Comparing cerebrovascular reactivity measured using BOLD and cerebral blood flow MRI: The effect of basal vascular tension on vasodilatory and vasoconstrictive reactivity

Abstract

Cerebrovascular reactivity (CVR) is an important metric of cerebrovascular health. While the BOLD fMRI method in conjunction with carbon-dioxide (CO₂) based vascular manipulation has been the most commonly used, the BOLD signal is not a direct measure of vascular changes, and the use of arterial-spin labeling (ASL) cerebral blood flow (CBF) imaging is increasingly advocated. Nonetheless, given the differing dependencies of BOLD and CBF on vascular baseline conditions and the diverse CO₂ manipulation types currently used in the literature, knowledge of potential biases introduced by each technique is critical for the interpretation of CVR measurements. In this work, we use simultaneous BOLD-CBF acquisitions during both vasodilatory (hypercapnic) and vasoconstrictive (hypocapnic) stimuli to measure CVR. We further imposed different levels of baseline vascular tension by inducing hypercapnic and hypocapnic baselines, separately from normocapnia by 4 mm Hg. We saw significant and diverse dependencies on vascular stimulus and baseline condition in both BOLD and CBF CVR measurements: (i) BOLD-based CVR is more sensitive to basal vascular tension than CBF-based CVR; (ii) the use of a combination of vasodilatory and vasoconstrictive stimuli maximizes the sensitivity of CBF-based CVR to vascular tension changes; (iii) the BOLD and CBF vascular response delays are both significantly lengthened at predilated baseline. As vascular tension can often be altered by potential pathology, our findings are important considerations when interpreting CVR measurements in health and disease.

Introduction

Cerebrovascular reactivity (CVR) is an important metric of cerebrovascular health. Experimentally, CVR is defined as the degree of vasodilatation or constriction in response to a vasoactive agent or task. CVR is an indication of vascular reserve (Ito et al., 2003) and autoregulatory efficiency (Nur et al., 2009), and CVR impairment has been observed in cerebral steno-occlusive vascular diseases (Mandell et al., 2011; Mandell et al., 2008), lacunar infarction (Birns et al., 2009; Mandell et al., 2011), microbleeding (Birns et al., 2009; Conijn et al., 2012) and cognitive decline in aging as well as dementia (Hurford et al., 2014; Kastrup et al., 1998; Kovács et al., 2010; Suri et al., 2014).

BOLD fMRI during carbon-dioxide (CO₂)-based vascular challenges is the most common CVR imaging approach (Kassner et al., 2010), and has been extensively cross-validated (Herzig et al., 2008). CO₂ is a potent vasodilator, triggering changes in vascular tension through the arterial baroreflex (Ainslie et al., 2008), and the vascular response to CO₂ is well established using transcranial Doppler ultrasound of arterial blood flow (Battisti-Charbonney et al., 2011). Typically, CVR is measured as the ratio between changes in the BOLD fMRI signal and endtidal CO₂ (PETCO₂) change, a surrogate for arterial CO₂ (Robbins et al., 1990).

While the BOLD fMRI method has been the most commonly used due to its high availability and ease of implementation, a well-known concern regarding this approach is that the BOLD signal is not a direct measure of cerebral blood flow (CBF). This concern has been a primary motivation for adopting CBF-based CVR measurement (Noth et al., 2006; Tancredi et al., 2012). While a linear relationship between BOLD- and CBF-based CVR was obtained during hypercapnic vasodilation in healthy controls (Mark et al., 2010) as well as patients (Mandell et al., 2008), these previous experiments imposed iso-oxia during hypercapnic challenges, not the case in the majority of CVR studies. As the BOLD and CBF signals are differentially sensitive to oxygen level changes (Bulte et al., 2007), failure to maintain iso-oxia could introduce biases in the CVR measurements, and the linearity would deteriorate. More importantly, this relationship is still unknown for vasoconstrictive paradigms. In fact, the agreement between BOLD- and CBF-based CVR measurements is potentially dictated by basal vascular tension and type of vascular challenge, as BOLD and CBF signals have differing dependences on baseline blood flow and vascular volume, among other variables (Bhogal et al., 2014).

Commonly, CVR-related vascular changes are induced through manually adjusted administration of blended gasses (Bandettini and Wong, 1997; Cohen et al., 2004; Yezhuvath et al., 2009), endtidal forcing (Poulin et al., 1996) or computerized prospective PETCO₂ targeting (Conklin et al., 2011; Han et al., 2011; Mandell et al., 2011; Mandell et al., 2008; Mark et al., 2010; Mikulis et al., 1989; Mutch et al., 2012; Prisman et al., 2008; Slessarev et al., 2007; Spano et al., 2013; Vesely et al., 2001). More recently, breath-holding-based hypercapnia is seen as an increasingly viable alternative (Bright and Murphy, 2013; Murphy et al., 2011). In addition, deep-breathing hypocapnia (Bright et al., 2009; Sousa et al., 2014) has also been proposed as an alternative vascular manipulation. Hypercapnia and hypocapnia result in vasodilation and vasoconstriction, respectively. However, these tasks do not furnish equivalent CVR measurements, as shown by Tancredi and Hoge (2013). Nonetheless, the relationship between simultaneous BOLD- and CBF-based CVR measurements has yet to be clarified for these widely varying types of vascular manipulations. Such knowledge is particularly useful in the context of altered vascular tension, as BOLD and ASL have diverging dependences on baseline conditions.

