the brain. The model proposed by Yang et al. introduce LVV as a node to link gCBV signal change and CSF_in to the fourth ventricle.

As the primary compartments containing CSF, the ventricles likely play a significant role in the glymphatic system, particularly in CSF secretion and flow¹³. For instance, hydraulic pressure from the ventricles could be a driving force for CSF movement¹⁴. However, the relationship between the ventricles and the dynamics of CSF_in to the fourth ventricle remains unclear. Recently, a mechanical model was proposed to explain the coupling between functional MRI (fMRI) hemodynamic signals and the CSF_in to the fourth ventricle⁶. In this model, the ventricles, especially the lateral ventricles, serve as the bridge connecting hemodynamic activity to CSF movement. The model hypothesizes that changes in CBV exert forces on the walls of the lateral ventricles, thereby driving CSF flow into and out of the fourth ventricle (dashed lines in Fig. 1). This model successfully accounts for the correlations and time delays observed between the derivative of gBOLD signal and CSF_in oscillations in the fourth ventricle during awake resting-state fMRI.

While the lateral ventricles are a key component in the proposed model, which posits an interaction between changes in CBV and the volume of the lateral ventricles, there is no direct measurement of lateral ventricle volume (LVV) in this context. Although changes in ventricle size over longer timescales—spanning months, years, or decades—are well documented in conditions such as aging¹⁵, alcoholism¹⁶, and various neurological diseases, including Alzheimer’s disease^17,18, little research has focused on dynamic changes occurring over much shorter timescales (seconds to minutes) and their potential coupling with CSF flow. To unlock the mechanism underlying the coupling between global BOLD signal and the CSF inflow to the fourth ventricle and to further validate the model proposed by Yang et al., we developed a method to measure dynamic changes in LVV using fMRI data. We validated it using existing fMRI datasets, characterized the dynamic changes in LVV, and explored their relationship with CSF_in and gBOLD signals. Beyond validating Yang et al.’s model and elucidating the complex interplay between the vascular and glymphatic systems, this approach may have broader clinical implications in non-invasively assessing changes in the glymphatic system and its relationship with neurovascular activity in both health and disease (e.g., Alzheimer’s disease, differentiating hydrocephalus subtypes (e.g., obstructive vs. communicating), and understanding changes during sleep).

Methods

Three independent MRI datasets were utilized in this study. We applied a segmentation tool on functional MRI data to acquire dynamic measures of LVV and characterize their properties across these datasets. The first exploratory dataset was acquired in-house on a small number of participants at high sampling rate to examine power spectrum of LVV while minimizing aliasing of physiological signals. The other two datasets, downloaded from OpenNeuro, were selected to assess the accuracy and reliability of our LVV measurement approach and investigate the coupling between LVV fluctuations, gBOLD signals, and CSF_in under broader conditions (e.g., rest vs. task; scanner differences). All experiments were performed in accordance with relevant guidelines and regulations.

Participants

For Dataset 1, five healthy participants without neurological disorders (3 female, ages 17–23 years) were scanned across various states of wakefulness and sleep. The study was approved by the institutional ethical review board of Indiana University. Written informed consent was obtained from all participants. Dataset 2 were from 32 participants (16 female, ages 19–29 years) that underwent MRI scans at Emory university. The OpenNeuro accession number is ds003540. All participants were healthy, unmedicated, and right-handed. This study was approved by the Institutional Review Board of Emory University. Dataset 3 were from 58 participants from Spain (29 female, age 32.01 ± 11.22 years) with MRI scans obtained at Barcelonaβeta Brain Research Center in Barcelona, Spain. None of the individuals had any neurological or mental health disorders. This research was approved by the Ethics Committee on Human and Animal Experimentation at the Universitat Autònoma de Barcelona. The OpenNeuro accession number is ds005375. The demographic information for all three datasets is listed in Table 1.

Table 1. Demorgraphic information for all datasets.

Dataset	Sample size	Sex		Age range (years)	Mean ± SD (years)
		M	F
1	5	2	3	17–23	20.00 ± 2.83
2	32	16	16	19–29	N/A
3	58	29	29	18–55	32.01 ± 11.22

MRI data

All scans in Dataset 1 were conducted on a 3 T Siemens Prisma-Fit scanner equipped with a 64-channel head-neck coil. The participants were instructed to close their eyes and rest during the scan. A T1-weighted anatomical image was acquired with the MPRAGE pulse sequence, followed by fast echo-planar imaging (EPI) scans with the following parameters: TR/TE = 415/31 ms, flip angle = 15°, in-plane resolution = 2.5 × 2.5 mm, slice thickness = 3 mm, and SMS acceleration factor = 7. The lowest slice was positioned either at the CSF inflow entrance of the fourth ventricle, or at the cerebral aqueduct leading into the fourth ventricle, as described in previous studies^5,19. The inclusion of Dataset 1 serves two purposes: (1) the shorter TR allows CSF inflow to induce larger MRI signal changes; and (2) it allows for investigation of cardiac and respiratory effects on LVV dynamics.

