BOLD-based cerebrovascular reactivity is influenced by baseline cerebral blood volume and oxygen metabolism in humans Authors: Gabriel Hoffmann, Lena Schmitzer, Jan Kufer, Matthias JP van Osch, Lena Václavů, Jan Kirschke, Christian Sorg, Stephan Kaczmarz, Christine Preibisch Published: 2026 ABSTRACT Cerebrovascular reactivity (CVR) is a key indicator of vascular health. It is typically assessed using a vasoactive stimulus such as hypercapnia while recording perfusion changes using blood oxygenation level dependent (BOLD) magnetic resonance imaging (MRI) data or measuring cerebral blood flow (CBF) directly, for example, with arterial spin labelling (ASL). Commonly used BOLD–CVR estimates, however, may be biased by differences in (baseline) cerebral blood volume (CBV). To investigate this, we acquired data from 50 participants at 3 T, obtaining BOLD–CVR, dynamic CBF–CVR, time-encoded CBF–CVR, relative CBV (rCBV) and relative oxygen extraction fraction. Additionally, dynamic simulations of BOLD signal changes were conducted. Our results showed a pronounced contrast between grey matter and white matter in BOLD–CVR, whereas CBF–CVR was similar across tissues. Both inter-subject and intra-subject analyses indicated that this grey matter/white matter contrast in BOLD–CVR primarily reflects differences in rCBV. Other potential factors, such as variations in arterial transit time or baseline CBF, did not show significant correlations. The simulations further supported these findings and, interestingly, pointed to a slow decrease of oxygen metabolism during hypercapnia. In conclusion, BOLD–CVR strongly depends on CBV, a relationship highly relevant for studies in pathologies with altered CBV, such as cerebrovascular diseases or brain tumors. Introduction Cerebrovascular reactivity (CVR) is considered to be an important marker of vascular health in various brain diseases, such as cerebrovascular diseases,1–5 multiple sclerosis 6 and glioma. 7 CVR generally describes the vasodilatory reserve of cerebral blood vessels and can be tested by a vasodilatory stimulus monitored by a physiology-sensitive readout, including blood oxygenation level dependent (BOLD), arterial spin labelling (ASL) or phase contrast (PC) magnetic resonance imaging (MRI).8,9 Hypercapnia (i.e. elevated levels of inspired CO2) or acetazolamide are frequently used to induce changes in vasodilation.10,11 BOLD–MRI and ASL provide voxel-wise values with whole-brain coverage, whereas PC–MRI is limited to measuring flow velocity in major brain feeding arteries. Due to its comparably high signal to noise ratio (SNR), high temporal resolution and decent image quality on a single subject level, BOLD–MRI-based CVR (BOLD–CVR) is generally preferred over cerebral blood flow (CBF)-based CVR (CBF–CVR) from ASL.12,13 The vasodilation-related BOLD signal change, however, does not only depend on the perfusion increase, as actually desired, but also on baseline physiology. 14 These complex dependencies have recently been investigated theoretically using a steady-state BOLD signal model. 9 Herein, the authors particularly propose that cerebral blood volume (CBV) is a major influencing factor in BOLD-based CVR measurements, a fact that has already been described in early investigations of BOLD-signal changes. 15 This may also explain study reports of higher CVR in grey matter (GM) than white matter (WM).16–18 Common BOLD-signal models also incorporate baseline oxygen extraction fraction (OEF) as a contributor to signal changes.19–21 Therefore, it appears plausible that BOLD-signal amplitudes—and consequently BOLD–CVR—may be biased by differences in baseline CBV and OEF. Lastly, it is still debated whether CO2, frequently regarded as an isometabolic stimulus, 22 especially in calibrated functional MRI (fMRI) experiments, may actually cause changes in cerebral oxygen metabolism. 