Specifically, vascular tension is an important determinant of the degree of vasoconstriction or vasodilation a blood vessel is capable of, and can be altered by age and neuropathology (Lüscher et al., 1991). Thus, understanding the relationship between CVR measurements and vascular tension would help interpret CVR measurements in patients as well as increase CVR data reliability in healthy controls. For instance, in post-stroke individuals, Zhao et al. (2009) reported a significantly lower increase in CBF during hypercapnic challenges but a greater reduction of CBF during hypocapnic challenges than the control group, which may have been attributable to stroke-related long-term alterations in vascular tension. In support of this theory, Ito et al. (2008) reported higher capacity for CBF reduction in a vasodilated state relative to the resting baseline in healthy controls, whereas Bright et al. (2010) identified a reduction in BOLD signal increase during a predilated baseline. The role of baseline vascular tone is more systematically demonstrated by Ghariq et al. (2014), who found that the BOLD CO₂ reactivity data fit better to a sigmoidal curve whereas the CBF versus PETCO₂ data fit better to a linear model. Nonetheless, to the best of our knowledge, there has yet to be a comparison of simultaneously measured BOLD- and CBF-based CVR for various vascular-tension conditions. CVR response delay is also a physiologically meaningful parameter (Blockley et al., 2011; Bright et al., 2010) that may become increasingly useful in assessing vascular dysfunction (Poublanc et al., 2013), but there is also limited information on the dependence of CVR response delay on vascular tension.

In this study, we use a dual-echo pseudocontinuous ASL technique to simultaneously assess the behavior of BOLD- and CBF-based CVR measurements under different vascular tensions and in response to different vascular stimuli. The results demonstrate that BOLD- and CBF-based CVR amplitude measurements are differentially sensitive to vascular tension and to the type of vascular stimuli. Our findings also demonstrate that the CO₂ response delay is highly sensitive to baseline vascular tension for both BOLD- and CBF-based CVR measurements. These findings have important implications for comparing BOLD- and CBF-based CVR studies in normal and patient populations.

Materials and methods

Participants

This study was carried out with 18 healthy adult volunteers (10 male, age 26.3 ± 6.5 years; range 18 to 36 years). Participants were recruited through the Baycrest Participants Database, which includes participants from Baycrest and the surrounding community. The study was approved by the Baycrest Research Ethics Board.

MRI acquisition

All images were acquired using a Siemens TIM Trio 3 Tesla System (Siemens, Erlangen, Germany). The scans employed 32-channel phased-array head coil reception and body-coil transmission. A T1-weighted MPRAGE 3D anatomical was acquired at an isotropic resolution of 1 mm. In addition, CBF and BOLD data were acquired concurrently using a dual-echo pseudo-continuous pCASL technique (Dai et al., 2008), with a repetition time (TR) of 4 s, echo times (TE)1/TE2 = 10/25 ms, field of view = 220 × 220 mm, 18 slices (ascending interleaved order), voxel size = 3.4 × 3.4 × 5.0 mm³, 100 frames, bandwidth = 2520 Hz/pixel and GRAPPA = 2. The labeling duration was 1500 ms, and the post-labeling delay was 1000 ms with a mean G_z of 0.6 mT/m which was selected to achieve transit time insensitivity.

Vascular tension manipulation and vascular challenges

All vascular manipulations were achieved by administering mixtures of O₂, CO₂ and medical air delivered using the RespirAct^™ breathing circuit (Thornhill Research, Toronto, Canada), designed to provide computerized and independent targeting of endtidal O₂ (PETO₂) and CO₂ (PETCO₂) pressure using the sequential gas delivery method (Slessarev et al., 2007). This method was chosen to maximize steady-state PETCO₂-targeting accuracy and stability while minimizing PETO₂ confounds during CO₂-based CVR measurements (Chen and Pike, 2010b; Prisman et al., 2008). Within the RespirAct framework, while breath rate was self-regulated, hypocapnia is achieved primarily through increased breathing depths, while hypercapnia is achieved through inhaling air with a higher than normal amount of CO₂. For this method, the CO₂ and O₂ capacity as well as consumption rate for each subject were estimated prior to the scan in order to compensate for the physiology-related resting PETCO₂ and CO₂ clearance variations between individuals. The detailed experimental protocol is illustrated in Fig. 1.

Fig. 1. The experimental paradigm.

PETCO₂ was sinusoidally modulated with a period of 120 s; 3 periods of sinusoidal PETCO₂ variations were induced, following a 1-minute baseline. This bipolar stimulus includes both vasodilatory (upper half of sinusoid) and vasoconstrictive (bottom half of sinusoid) components. This sinusoidal manipulation was applied at the subject’s natural baseline (normocapnic) as well as at hypercapnic and hypocapnic baselines, which are separated from the normocapnic baseline by 4 mm Hg.

During the pCASL scans, PETCO₂ was sinusoidally modulated (Blockley et al., 2011) at each baseline PETCO₂ with a period of 120 s; 3 periods of sinusoidal PETCO₂ variations were induced, following a 1-minute baseline. We chose such a bipolar paradigm as it includes both vasodilatory (upper half of sinusoid) and vasoconstrictive (bottom half of sinusoid) components.

This sinusoidal manipulation was applied at the subject’s natural baseline (normocapnic) as well as at hypercapnic and hypocapnic baselines, which are both separated from the normocapnic baseline by 4 mm Hg. The ordering of the different baselines was randomized to minimize biases. These mild PETCO₂ changes result in slight changes in the subject’s vascular tone without noticeable changes in cerebral oxidative metabolism, a potential confound for CO₂-based vascular assessment (Chen and Pike, 2010a).

Image preprocessing

Functional images, including the tag and control images in the pCASL data, and T₁-weighted anatomical images were separately preprocessed using SPM8 (Wellcome Trust Centre for Neuroimaging, London, UK). In each of the tag and control series, the first four time frames were discarded to ensure MR steady state. Preprocessing for pCASL images included retrospective head-motion correction, slice-timing correction using sinc interpolation, spatial transformation into a Montreal Neurological Institute (MNI) space, and spatial smoothing with a 6-mm full-width at half-maximum (FWHM) Gaussian kernel. Anatomical images were co-registered with their corresponding realigned functional data and segmented into gray and white matter tissue probability maps using the unified segmentation procedures (Ashburner and Friston, 2005).

Data analysis

Analyses were performed using in-house software written in MATLAB (www.mathworks.com) and FSL (www.fmrib.ox.ac.uk/fsl).