Dataset 2 was the Emory multiband dataset²⁰with participants scanned on a 3 T MR scanner (Siemens, Prisma-Fit) using a 32-channel head coil, with various multiband factors (1, 2, 3, 4, 6, 8, 9, and 12). The participants remained at resting state during the scans, each lasting 6 min. Since image contrast and signal-to-noise ratio (SNR) varied with multiband factors, not all scans were ideal for the quantification of LVV. We selected imaging data with a multiband factor of 4 (MB4), as it offers a good balance between SNR, contrast, and sampling rate. The TR of this dataset is 1.44 s, with a 2 mm isotropic resolution. We refer to this dataset as 2A. For comparison, we also analyzed data from single-band (SB) scans with a TR of 3 s and isotropic resolution of 3 mm, which we call dataset 2B. Dataset 2A contains data from 31 participants (one subject’s images could not be read), with each participant contributing 248 volumes. Dataset 2B includes 32 participants, each with 114 volumes. Additionally, each participant has a separate T1-weighted (T1w) image with 1 mm isotropic resolution. In both Dataset 2A and 2B, only one participant had a mean framewise displacement exceeding 0.2 mm or had 1/6th of scans exceeding 0.25 mm in framewise displacement²¹.

Dataset 3 was the POLEX dataset²². The participants performed three tasks. We used the data associated with the cyberball task, in which participants engage in a ball-tossing game with three virtual avatars. The fMRI images were acquired on a Philips Ingenia 3 T CX scanner, with a TR of 1.75 s and a resolution of 3 × 3 × 3.1 mm³. The dataset contains 261 volumes in total. The T1w image has a resolution of 0.65 × 0.65 × 1.0 mm³. The scan duration for the cyberball task fMRI is 10.5 min. Framewise head displacement was less than 0.5 mm²³.

Data processing

Quantification of lateral ventricle volume

The quantification of lateral ventricle volume was performed using the new functionality “SynthSeg” of FreeSurfer software package (version 7.3.1; https://surfer.nmr.mgh.harvard.edu/). SynthSeg generates masks for the left and right lateral ventricles, as well as the gray and white matter. It is an unsupervised segmentation tool based on convolutional neural networks, trained with fully randomized synthetic scans featuring different orientations, contrasts, resolutions, and artifacts^24,25. This training makes SynthSeg robust to variations in contrast and resolution. In FreeSurfer software package, SynthSeg is executed using the command “mri_synthseg,” and we employed the “–robust” flag to further improve its robustness.

For Dataset 1, the EPI images were first motion-corrected, and a moving-average approach was employed to enhance the signal-to-noise ratio (SNR) by averaging six continuous volumes. The resulting images were denoised using 3D FFDNet²⁶, after which SynthSeg was applied to the denoised images. With TR = 0.415 Hz, the Nyquist frequency is 1.2 Hz. Since sliding-window reduced the effective sampling rate by sixfold—potentially introducing aliasing of respiratory and cardiac signals,—a parallel analysis was performed on Dataset 1 without sliding-window averaging, solely for power spectrum density (PSD) assessment.

For Datasets 2 and 3, SynthSeg was applied directly to the original functional images without any preprocessing to generate a segmented image for each functional volume. For each segmented image, the lateral ventricle volume (LVV) was calculated as the sum of total voxels within these masks multiplied by the voxel size. All datasets were visually inspected for segmentation accuracy. One subject in Dataset3 (ID: 1181) was excluded in this analysis due to an extensive signal void in lateral ventricle.

The accuracy of SynthSeg was initially evaluated using high-resolution T1w anatomical images by comparing its results with those from FreeSurfer segmentation for each subject in Datasets 2 and 3. The LVV derived from FreeSurfer segmentation served as the ground truth for this comparison. To account for the relatively low spatial resolution of the EPI-based functional images across all datasets, we further assessed the accuracy and reliability of SynthSeg in extracting LVV from these lower-resolution functional images. Specifically, we compared the average LVV values obtained from the functional images with the ground truth.

A time series of LVV was obtained for each fMRI time series. All LVV time series were demeaned (mean-subtracted) and subsequently filtered using a bandpass filter with a frequency range of [0.01, 0.1] Hz for further analysis except in computing the PSD^6,7. The standard deviation of the bandpass filtered LVV time series was computed to characterize the magnitude of its low-frequency fluctuation over time.

Computation of global BOLD signal

The global BOLD (gBOLD) signal was computed as the mean MRI signal within the gray matter mask derived from SynthSeg. The gBOLD signals were then demeaned and subjected to a bandpass filter with a frequency range of 0.01, 0.1 Hz. The coefficient of variation (CV) of gBOLD was calculated as the ratio of the standard deviation of the bandpass filtered gBOLD time series to the mean value of original gBOLD signals.