Regions of interest

ROIs were derived from subject-specific native space FreeSurfer tissue parcellations (Fischl et al., 2004) (FreeSurfer software is publically available at http://surfer.nmr.mgh.harvard.edu). Binary masks were also generated for cortical regions, namely the frontal, temporal, parietal, limbic and occipital regions as well as the hippocampus. We also examined subcortical gray matter regions, including the amygdala, caudate, pallidum, putamen and thalamus. Additionally, we created ROIs for global gray matter and white matter. ROIs were then down-sampled to match the resolution of the fMRI data using FreeSurfer.

BOLD- and CBF-based CVR and CVR-delay calculations

BOLD and CBF time series were generated from the pCASL tag and control images. More specifically, the modulated CBF component, which is less affected by the BOLD-weighted tissue component, was extracted by high-pass filtering the pCASL signal, followed by demodulation, as introduced by Chuang et al. (2008). BOLD contamination in the CBF time series was minimized as described by Tak et al. (2014).

Upon obtaining these time series data, we applied three levels of data control to boost the reliability of the CVR values used in the statistical analysis: (i) control for image-intensity; (ii) control for temporal outliers; and (iii) control for outliers in ROIs.

• 1Control for image intensity values: Due to the slightly lower signal-to-noise ratio (SNR) of the ASL data in comparison to the BOLD data, we excluded voxels with insufficient SNR by stipulating that at least 60% of the time course has physically positive tag-control difference, given that all data are acquired from healthy adults thus are unlikely to produce negative CBF values. Subsequently, physically negative CBF data was removed and both the CBF and BOLD signals are demeaned to obtain the percent change in the signal relative to the baseline (mean) of the signal.
• 2Control for temporal outliers: For each voxel, outlier removal was performed by removing points from the time courses that fulfill Student’s criterion or a Cook’s distance greater than 4/N, where N is the number of time points. A linear model was applied to regress the PETCO₂ time course and the voxel-wise BOLD or CBF time course, and CVR was calculated as the slope of the linear model. The spatially specific CVR was thresholded at 0 to minimize biases introduced by noise. Additionally, we excluded data associated with any spiking in the PETCO₂ time course.
• •In calculating CVR, the CO₂-response delay in BOLD and CBF time courses were taken into account at each voxel. This delay was estimated from the phase difference between the BOLD or CBF time courses and that of PETCO₂, enabling the alignment of the fMRI and PETCO₂ time series for maximal correlation. For this alignment, both PETCO₂ and fMRI data were upsampled to a common frequency, such that delays that are shorter than TR could be measured. Following the alignment, the PETCO₂ was interpolated to the pCASL TR and detrended at half the total duration of the respiratory paradigm.

  • Response to bipolar stimulus: derived by the linear fit of both the upper and lower parts of the MRI signal response to sinusoidal PETCO₂ variations;

  • Response to vasodilatory stimulus: derived by the linear fit of the part of the BOLD or CBF signal associated with the upper portion of the PETCO₂ sinusoid;

  • Response to vasoconstrictive stimulus: derived by the linear fit of the part of the BOLD or CBF signal associated with the lower portion of the PETCO₂ sinusoid;

• 3Control for outliers in ROIs: We computed the regional mean CVR and standard deviation across each ROI. Specifically, in order to generate representative regional CVR values unbiased by outliers, we first removed outlier voxels from each ROI using the non-parametric algorithm based on the Tukey’s box-plot method (Patterson, 2012; Tukey, 1977); this method suits the present application as it does not assume a normal distribution for the voxel-wise CVR values. Subsequently, we performed linear regression to assess the relationship between BOLD- and CBF-based CVR and CVR delay for all ROIs.

Statistical analysis

The inter-subject mean CBF-derived and BOLD-derived CVR were regressed for the brain lobes and whole brain gray matter (GM) ROIs. Before regression, we implemented the first 2 levels of data control, and removed the outlier data points based on Student’s criterion. In addition, we used multi-variate and univariate ANOVA as well as Student’s t-test to assess the dependence of CVR and CVR-delay values on

• The level of vascular tension: modulated through a hypercapnic, normocapnic (resting) and hypocapnic baseline;

• The type of vascular stimulus: vasodilatory, vasoconstrictive and bipolar vasodilation–constriction stimulus.

All 3 levels of data control were implemented prior to the ANOVA and t-tests.

Results

The PETCO₂ levels corresponding to the three vascular tension levels are shown in Fig. 2a. Prospective targeting permitted us to realize our designated PETCO₂ changes, such that the mean PETCO₂ was 36.2 ± 1.6, 40.1 ± 1.4 and 44.1 ± 1.5 mm Hg for the hypocapnic, normocapnic and hypercapnic baselines, respectively. The corresponding PETO₂ values were 113.5 ± 1.8, 113.0 ± 3.3 and 112.5 ± 4.7 mm Hg, respectively. The sinusoidal PETCO₂ modulation (Fig. 2b) induced robust CBF and BOLD signal changes (sample data shown in Fig. 2c).

Fig. 2. Sample PETCO₂ and fMRI data.

(a) The PETCO₂ levels corresponding to the three vascular tension levels are shown with error bars indicating standard error across subjects. (b) Prospective targeting permitted us to realize our designated PETCO₂ changes of approximately 4 mm Hg in either direction, relative to the participant’s natural baseline. (c) The sinusoidal PETCO₂ modulation (b) induced robust CBF and BOLD signal changes.

Agreement between BOLD- and CBF-based CVR

CVR maps were successfully generated using both BOLD and CBF data, as shown in Fig. 3, with BOLD maps exhibiting superior signal-to-noise (SNR) than CBF. BOLD- and CBF-based CVR measurements were significantly correlated in the majority of brain regions. This was true for the CVR amplitude (Fig. 4) as well as response delay (Fig. 5). In spite of the existence of linear relationships between BOLD and CBF-based CVR measurements, these measurements have different sensitivities to vascular tone and to the nature of the vascular stimulus.

Fig. 3. Sample BOLD and CBF CVR maps across all vascular tension levels.

BOLD-based CVR maps in response to the bipolar sinusoidal stimulus. Qualitatively lower signal-to-noise ratios (SNRs) can be discerned in the CBF-based CVR maps compared to the BOLD, and the response at the hypocapnic (preconstricted) baseline appears to be the lowest compared to the other two basal vascular tension levels.

Fig. 4. The relationship between BOLD- and CBF-based CVR measurements across main regions-of-interest (ROIs).

Red, black and blue represent hypercapnic, normocapnic and hypocapnic baselines, respectively. In most regions, the BOLD and CBF CVR values were significantly correlated. Note that the plotted values correspond to regional averages prior to ROI-outlier removal.

Fig. 5. Relationship between BOLD- and CBF-based CVR delay measurements across ROIs.

Again, red, black and blue represent hypercapnic, normocapnic and hypocapnic baselines, respectively. The BOLD and CBF CVR delay values were significantly correlated in most brain regions.

The effect of vascular stimulus type on CVR measurement

At normocapnic (resting) vascular tension, after performing all 3 levels of data control, we found the group-average BOLD CVR to be 0.05 ± 0.005% BOLD/mm Hg in the cortex and 0.03 ± 0.002% BOLD/mm Hg in the white matter, taken across all participants. In the case of CBF-based CVR, the cortical and white matter group averages were 1.9 ± 0.4 and 1.1 ± 0.3% CBF/mm Hg, respectively. With regards to the effect of the choice of vascular stimulus, the results are illustrated in Fig. 6, and the ANOVA assessments in Table 1. When all vascular tension levels were considered, we found no significant distinction in the BOLD-based measurements, i.e. BOLD CVR = 0.04 ± 0.007, 0.03 ± 0.007 and 0.05 ± 0.008% BOLD/mm Hg, respectively, as the cortical average for vasodilatory, vasoconstrictive and bipolar stimuli. However, for CBF, the vasodilatory stimulus to elicit higher responses was comparable to vasoconstrictive and bipolar stimuli in multiple brain regions, including the frontal, limbic and occipital lobes as well as the amygdala and hippocampus (p < 0.05).

Fig. 6. The dependence of BOLD and CBF CVR measurements on the type of vascular stimulus.

Compared to CBF measurements, BOLD measurements of CVR were insensitive to the type of stimulus employed. The asterisks indicate statistically significant trends and differences (based on ANOVA results). The plotted values correspond to regional averages after ROI-outlier removal.

Table 1. The dependence of CVR estimate amplitude on basal vascular tension and the type of vascular challenge.

Summarized here are assessments by the multi-variate ANOVA. We did not find significant interactions between vascular tension and the type of vascular stimulus.

	BOLD	CBF
Effect of vascular stimulus type
Cortex	–	Frontal lobe (p = 0.05)
		Temporal lobe (p = 0.04)
		Occipital lobe (p = 0.04)
Subcortex	–	Hippocampus (p = 0.05)
		Amygdala (p = 0.01)
White matter	–	–
Effect of basal vascular tension: bipolar stimulus
Cortex	Frontal lobe (p = 0.04)	Limbic lobe (p = 0.008)
	Limbic lobe (p = 0.04)	Parietal lobe (p = 0.01)
	Parietal lobe (p = 0.05)	Temporal lobe (p = 0.02)
	Temporal lobe (p = 0.009)	Occipital lobe (p = 0.01)
	Occipital lobe (p = 0.009)	Whole cortex (p = 0.02)
	Whole cortex (p = 0.02)
Subcortex	Amygdala (p = 0.05)	Putamen (p = 0.03)
	Thalamus (p = 0.05)	Thalamus (p = 0.05)
White matter	All regions (p = 0.04)	All regions (p = 0.02)
Effect of vascular tension: vasodilatory stimulus
Cortex	Temporal lobe (p = 0.05)	Parietal lobe (p = 0.05)
Subcortex	Thalamus (p = 0.01)	–
White matter	–	–
Effect of vascular tension: vasoconstrictive stimulus
Cortex	Frontal lobe (p = 0.05)	Temporal lobe (p = 0.05)
	Temporal lobe (p = 0.05)
	Occipital lobe (p = 0.04)
	Whole cortex (p = 0.05)
Subcortex	–	–
White matter	–	–

The effect of baseline vascular tension on CVR measurement

The effect of vascular tension on CRV estimation is also summarized in Table 1. When considering both vasodilatory and vasoconstricting stimuli (Fig. 7), we found that BOLD-based CVR was highest during hypercapnic (predilated) baseline, followed by hypocapnic (preconstricted) baseline, with normocapnic (resting) baseline being associated with the lowest CVR values. Similar trends were also observed for CBF CVR measurements throughout the cortex and deep-gray structures. The white matter did not show a significant vascular-tension effect.

Fig. 7. The dependence of BOLD and CBF CVR measurements on vascular baseline (basal vascular tension): bipolar stimulus.

BOLD and CBF were similarly sensitive to vascular tension, although different behaviors are observed. The asterisks indicate statistically significant trends and differences (based on ANOVA results). The plotted values correspond to regional averages after ROI-outlier removal.

We illustrate the dependence of the vasodilatory response to vascular tension in Fig. 8. In regard to BOLD CVR, we found that a normocapnic baseline was associated with the highest vasodilatory response, with both predilation (hypercapnic baseline) and preconstriction (hypocapnic baseline) both leading to a lower vasodilatory CVR. This was a significant effect in the temporal lobe and thalamus. In contrast, CBF-based vasodilatory CVR measurements were, highest in the predilated and resting baselines (at similar levels), and lowest in the preconstricted state. This effect was significant in the parietal lobe alone.

Fig. 8. The dependence of BOLD and CBF CVR measurements on vascular baseline (basal vascular tension): vasodilatory stimulus.

The normocapnic (resting) baseline was generally associated with a higher BOLD response to vasodilatory stimulation, while the effect is much less pronounced in CBF. The asterisks indicate statistically significant trends and differences (based on ANOVA results). The plotted values correspond to regional averages after ROI-outlier removal.

The influence of vascular tension on the vasoconstrictive stimulus is illustrated in Fig. 9. The predilated state was associated with higher BOLD and CBF responses than resting and preconstricted baselines. These effects were dominant in the cortex (frontal, temporal and occipital regions) and not significant in the subcortical regions. Once again, CBF-based CVR measurements exhibited less vascular-tension dependence than BOLD-based CVR.