Extraction of CSF inflow (CSF_in)

The low-frequency oscillation of CSF_in to the fourth ventricle was first discovered by Fultz et al. using a fast-imaging technique⁵. The CSF inflow is characterized by the enhancement of the MRI signal in the CSF voxels in the few lower slices, as incoming fresh spins from CSF flow are immune to the saturation effect. This signal enhancement persists even at longer TRs (e.g., 2 s)¹⁹ due to the long T1 relaxation time of CSF (~ 4 s).

CSF_in was extracted only for Datasets 1, 2A and 3, as the TR of 3 s in Dataset 2B is not suitable for this analysis. The CSF mask was manually outlined on the lowest slice using the first volume of the fMRI 4D data. Three subjects in Dataset 2A were excluded due to uncertainty of the CSF masks. The mean MRI signal within the CSF mask for all functional volumes was then computed, which is assumed to be correlated with the CSF inflow to the plane of the lowest slice. Throughout this paper, CSF_in, referred as CSF movement towards the fourth ventricle, is represented by the mean MRI signal in the CSF mask. The CSF signal was also demeaned and bandpass filtered within the frequency range of [0.01, 0.1] Hz.

Correlation analysis

To evaluate the performance of SynthSeg, we used the LVV value derived from the T1w image using FreeSurfer segmentation as the ground truth. First, we computed the correlation coefficients between the LVV extracted from the T1w image using SynthSeg and the ground truth across all subjects. Next, we calculated the mean LVV values extracted from functional images using SynthSeg and assessed their correlation with the ground truth.

Correlation analysis was conducted on the time series of LVV-gBOLD, gBOLD-CSF_in and after applying bandpass filtering in the frequency range of 0.01 Hz to 0.1 Hz for each subject. Previous studies have reported small lags of CSF_in relative to . To account for potential time shifts between these signals, we systematically applied temporal shifts in increments of the repetition time (TR) within a ± 10-s window to identify the maximum correlation or anticorrelation. Additionally, we calculated the correlation between the coefficient of variation (CV) of gBOLD and the standard deviation of LVV across all subjects.

We aimed to analyze only the low frequency oscillations between 0.01 and 0.1 Hz for several reasons:

The frequency range of 0.01–0.1 Hz is commonly used to study resting-state functional connectivity, which is believed to reflect spontaneous brain activity and intrinsic functional networks²⁷.

We hypothesized that the LVV oscillation is related to the vasomotion, which lies in the low-frequency range of 0.01–0.1 Hz²⁸.

The low frequency oscillation of CSF inflow and its coupling to global BOLD signal has been observed and studied widely, the focus on the low frequency range makes our results comparable with others^7,29.

Results

Figure 2 presents examples of the segmented lateral ventricles using SynthSeg on representative images from Datasets 1, 2A, 2B, and Dataset 3. In all these datasets, the MRI signal within the lateral ventricle appears brighter than the surrounding tissue. SynthSeg successfully outlined the lateral ventricle automatically, and the segmentation results appear visually promising.

Fig. 2 [Images not available. See PDF.]

Example of the segmented lateral ventricles using SynthSeg for representative images from Datasets 1 (a), 2A (b), 2B (c), and Dataset 3 (d).

A detailed analysis of SynthSeg demonstrated that both FreeSufer segmentation and SynthSeg yielded highly similar results in quantifying the LVV from T1w images, as evidenced by the strong correlations (r(29) = 0.995 for Dataset 2A; r(55) = 0.998 for Dataset 3) displayed in Fig. 3 Moreover, a strong correlation was observed between the LVVs derived from T1w images using FreeSurfer and those obtained from EPI images using SynthSeg (r(29) = 0.970 for Dataset 2A; r(55) = 0.980 for Dataset 3), despite an overall overestimation of LVV in the EPI images by SynthSeg.

Fig. 3 [Images not available. See PDF.]

Correlation between lateral ventricle volume (LVV) obtained from FreeSurfer segmentation (ground truth) on the T1w anatomical image and LVV derived using SynthSeg from T1w image and EPI images for Dataset 2A (top) and Dataset 3 (bottom).

In Dataset 1, framewise head motion was below 0.1 mm for all participants. Figure 4 displays the PSD of the LVV time series–computed without sliding window in preprocessing –for all five subjects. Respiration or cardiac peaks are not clearly visible in the PSD. The PSD is predominantly characterized by low-frequency components below 0.2 Hz except for sub-02, which is much noisier due to smaller ventricle anatomy. Subjects 2, 4, and 5 reported falling asleep toward the end of the scan; however, no clear differences in brain states were reflected in the PSD. PSD analysis was not performed for other datasets due to their long TR values.

Fig. 4 [Images not available. See PDF.]

Power spectral density (PSD) of the LVV time series for all five subjects in Dataset 1 (note that data is not temporally filtered). There are no prominent peaks corresponding to cardiac or respiratory activities.