Fig. 9. The dependence of BOLD and CBF CVR measurements on vascular baseline (basal vascular tension): vasoconstrictive stimulus.

For both BOLD and CBF, the hypercapnic (predilated) baseline was associated with a higher CVR. The asterisks indicate statistically significant trends and differences (based on ANOVA results). The plotted values correspond to regional averages after ROI-outlier removal.

The effect of baseline vascular tension on CO₂ response delay

Finally, the impact of basal vascular tension on the CO₂ response delays is summarized in Fig. 10 and Table 2. We opted to use the response to the bipolar stimulus alone in order to maximize the accuracy of delay estimation. Across all baseline conditions, we found BOLD-based CVR delay to be 10.4 ± 1.4 s for the frontal lobe, 10.8 ± 1.2 s for the parietal lobe, 10.9 ± 1.2 s for the temporal lobe, 9.9 ± 1.1 s for the limbic lobe, 11.0 ± 1.1 s for the occipital lobe and 11.7 ± 1.2 s for the thalamus. A similar regional variation was seen in the CBF-based estimate, although more pronounced. The average BOLD-based CVR delay was 10.2 ± 1.1 s in the cortex and 10.6 ± 1.2 s in the white matter, longer than the average CBF-based cortical and white-matter delays of 9.2 ± 1.7 s and 9.5 ± 1.9 s, respectively. During the predilated baseline, these delays increased significantly relative to the resting baseline, by as much as 8.2 s across all brain regions. In fact, we found that the CO₂ response delay was the largest at the predilated baseline, while the difference between preconstricted and resting baselines was not statistically significant; this was true for both BOLD and CBF.

Fig. 10. The dependence of BOLD and CBF CVR delay measurements on vascular baseline (basal vascular tension).

For both BOLD and CBF, the predilated (hypercapnic) baseline was associated with the longest CO₂ response delays. The asterisks indicate statistically significant trends and differences (based on ANOVA results). The plotted values correspond to regional averages after ROI-outlier removal.

Table 2. The dependence of CVR delay estimates on baseline vascular tone.

assessment by univariate ANOVA.

Effect of basal vascular tone	BOLD	CBF
Cortex	Frontal lobe (p = 2e-6)	Frontal lobe (p = 2e-5)
	Limbic lobe (p = 7e-7)	Limbic lobe (p = 3e-5)
	Parietal lobe (p = 3e-6)	Parietal lobe (p = 0.0008)
	Temporal lobe (p = 2e-7)	Temporal lobe (p = 0.0003)
	Occipital lobe (p = 4e-8)	Occipital lobe (p = 0.0005)
Subcortex	Amygdala (p = 0.0006)	Hippocampus (p = 0.002)
	Hippocampus (p = 4e-6)	Pallidum (p = _0.001)
	Pallidum (p = 6e-5)	Putamen (p = _0.0003)
	Putamen (p = 7e-6)	Thalamus (p = 0.0002)
	Thalamus (p = 5e-6)
White matter	All regions (p = 2e-6)	All regions (p = 5e-7)

Discussion

In this work, we used simultaneous BOLD-CBF measurements during vasodilatory and vasoconstrictive tasks to measure CVR. These tasks were in the form of PETCO₂ manipulations, well known for inducing robust intravascular CO₂ and blood-flow changes. We further imposed three levels of baseline vascular tension to emulate arterial pressure changes characteristic of normal vascular variability and potential disease conditions. In each individual, we modulated vascular tension through different capnic states, focusing on mild vascular-tension changes so as to avoid inducing non-vascular confounds. Even within this small range of variations, we saw significant and diverse dependencies in both the BOLD and CBF measurements of CVR. These findings are expected to be important considerations when interpreting CVR measurements.

The effect of vascular stimulus type on CVR measurement

The first main message of this work is that CBF-based CVR measurements are far more sensitive to the type of vascular stimulus than BOLD-based CVR, with higher CVR during vasodilatory tasks. Indeed, when comparing BOLD-based CVR measurements made using breath-holding hypercapnia and deep-breathing hypocapnia, Bright et al. (2009) noted that vasoconstrictive responses are generally lower than vasodilatory responses. A similar trend was described by Tancredi and Hoge (2013), with hyperventilation-induced hypocapnia resulting in lower CBF responses than either breathholding or prospectively-targeted hypercapnia. Such observations are consistent with the sigmoidal model of the dependence of blood flow on CO₂ (Battisti-Charbonney et al., 2011; Claassen et al., 2007; Tancredi and Hoge, 2013), derived based on transcranial Doppler data from healthy adults, describing diminished arterial blood-flow response sensitivity to CO₂ below and above a certain PETCO₂ range. This model was recently validated by Sobczyk et al. (2014) using BOLD data. We attribute the lack of similar effects in our BOLD data to our particular vascular tasks, which induce much smaller PETCO₂ changes than expected from previous studies. Surprisingly however, the bipolar stimulus elicited lower CBF CVR than either the vasodilatory or the vasoconstrictive stimulus. This was a robust finding in all brain regions examined, and indicates that CBF CVR values obtained from different stimulus types should not be compared. This finding could be explained by the lower CBF vasoconstrictive reactivity compared to the vasodilatory reactivity, potentially due to SNR limitations in ASL data in low-flow conditions. Such an experimental limitation would potentially result in a non-linear CBF dependence on PETCO₂ from low to high values, and fitting a linear model to such data would result in a shallower slope than fitting to the upper or lower half alone.

The effect of baseline vascular tension on CVR measurement

The second main message of this paper is that it is important to be aware of heterogeneity in basal vascular tension when comparing CVR measurements across individuals. Given the disparities in the dependence of the BOLD and CBF signals on vascular baseline, vascular tension is key to the interpretation of MRI-based CVR measurements.

When using the bipolar stimulus, the BOLD CVR values were generally lower at the preconstricted (hypocapnic) baseline (Figs. 7a). A similar trend was also observed in the CBF data. This is in fact also consistent with predictions of the sigmoidal model, describing diminished arterial blood-flow response sensitivity to CO₂ below and above a certain PETCO₂ range. Vascular tension reflects arterial pressure, and both the predilated and preconstricted baselines are associated with elevated arterial blood pressure (Carrera et al., 2009). These in turn modulate the ability of a blood vessel to dilate and contract passively by the change in arterial blood pressure. Using transcranial Doppler ultrasound, Carrera et al. (2009) found CVR to be reduced in the presence of increased arterial pressure, being significantly higher in a mildly hypocapnic state. Furthermore, our findings agree with those of Tancredi and Hoge, showing that even moderately hypocapnic baseline can bring CBF-CVR into the non-linear CO₂ response regime, resulting in significantly diminished CVR compared to normocapnic and mildly hypercapnic conditions.

In practice, most existing CVR studies have not used bipolar stimuli, but rather vasodilatory stimuli alone (Kassner et al., 2010; Mandell et al., 2011; Mandell et al., 2008; Mark et al., 2010). Our finding of the vascular-tension dependence of BOLD and CBF vasodilatory CVR is relevant to the majority of CVR results. The higher BOLD dilatory response in the resting baseline compared to at the predilated baseline is in agreement with findings by Bright et al. (2010), and is potentially attributable to altered vascular compliance with arterial pressure changes. We note that our finding of a higher BOLD-CVR at normocapnia than at hypercapnia differs from the pattern observed in CBF-CVR estimates. This may be a reflection of the physical differences in the origins of the BOLD and CBF signals, implying that BOLD- and CBF-CVR likely do not share the same linear and non-linear regimes, as will be discussed in a later section. Interestingly, within the range of PETCO₂ values used in our study, the CBF dilatory response was slightly less affected by vascular tension, suggesting that CBF CVR measurements are potentially more robust than BOLD.

There is an increasing interest to use hypocapnia for CVR measurement, particularly as the recently proposed cued deep-breathing paradigm is more patient-friendly than existing hypercapnia paradigms (Bright et al., 2009; Sousa et al., 2014). In this regard, a significant trend towards higher vasoconstrictive BOLD CVR at the predilated baseline was observed. A similar but less significant trend was observed in CBF measurements. Such trends support previous findings (Bright et al., 2010; Ito et al., 2008). Ito et al. used a spectrum of intrinsic variations in resting CBF to investigate the dependence of CBF response on vascular baseline, and found that in most brain regions, a higher basal CBF was associated with a higher vasoconstriction capacity and a lower vasodilatory capacity. Similar findings were reported by Liu et al. (2012) in large cerebral arteries.

The sensitivity of CBF CVR to vascular tension is lower but in keeping with findings by Ito et al. We note, nonetheless, that Ito et al. reported considerable spatial dependence in this relationship, and in fact, the frontal and occipital lobes seemed to show opposite trends. We noted a substantial degree of regional heterogeneity in the trends observed in our data as well, which may obscure the global trend. We also note that in our study, instead of relying on cross-sectional basal CBF differences, we modulated basal vascular tension in each individual through computerized CO₂ targeting. We argue that our approach minimizes the confounding influence of potential neurovascular properties (such as CVR itself) that may have led to these inter-subject CBF differences in the first place. These methodological differences may help to explain some of the differences across studies. Last but not least, the insensitivity of dilatory CBF reactivity to vascular tension may also be attributed to the lower SNR in the CBF data. Paradoxically, in our data, the hypocapnic baseline did not result in a higher vasodilatory BOLD reactivity compared to the normocapnic baseline. As described earlier, such behavior may indicate that the arterial blood pressure increase associated with our hypocapnic baseline condition has begun to compromise arterial compliance.

The effect of baseline vascular tension on CO₂ response delay

In this work, we used the sinusoidal stimulus proposed by Blockley et al. (2011). However, we did not use the frequency spectrum-based technique for response delay estimation, as low-frequency physiological noise interfered within the frequency bandwidth of our stimulus. Rather, our delay-estimation method is similar to that of Thomas et al. (2014). Nonetheless, our findings of earlier response in the frontal lobe and later response in the occipital lobe as well as the subcortical gray matter structures are in agreement with previous findings (Blockley et al., 2011; Bright et al., 2009). We do not quantitatively compare our delay estimates with those of Blockley et al. as the latter results were obtained at 7 T.

Our third major finding in this work is that vascular response delay is longest at a predilated vascular baseline. We observed the longest delay at the hypercapnic baseline in both BOLD and CBF data. As the CVR response to the vasodilatory stimulus is generally higher, the delay estimates may have been dominated by the vasodilatory response, and thus show the CO₂ response to be more sluggish when the vessels are predilated. Such a finding could be evidence of the behavior described by the arteriolar-compliance model (Behzadi and Liu, 2005), which predicts that the baseline vascular state would affect CVR by changing the relative contributions of an active smooth muscle component and a passive connective tissue component to the overall vascular reactivity. According to this model, one would expect the vascular response to any vasoactive agent to be slower in a predilated state, as the slower passive-tissue component dominates at larger vascular radii. A potential alternative explanation is that as basal vascular diameter increases, more incoming blood is required to achieve a certain fractional vascular signal increase, resulting in the appearance of an increasing BOLD and CBF CVR delay. However, a contradicting observation to both theories is that the preconstricted baseline did not result in a significant reduction in CVR delay. Further investigation is necessary to clarify the nature of our observation.