Figure 5 shows the normalized gBOLD signal, CSF_in, LVV fluctuations, and over time for a representative subject from Dataset 2A (TR = 1.44 s). The correlation coefficients indicate strong relationships between these variables: − 0.96 between LVV and gBOLD (with gBOLD lagging by 1 TR), 0.79 between and CSF_in (with CSF_in lagging 1 TR), and − 0.79 between gBOLD and CSF_in (with gBOLD lagging 4 TR). Additionally, visual changes in lateral ventricle size can be observed.

Fig. 5 [Images not available. See PDF.]

Fluctuations of lateral ventricle volume (LVV, green), global BOLD (gBOLD) signal (blue), MRI signal changes induced by CSF inflow (CSF_in, red), and for a representative subject from Dataset 2A (subject 29). All signals are normalized to the range [0 1] and bandpass filtered in the frequency range of 0.01–0.1 Hz. The correlation coefficients are -0.96 between LVV and gBOLD, 0.79 between and CSF_in, and − 0.79 between gBOLD and CSF_in. A snapshot of the extracted lateral ventricle using SynthSeg is displayed for two time points. The change of lateral ventricle size can be perceived visually.

The correlation analysis results for LVV, gBOLD, CSF_in, and in Dataset 1 are summarized in Table 2, while results for Dataset 2A (rsfMRI) and Dataset 3 (task fMRI) are presented in Fig. 6. A strong anti-correlation between LVV and gBOLD was observed across all datasets. Specifically, in Dataset 1, the correlation between LVV and gBOLD is − 0.83 ± 0.06, with an average gBOLD lag of 0.25 s; In Dataset 2A, the correlation was − 0.65 ± 0.20, with an average gBOLD lag of 1.2 s. In Dataset 3, the correlation was − 0.61 ± 0.14, with an average gBOLD lag of 0.34 s.Table 2

Measurements of the lateral ventricle volume (LVV) and correlations between low-frequency dynamic changes in LVV, gBOLD, , and CSF_in signals from Dataset 1.

Subject	Total LVV (cm3)	Range of LVV change (%)	Correlation	Correlation (LVV-gBOLD)	Correlation (CSF-gBOLD)
1	18.5	6.1	0.73	− 0.86	− 0.76
2	7.5	8.5	0.74	− 0.85	− 0.77
3	14.5	5.8	0.58	− 0.73	− 0.67
4	15.1	9.1	0.63	− 0.88	− 0.64
5	13.2	6.2	0.68	− 0.83	− 0.58

Fig. 6 [Images not available. See PDF.]

Correlation analysis results for lateral ventricle volume (LVV), global BOLD (gBOLD), and CSF inflow (CSF_in) in Dataset 2A (resting-state fMRI) and Dataset 3 (task fMRI) visualized in box plots. Global BOLD signal is negatively correlated with both LVV and CSF_in. Notably, the magnitude of the anticorrelation between LVV and gBOLD is significantly larger than that between CSF_in and gBOLD in both datasets. A moderate positive correlation is observed between and CSF_in.

Consistent with previous research, a moderate anti-correlation between CSF_in and gBOLD was observed across datasets. The correlations were − 0.68 ± 0.08 for Dataset 1, − 0.50 ± 0.19 for Dataset 2A, and − 0.47 ± 0.13 for Dataset 3. These correlations were less pronounced (less negative) than those observed between LVV and gBOLD.

Finally, a positive correlation was found between and CSF inflow, with correlation coefficients of 0.67 ± 0.07 for Dataset 1, 0.48 ± 0.19 for Dataset 2A and 0.40 ± 0.15 for Dataset 3. Notably, CSF inflow lagged behind LVV changes, with mean lag values of 2.4 s in Dataset 1, 1.4 s in Dataset 2A, and 1.5 s in Dataset 3.

Figure 7 demonstrates strong correlation between that the magnitude of LVV fluctuations, characterized by the standard deviation, and the variation of the BOLD signal, characterized by the coefficient of variance (r(29) = 0.73, p < 0.001 for Dataset 2A [resting-state fMRI] and r(55) = 0.64, p < 0.001 for Dataset 3 [task-based fMRI]).

Fig. 7 [Images not available. See PDF.]

Scatter plots showing the relationship between the coefficient of variation (CV) of the BOLD signal and lateral ventricle volume (LVV) fluctuations, characterized by their standard deviation. The correlation coefficients are 0.73 for resting-state fMRI data (Dataset 2) and 0.64 for task fMRI data (Dataset 3).

The LVV values derived from Datasets 2B align closely with those derived from Dataset 2A. Although there is a 10% difference in the average absolute values, the correlation between them remains remarkably high (r(29) = 0.993, p < 0.001) across 31 participants. Additionally, the magnitude of LVV fluctuations exhibits moderate consistency between the two data acquisitions (r(29) = 0.45, p = 0.011), comparable to the correlation observed for the size of BOLD signal (r(29) = 0.43, p = 0.016 ).

Discussion

Here, we developed and applied a new method of extracting lateral ventricle volume (LVV) from EPI images in a large variety of typical fMRI datasets. We investigated the dynamic change of LVV in resting state fMRI and task fMRI experiments and found a strong anticorrelation between LVV and gBOLD signal, as well as a moderate correlation between LVV and CSF inflow. As a critical compartment for CSF secretion/storage and a buffer in CSF flow, the volume change of lateral ventricles provides a novel perspective for understanding the functioning of the glymphatic system.

SynthSeg is a robust segmentation tool based on convolutional neural network that can deal with MRI images of any contrast and resolution, even with poor tissue contrast and low SNR^24,25. On high-resolution T1w images, SynthSeg and FreeSurfer segmentation achieves highly similar results with an exceptional correlation of 0.997. The accuracy degrades slightly when applied to EPI-based functional images. SynthSeg tends to overestimate the LVV value systematically. Nevertheless, the correlation with the ground truth remains high (> 0.97), along with the strong correlation (0.993) between different runs with varying acquisition parameters in Dataset 2, demonstrating strong consistency. Additionally, the mean standard deviation of LVV is 140 mm³, representing approximately 1% of the total LVV, which demonstrates good precision in the measurements. Together, these findings confirm that the LVV quantified from functional images using SynthSeg is reliable for quantitative analysis.

In both resting-state and task fMRI experiments, we observed considerable variability of lateral ventricle size across participants, ranging from 5 to 42 cm³ (Fig. 3), as well as in the dynamic change of LVV (Fig. 7). The standard deviation of LVV spans from 0.05 to 0.35 cm³. These dynamic changes in LVV, which have not been previously reported, are unlikely due to imaging artifacts or errors from SynthSeg. Naturally, the question arises: what drives these changes? Since the ventricle is a cavity filled with CSF and surrounded by brain tissue, the observed changes must stem either from CSF flow or from movement of the brain parenchyma. The delayed CSF inflow to the fourth ventricle suggests that CSF flow is a consequence of LVV changes rather than the driving force behind them.

The LVV dynamics are predominantly observed in the low-frequency range below 0.2 Hz in Dataset 1 (Fig. 4). Low-frequency oscillations (LFOs) of cerebral blood volume (CBV) and cerebral blood flow (CBF) have long been observed in studies of spontaneous BOLD fluctuations, typically below the frequency of 0.15 Hz³⁰. As regional CBV increases, it is accompanied by a corresponding regional decrease in CSF space³¹. According to the Monro-Kellie doctrine, which assumes intracranial volume is constant, changes in global CBV will induce changes in LVV to balance in intracranial pressure. This hypothesis is supported by several observations:

• 1. The anticorrelation between LVV and gBOLD: Strong anticorrelation (< − 0.6) between LVV and gBOLD has been consistently observed across all three datasets. BOLD is not a direct measure of neuronal activity, it is a blood-related composite signal reflecting changes in blood flow, volume, and oxygenation. Given the well-known positive coupling between CBV and BOLD (Fig. 1), the observed anticorrelation between LVV and gBOLD explains the inverse relationship between LVV and CBV. The small lag of gBOLD signal to LVV (~ 1 s) also coincides with previous findings that BOLD signal lags both CBV and CBF signals³².

• 2. The positive correlation between LVV variation and gBOLD fluctuation magnitude: A positive correlation exists between the magnitude of LVV variation and the size of gBOLD fluctuations. This observation, along with the inverse coupling between gBOLD and LVV in individual subjects, suggests that LVV contraction is driven by the expansion of brain tissues caused by arterial vasodilation, which simultaneously leads to an increased BOLD signal.

• 3. Validation of Yang et al.’s model⁶: Yang et al. proposed a mechanical model in which the LVV acts as an intermediary structure, transmitting forces from vasodilation or vasoconstriction to drive the CSF flow in/out of the fourth ventricle. Experimental data have validated this model, further supporting the relationship between CBV changes and LVV dynamics.

Thus, CBV is at least one significant source driving LVV changes, providing critical insights into the mechanisms of neurovascular coupling underlying the glymphatic system.

A positive correlation was observed between time series of and CSF_in to the fourth ventricle. The correlation peaks when CSF takes a lag in the order of seconds. Compelling evidence shows that CSF flow is primarily driven by cerebrovascular activity locally^8,9,33. The macroscopic CSF flow within the brain is very complex. It consists of several paths¹. Within the ventricles, CSF flows from lateral ventricles to the third ventricle and flow through the cerebral aqueduct to the fourth ventricle.