Basal vascular tension variations in health and disease

In the absence of external stimuli, CBF stays constant over a range of mean arterial pressure (MAP) values as part of cerebral autoregulation (Paulson et al., 1990). However, vascular diameter and vascular tension can be modulated in a dose-dependent fashion by vasoactive agents such as caffeine (Rack-Gomer and Liu, 2012). Moreover, the arterial CO₂ fluctuations introduced in this study are a part of normal daily activities, resulting in a major source of vascular tension changes. According to the detailed characterization by Battisti-Charbonney et al., the linear regime for vasodilation-mediated blood flow velocity response to CO₂ in a healthy adult can span no more than 4 mm Hg above and below normocapnia, beyond which MAP will no longer remain unchanged. Thus, as our multi-baseline and sinusoidal stimulation approach probes PETCO₂ values up to 8 mm Hg above and below normal baseline, we involve both the constant-MAP and variable-MAP regimes of CBF regulation. As shown by Carrera et al., a ΔPETCO₂ of 8 mm Hg could induce a MAP increase of up to 10%, whether for hypercapnia or hypocapnia. A similar degree of MAP reduction can result from less than 30 min of physical exercise (Pescatello et al., 1991; Sun et al., 2014).

More importantly, vascular tension is a common indicator of vascular health. Hypertension entails increased MAP, reduced vascular compliance, and is widely identified with atherosclerosis, diabetes, heart disease and obesity (Benetos, 2008; van Popele et al., 2001). In diabetes mellitus, for instance, MAP is on average 10% higher than in age-matched controls, whereas in clinically hypertensive patients, MAP can measure up to 40% higher (Megnien et al., 1992). The hypertension and vascular lumen narrowing seen in atherosclerosis can be simulated by our hypocapnic baseline in terms of the hypoperfusion and reduced vasoconstrictive compliance, with the complication that vasodilatory response is also compromised due to vascular stiffening in atherosclerosis (van Popele et al., 2001). On the other hand, cerebral hyperperfusion syndrome, which has been associated with seizures (Ho et al., 2000), is analogous to our hypercapnic baseline. Even healthy aging, older adults have been associated with an average MAP of up to 4% higher than middle-aged adults (Sesso et al., 2000).

BOLD-based vs. CBF-based CVR

We note that after ROI-wise outlier removal, our mean BOLD-based CVR values reported here became much lower than those found in the literature (e.g. cortical mean CVR reduced from 0.13 to 0.05% BOLD/mm Hg at normocapnia). For instance, using various hypercapnic manipulations, Stefanovic et al. (2006) (1.5 T), Yezhuvath et al. (2009) (3 T), Kassner et al. (2010) (1.5 T) and Mark et al. (2010) (3 T) listed BOLD CVR values of 0.1–0.2% BOLD/mm Hg, while Bright et al. (2009) reported values up to 0.5%/mm Hg using hypocapnia. We note that Kassner et al. reported a mean normocapnic cortical BOLD CVR of 0.14%/mm Hg using a precise PETCO₂ targeting paradigm very similar to our own, but upon closer inspection, the CVR maps shown by Kassner et al. seem to reflect parenchymal CVR values much lower, primarily between 0.01 and 0.025%/mm Hg (at 1.5 T). Similarly low values were shown by Vesely et al. (2001), more akin to our own findings of mean CVRs of 0.05% BOLD/mm Hg at 3 T. The most logical explanation is that the reported high mean-CVR values were driven by large vessel effects in BOLD, which we were able to suppress through our ROI-wise outlier exclusion strategy. Thus, our measurements rather reflect the parenchymal contribution to the BOLD CVR measurements. Incidentally, CBF-based CVR values did not seem to be as sensitive to the ROI-wise outlier exclusion, suggesting that CBF-based measurements are less prone to large-vessel effects, as they were less weighed towards veins. This is an important distinction between BOLD and CBF that will likely become critical in distinguishing parenchymal and vascular health.

By nature, CBF measured using ASL takes into account both vasodilatory CBF and velocity-driven CBF increases. This is not the case with BOLD. Based on the BOLD signal model (Davis et al., 1998; Hoge et al., 1999), the BOLD signal increases due to the dilution of deoxyhemoglobin by CBF increases. Thus, when deoxyhemoglobin dilution reaches a maximum, further increases in CBF would not be reflected in the BOLD signal. Conversely, when the lowest deoxyhemoglobin-concentration steady state has been reached, no further reductions in CBF would be reflected in BOLD. These scenarios were compellingly illustrated by Sobczyk et al. (2014) in patients with steno-occlusive disease. Both of these states would entail MAP changes (see earlier discussion), to which BOLD is theoretically insensitive. Thus, analogous to the case presented by Regan et al. (2014), BOLD is likely to underestimate CVR when MAP is not constant. We simulated these situations using the vasoconstrictive stimulus in the hypocapnic baseline as well as vasodilatory stimulus in hypercapnic baseline, and indeed observed such an effect, as illustrated in Fig. 11. BOLD-based CVR to the vasodilatory stimulus was lower at hypercapnic baseline compared to normocapnic baseline, but CBF maintained a larger linear range with increasing PETCO₂, consistent with ultrasound recordings (Battisti-Charbonney et al., 2011). However, we did observe a trend of lower CBF response to vasoconstrictive than to vasodilatory stimulus, consistent with previous MRI findings (Tancredi and Hoge, 2013). We attribute this primarily to potential transit-time effects in the ASL technique, and secondarily to SNR limitations in low-flow conditions. Potential remedies include novel techniques such as velocity-selective ASL (Wu and Wong, 2006).

Fig. 11. Conceptualized illustration of CVR measurement dependence on basal vascular tension.

The key features of the plots are marked with numbers, and explained in the right half of the figure. The blue and pink shading represent the PETCO₂ spans associated with the sinusoidal manipulation at the hypocapnic (preconstricted) and hypercapnic (predilated) baselines, respectively. In this work, BOLD-based response to the vasodilatory stimulus was lower at hypercapnic baseline compared to normocapnic baseline, consistent with a nonlinear BOLD response to changing baseline PETCO₂. However, CBF maintained a larger linear range with increasing PETCO₂.