In the pioneering study by Fultz et al., the dynamic change of CSF_in was assumed to be driven by CBV based on the coupling between CSF_in and gBOLD signals. Following this, two mechanistic pathways have been proposed to explain the slow-wave oscillation of CSF to the fourth ventricle during sleep and its correlation with global BOLD¹⁹, in which CBV is the key component (node) in both pathways. However, none of these studies provided strong evidence linking CBV fluctuation to CSF_in oscillation. Yang et al. attempted to explain their relationship with a mechanical model⁶, in which a critical hypothesis is that the CSF inflow to the fourth ventricle is resultant from the compression of LVV. A change in LVV implies a displacement of the CSF, which must be coupled with the CSF flow. Therefore, the positive coupling between and CSF_in, as well as the lag of CSF_in observed in our results, support this hypothesis in Yang’s model: a global CBV change lead to the contraction or expansion of the lateral ventricles, subsequently driving bulk CSF flow into or out of the fourth ventricle. Thus, both LVV and CSF_in oscillations are likely driven (at least partially) by a common source of global CBV oscillations.

The dynamics of LVV provide valuable insights that refine and expand upon the mechanical model proposed by Yang et al.⁶, which has been validated by the experiment data. The model can be fine-tuned by incorporating the volume changes of the lateral ventricles as well as the time delay between LVV and gBOLD and that between CSF_in and . Furthermore, more sophisticated computational models based on fluid dynamics have been proposed to better capture the complex behavior of CSF flow within the brain^34,35. These models require precise physiological parameters and physical boundary conditions to reduce discrepancies between model predictions and real-world measurements. While previous studies primarily regarded the lateral ventricles as sites for CSF production and structural constraints, emerging features such as the low-frequency oscillations of LVV and their amplitudes provide new parameters for refined models. These features can serve as both input variables for model development and benchmarks for validating model predictions.

Reduced coupling between gBOLD and CSF_in has been reported for various neurological disorders, including Alzheimer’s disease, Parkinson’s disease, traumatic brain injury, frontotemporal dementia, depression, and small vessel disease^{36, 37, 38, 39–40}. Given the role of CSF in brain waste clearance, understanding the physiological basis of this coupling and its alterations in these conditions is crucial. As a critical link between gBOLD and CSF_in, the lateral ventricles offer unique insights into how global brain activity influences waste clearance. Compared to measuring CSF_in, LVV quantification provides several advantages:

1. Absolute measurement: LVV is a physical measurement with an absolute value, whereas CSF inflow is indirect and challenging to quantify.

2. Compatibility with standard fMRI protocols: LVV can be derived from standard BOLD fMRI data acquisition without requiring significant modifications to the imaging protocol.

3. Robustness to motion: LVV measurement is less sensitive to head motion while CSF inflow highly susceptible to effects of movement, as it is measured at the boundary slice.

4. Automation and retrospective analysis: The process is fully automated and can be applied retrospectively to previously collected data.

Direct measurement of LVV dynamics allows for the characterization of fluctuation magnitude, power spectral density, rate of change ( ), and the coupling between LVV changes, gBOLD, and other parameters. These features may provide insightful information for various neurological conditions or aging. For instance, brain’s stiffness increases with age⁴¹, which might be revealed by smaller LVV fluctuations. Similarly, this approach may prove useful in understanding changes in CSF flow dynamics during sleep⁵.

This study has several limitations. First, the sample size of Dataset 1 is relatively small, which may weaken the conclusion that low-frequency oscillations are dominant. It is also possible that the PSD depends on the arousal state. However, this does not affect the analysis within the low-frequency range, as supported by previous studies on gBOLD and CSF inflow^5,7. Second, the accuracy of CSF inflow measurements in Datasets 2 and 3 is limited. The procedure relies on manual segmentation of CSF signals in the lowest slice, which can be ambiguous due to susceptibility artifacts in EPI. Additionally, the lowest slice position with respect to the fourth ventricle varies across subjects. Furthermore, the averaged CSF signal does not necessarily reflect actual flow dynamics because the CSF flow is not laminar. Third, while SynthSeg is applied to individual images and is less sensitive to head motion, abrupt head movements remain a confounding factor as they can alter the local magnetic field and affect image contrast.

Conclusions

We have developed a novel method for reliably quantifying lateral ventricle volume (LVV) in fMRI data. Using three independent datasets, including resting-state and task-based fMRI, we investigated dynamic changes in LVV and their relationships with gBOLD and CSF inflow to the fourth ventricle. Our results demonstrate significant couplings between LVV dynamics and low-frequency oscillations of gBOLD and CSF inflow, supporting the hypothesis that LVV fluctuations, driven by changes in global cerebral blood volume, play a regulatory role in CSF movement into and out of the fourth ventricle. These findings establish a foundation for further exploration of the role of LVV dynamics in aging and neurological disorders.

Acknowledgements

The authors thank Macie Schmitt for her help in piloting the data acquisition. This work was supported in part by NIH R01MH110630 to DPK.

Author contributions

HC conceived the idea, performed data collection and analysis, and drafted the manuscript. HC and DPK developed the idea collaboratively. DPK provided critical feedback and contributed to shaping the research, analysis, and manuscript.

Data availability

Dataset 1, comprising data from three subjects we collected in-house, is shared on brainlife: https://brainlife.io/project/67faec78da629ed3b136e272/dataset. Datasets 2 and 3 are open-access datasets available publicly: Dataset 2: https://openneuro.org/datasets/ds003540/versions/1.0.1. Dataset 3: https://openneuro.org/datasets/ds005375/versions/1.0.0

Declarations

Competing interests

The authors declare no competing interests.

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. Hladky, S; Barrand, M. Regulation of brain fluid volumes and pressures: Basic principles, intracranial hypertension, ventriculomegaly and hydrocephalus. Fluids Barriers CNS; 2024; 21, 57. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39020364][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11253534]

2. Laganà, M et al. Blood and cerebrospinal fluid flow oscillations measured with real-time phase-contrast MRI: Breathing mode matters. Fluids Barriers CNS; 2022; 19, 100. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36517859][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9749305]

3. Mestre, H et al. Flow of cerebrospinal fluid is driven by arterial pulsations and is reduced in hypertension. Nat. Commun.; 2018; 9, 4878.2018NatCo..9.4878M [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30451853][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6242982]

4. Klose, U; Strik, C; Kiefer, C; Grodd, W. Detection of a relation between respiration and CSF pulsation with an echoplanar technique. J. Magn. Reson. Imaging; 2000; 11, pp. 438-444.1:STN:280:DC%2BD3c3ivFGltw%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/10767073]

5. Fultz, N et al. Coupled electrophysiological, hemodynamic, and cerebrospinal fluid oscillations in human sleep. Science; 2019; 366, pp. 628-631.2019Sci..366.628F1:CAS:528:DC%2BC1MXitV2rt7fL [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31672896][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7309589]

6. Yang, H et al. Coupling between cerebrovascular oscillations and CSF flow fluctuations during wakefulness: An fMRI study. J. Cereb. Blood Flow Metab.; 2022; 42, pp. 1091-1103. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35037498][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9125495]

7. Han, F et al. Reduced coupling between cerebrospinal fluid flow and global brain activity is linked to Alzheimer disease-related pathology. PLoS Biol.; 2021; 19, 1:CAS:528:DC%2BB3MXht1Oisb%2FN [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34061820][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8168893]e3001233.

8. Holstein-Rønsbo, S et al. Glymphatic influx and clearance are accelerated by neurovascular coupling. Nat. Neurosci.; 2023; 26, pp. 1042-1053. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37264158][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10500159]

9. van Veluw, S et al. Vasomotion as a driving force for paravascular clearance in the awake mouse brain. Neuron; 2020; 105, pp. 549-561. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31810839]

10. Kedarasetti, RT; Drew, PJ; Costanzo, F. Arterial vasodilation drives convective fluid flow in the brain: A poroelastic model. Fluids Barriers CNS; 2022; 19, 34. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35570287][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9107702]

11. Drew, P. Vascular and neural basis of the BOLD signal. Curr. Opin. Neurobiol.; 2019; 58, pp. 61-69.1:CAS:528:DC%2BC1MXhtlOlsbvO [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31336326][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6859204]

12. Simon, A; Buxton, R. Understanding the dynamic relationship between cerebral blood flow and the BOLD signal: Implications for quantitative functional. Neuroimage; 2015; 116, pp. 158-167. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25862267]

13. Bothwell, S; Janigro, D; Patabendige, A. Cerebrospinal fluid dynamics and intracranial pressure elevation in neurological diseases. Fluids Barriers CNS; 2019; 16, 9. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30967147][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6456952]

14. Wichmann, T; Damkier, H; Pedersen, M. A brief overview of the cerebrospinal fluid system and its implications for brain and spinal cord diseases. Front. Hum. Neurosci.; 2022; 15, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35126070][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8813779]737217.

15. Kaye, J; DeCarli, C; Luxenberg, J; Rapoport, S. The significance of age-related enlargement of the cerebral ventricles in healthy men and women measured by quantitative computed X-ray tomography. J. Am. Geriatr. Soc.; 1992; 40, pp. 225-231.1:STN:280:DyaK387ms1aisQ%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/1538040]

16. Wobrock, T et al. Effects of abstinence on brain morphology in alcoholism: A MRI study. Eur. Arch. Psychiatry Clin. Neurosci.; 2009; 259, pp. 143-150. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/19165528][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3085767]

17. Lee, J; Heo, D; Choi, K; Kim, H. Impact of the ventricle size on Alzheimer’s disease progression: A retrospective longitudinal study. Dement. Neurocogn. Disord.; 2024; 23, pp. 95-106. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38720825][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11073924]

18. Sayo, A; Jennings, R; Van Horn, J. Study factors influencing ventricular enlargement in schizophrenia: A 20 year follow-up meta-analysis. Neuroimage; 2012; 59, pp. 154-167. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/21787868]

19. Picchioni, D et al. Autonomic arousals contribute to brain fluid pulsations during sleep. Neuroimage; 2022; 249, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35017126]118888.