Recommendations and significance

First, as the dependence of CVR measurements on baseline vascular tension becomes an important consideration when PETCO₂ enters the non-linear regime (Battisti-Charbonney et al., 2011), our results reinforce support for the recommendation by Tancredi and Hoge that CVR studies should be conducted within the linear range of PETCO₂ dependence, which is above 30 and below 45 mm Hg (Battisti-Charbonney et al., 2011; Tancredi and Hoge, 2013). However, even with the small range of PETCO₂ variations used in our study, we still observed vascular-tension dependence in the BOLD CVR measurements. On that note, BOLD and CBF, due to their fundamental differences in signal origin, are associated with different linear and non-linear PETCO₂ regimes. In fact, as shown very recently by Ghariq et al. (2014), the linear regime of the CBF CVR measurements may be more extensive than that of BOLD, which was demonstrated as not being solely driven by the differential SNR between BOLD and CBF. Thus, our results support the view that CBF may provide somewhat less biased CVR measurements than BOLD.

Secondly, vasodilatory and vasoconstrictive stimuli exhibited similar sensitivities to basal vascular tension. Bipolar stimuli, such as the sinusoidal paradigm used here, were also affected by vascular tension, but to a greater extent than either of the other two paradigms. Furthermore, CBF-based CVR measurements obtained using a bipolar stimulus were the most sensitive to vascular tension, and may be the best choice when sensitivity is desired. However, CBF-based CVR to the vasodilatory stimulus is the least affected by vascular tension, and may be a good choice when tension information was unavailable. Finally, in view of these dependencies, we recommend that MAP measurements be incorporated into future CVR studies to provide a means to account for biases.

The above findings relate to a healthy vasculature among relatively young adults, in whom control of CO₂-elicited vascular response has not been impaired. This may not always be the case in CVR studies. For example, aging is associated with increasing blood pressure (potentially diminished ability to vasodilate further) despite decreasing CBF, as well as increasing vascular stiffness (Tarumi et al., 2014). CVR reduction in the elderly may therefore be exaggerated, as this reduction may be due both to increased vascular stiffness and to the lengthened vascular response delay in a vasodilated state. In addition, CVR impairment as reflected by a comparison of vasodilatory and vasoconstrictive responses may help to dissect the effects of blood pressure increase and vascular stiffness.

While vascular stiffness is also associated with diabetes (van Elderen et al., 2011) and cognitive impairment (Scuteri et al., 2005), current clinical applications of CVR measurements have mostly been found in the study of stroke (Zhao et al., 2009) and occlusive diseases. Notably, regional vasodilation is characteristic of carotid occlusion (Kavec et al., 2004), which could be in part emulated by our predilated vascular baseline. An observation of enhanced vasoconstrictive and diminished vasodilatory response in such patients in spite of compromised vasodilatory response may suggest that CVR is not truly impaired.

Potential caveats

When comparing our work with previous studies, we have found disparities potentially attributable to the definition of “hypercapnia” and “hypocapnia”. As we noted earlier, vascular response becomes diminished when vascular tension encroaches on the non-linear PETCO₂-response range. However, in view of the diverse implementations of hyper- and hypocapnia in the literature, it is difficult to closely compare results from different studies. For instance, most relevant studies have used the breathing of carbogen–air mixtures to achieve different capnic conditions (Bright et al., 2010; Ito et al., 2008), with no control of the actual PETCO₂ levels achieved or minimization of concurrent endtidal oxygen changes. Conversely, in our methodology, we strictly controlled PETCO₂ to vary within a ±4 mm Hg range. In addition, we constrained our CO₂ manipulations to lower levels than were used in previous studies, and minimized concurrent changes in endtidal pressure of oxygen, which has been known to modulate BOLD and CBF signals, though mildly in comparison with CO₂ (Bulte et al., 2007; Prisman et al., 2008).

We recognize that in minimizing metabolic confounds with our moderate hypercapnic and hypocapnic manipulations, we may be limited in our ability to truly simulate disease conditions, which typically entail much greater changes in MAP (see earlier Discussion). We also recognize that higher capnic challenges than ours have been commonly used in CVR studies employing prospective targeting, involving ΔPETCO₂ of up to 10 mm Hg above baseline (Kassner et al., 2010; Mandell et al., 2011; Sobczyk et al., 2014; Tancredi and Hoge, 2013). However, given our knowledge of known limitations of BOLD fMRI, we feel the need to adhere to such a limited range. We feel nonetheless that our conceptual demonstration of potential biases in CVR measurement in this limited range offers valid insight into measurement biases associated with disease conditions. Our future work will be directed towards the use of more advanced imaging techniques to probe higher MAP ranges.

Finally, a well known confound of a study such as ours is the low SNR of ASL data. Such concerns were noted earlier with regards to the ability of ASL to capture CBF changes in low-flow conditions, i.e. at a hypocapnic baseline. In this work, typical temporal SNR (mean effect size/stdev, bipolar stimulus) for the BOLD signal is ~7.0 for gray matter and 3.0 for white matter. For CBF data, the SNR is 3.5 and 1.5, respectively. Moreover, we noted that more voxels in the white matter were associated with negative, noise-driven CBF-CVR estimates, and these were excluded from our analyses. However, as Ghariq et al. (2014) demonstrated, a linear relationship better explains the variation of CBF-CVR with basal PETCO₂ than a nonlinear one, even when accounting for the lower CBF SNR. Despite our conclusion of higher robustness of CBF CVR measurements against vascular tone and type of vascular stimuli, the use of ASL for CVR measurement would perhaps be difficult if voxel-wise comparisons are desired, and may need to be based on satisfactory power analysis to ensure that the effects of interest would be captured in spite of the lower SNR.

Conclusion

Accurate and reproducible CVR mapping tools have important implications for studying cerebrovascular health. In this study, we provide empirical evidence that while BOLD-derived CVR and CBF-derived CVR are significantly correlated in most regions, they do not behave in the same manner under altered vascular tension. This finding is potentially attributable to the different physical origins of the BOLD and CBF signals. In addition, this vascular-tension dependence differs between vasodilatory and vasoconstrictive responses, necessitating caution in the choice of vascular paradigm and when comparing across studies. Finally, we explore the dependence of CVR response delay on vascular tension, and find predilation to lead to slower CO₂ reactivity. Thus, this study addresses differences between vascular stimuli, different vascular tension levels and the choice of BOLD vs. CBF when measuring CVR, and is directly relevant to CVR-based imaging research into cerebrovascular health.