20. Risk, B. et al. (OpenNeuro Dataset ds003540, 2021).

21. Risk, B et al. Which multiband factor should you choose for your resting-state fMRI study?. Neuroimage; 2021; 234, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33744454]117965.

22. Marcos-Vidal, L. et al. (OpenNeuro Dataset ds005375, 2024).

23. Marcos-Vidal, L et al. When group grievances become personal: The neural correlates of group and personal rejection. Cogn. Affect. Behav. Neurosci.; 2025; [DOI: https://dx.doi.org/10.3758/s13415-024-01257-x] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39776065][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12130114]

24. Billot, B et al. SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Med. Image Anal.; 2023; 86, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36857946][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10154424]102789.

25. Billot, B et al. Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets. Proc. Natl. Acad. Sci. U S A; 2023; 120, 1:CAS:528:DC%2BB3sXmtVCis7c%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36802420][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9992854]e2216399120.

26. Zhang, K; Zuo, W; Zhang, L. FFDNet: Toward a fast and flexible solution for CNN based image denoising. IEEE Trans. Image Process.; 2018; 3818684 [DOI: https://dx.doi.org/10.1109/TIP.2018.2839891] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30596577][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5844189]

27. Van Dijk, K et al. Intrinsic functional connectivity as a tool for human connectomics: Theory, properties, and optimization. J. Neurophysiol.; 2010; 103, pp. 297-321. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/19889849]

28. Liu, S; Zhao, L; Liu, Y. The origin of vasomotion and stochastic resonance in vasomotion. Front. Bioeng. Biotechnol.; 2022; 10, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35309989][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8924506]819716.

29. Han, F et al. Global brain activity and its coupling with cerebrospinal fluid flow is related to tau pathology. Alzheimers Dement.; 2024; 20, pp. 8541-8555.1:CAS:528:DC%2BB2cXisV2ju73I [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39508716][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11667553]

30. Cheng, R; Shang, Y; Hayes, DJ; Saha, S; Yu, G. Noninvasive optical evaluation of spontaneous low frequency oscillations in cerebral hemodynamics. Neuroimage; 2012; 62, pp. 1445-1454. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/22659481]

31. Piechnik, S; Evans, J; Bary, L; Wise, R; Jezzard, P. Functional changes in CSF volume estimated using measurement of water T2 relaxation. Magn. Reson. Med.; 2009; 61, pp. 579-586.1:STN:280:DC%2BD1M7mvVKmsQ%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/19132756]

32. Shen, Q; Ren, H; Duong, T. CBF, BOLD, CBV, and CMRO(2) fMRI signal temporal dynamics at 500-msec resolution. J. Magn. Reson. Imaging; 2008; 27, pp. 599-606. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/18219630][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2900800]

33. Wang, Y et al. Cerebrovascular activity is a major factor in the cerebrospinal fluid flow dynamics. Neuroimage; 2022; 258, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35688316]119362.

34. Fillingham, P et al. Patient-specific computational fluid dynamic simulation of cerebrospinal fluid flow in the intracranial space. Brain Res.; 2022; 1790, 1:CAS:528:DC%2BB38XhsFKjsLjE [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35667413]147962.

35. Vandenbulcke, S; De Pauw, T; Dewaele, F; Degroote, J; Segers, P. Computational fluid dynamics model to predict the dynamical behavior of the cerebrospinal fluid through implementation of physiological boundary conditions. Front. Bioeng. Biotechnol.; 2022; 10, 1040517. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36483773][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9722737]

36. Liu, X. Decoupling between brain activity and cerebrospinal fluid movement in neurological disorders. J. Magn. Reson. Imaging; 2024; 60, pp. 1743-1752. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37991132]

37. Jiang, D et al. Regional glymphatic abnormality in behavioral variant frontotemporal dementia. Ann. Neurol.; 2023; 94, pp. 442-456.1:CAS:528:DC%2BB3sXht1KrtLnJ [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37243334][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10657235]

38. Wang, Z et al. Reduced coupling of global brain function and cerebrospinal fluid dynamics in Parkinson’s disease. J. Cereb. Blood Flow Metab.; 2023; 43, pp. 1328-1339.1:CAS:528:DC%2BB3sXhsFagurfO [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36927139][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10369155]

39. Zhang, Y et al. Reduced coupling between the global blood-oxygen-level-dependent signal and cerebrospinal fluid inflow is associated with the severity of small vessel disease. Neuroimage Clin.; 2022; 36, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36252555][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9668594]103229.

40. Zhang, Y et al. Reduced coupling between global signal and cerebrospinal fluid inflow in patients with depressive disorder: A resting state functional MRI study. J. Affect. Disord.; 2024; 354, pp. 136-142. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38484877]

41. Hiscox, L; Schwarb, H; McGarry, M; Johnson, C. Aging brain mechanics: Progress and promise of magnetic resonance elastography. Neuroimage; 2021; 232, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33617995]117889.