14 To study these complex relations, we experimentally compared BOLD–MRI- and ASL-based CVR measurements and investigated the dependence of BOLD–CVR on baseline physiology, in particular CBV and OEF. The presented analyses were conducted across two separate datasets: the first to explore relationships between BOLD–CVR and dynamic CBF-CVR (dynCBF–CVR), as well as general signal characteristics, and the second to specifically investigate the dependence of BOLD–CVR on baseline physiology, including CBV and OEF. In addition, we performed dynamic BOLD signal simulations to investigate potential dependencies on baseline physiology and explore the dependence of temporal characteristics of hypercapnia-induced BOLD signal changes on cerebral oxygen metabolism. We hypothesized that BOLD–CVR is significantly influenced by baseline CBV as well as baseline OEF. Material and methods Study population We acquired data from 50 healthy young and elderly subjects. For comparing CVR derived from BOLD–MRI (BOLD–CVR) and dynCBF–CVR, we scanned 21 subjects in dataset 1 (age = 29.1 ± 8.6 years, 10 female). Three subjects had to be excluded from the analyses due to incomplete data acquisition; three further subjects were excluded due to apparent image artefacts. To evaluate the dependency of BOLD–CVR on CBV and OEF, we further acquired data from 29 subjects in dataset 2 (age = 71.2 ± 7.5 years, 18 female). In this elderly cohort, three subjects had to be excluded due to incidental findings. Also, we had to exclude four subjects with incomplete data acquisition (due to claustrophobia during MRI or CO2 intolerance). Lastly, we excluded two subjects due to non-optimal CO2 application (i.e. no effective seal of the face mask) and strong imaging artefacts (i.e. motion). The study design was approved by the local ethics committee (TUM Ethics Committee, TUM School of Medicine and Health, approval numbers 472/16S for dataset 1 and 293/16S for dataset 2) and was in line with the Helsinki declaration. All subjects gave written consent to participate. Figure 1 illustrates which sequences were recorded for each dataset (dataset 1 in green, dataset 2 in orange, dashed lines if data are available for both datasets). MRI protocol and derived parameters. MRI measurements performed in two subject groups are framed in different colours (dataset 1, green; dataset 2, orange) and included CVR mapping by means of single PLD pCASL, BOLD–MRI and teASL. The breathing paradigm alternated 3 min of medical air (AIR) and hypercapnia (5% CO2) for the dynamic acquisitions and 5 min each for teASL under AIR and CO2. For dynCBF–CVR and BOLD–CVR, the contrast ratio ΔCVRi between GM and WM was calculated. TeASL yielded baseline CBF, teCBF–CVR, as well as ATT. Relaxometry included T2* and T2 mapping by means of mGRE and mGraSE, respectively, yielding R2′. This was combined with rCBV from DSC MRI to calculate the rOEF. Additionally, the CBV difference (ΔrCBV) between GM and WM was determined. CVR: cerebrovascular reactivity; PLD: post labelling delay; pCASL: pseudo-continuous arterial spin labelling; BOLD: blood oxygenation level dependent; teASL: time-encoded ASL; dynCBF–CVR: dynamic CBF–CVR; GM: grey matter; WM: white matter; ATT: arterial transit time; mGRE: multi-echo gradient echo; mGraSE: multi-echo gradient and spin-echo; rCBV: relative cerebral blood volume; DSC: dynamic susceptibility contrast; rOEF: relative oxygen extraction fraction. Data acquisition Data were acquired on a 3 T clinical MRI scanner (Ingenia Elition X; Philips Healthcare, Best, The Netherlands) using a 32-channel head coil. We applied a multi-contrast MRI protocol together with CO2 application and end-tidal gas sampling. The MRI protocol and the derived parameters are visualized in Figure 1: Gas application for CVR mapping used medical air and hypercapnia (HC, 5% CO2), which were supplied by a gas-mixer (AltiTrainer, SMTec, Switzerland) via a sealed face mask. An alternating block design was used, with 3-min blocks while continuously acquiring BOLD–CVR and dynCBF–CVR data. For CVR from time encoded arterial spin labelling (teASL), we performed two separate acquisitions (baseline and hypercapnia) as continuous acquisition is not possible with this sequence. To accommodate the time-encoded pCASL measurement, we used slightly longer blocks of 5 min. End-tidal CO2 (etCO2) was monitored by a gas analyzer (ML206; AD Instruments, New Zealand). Gas application for CVR mapping used medical air and hypercapnia (HC, 5% CO2), which were supplied by a gas-mixer (AltiTrainer, SMTec, Switzerland) via a sealed face mask. An alternating block design was used, with 3-min blocks while continuously acquiring BOLD–CVR and dynCBF–CVR data. For CVR from time encoded arterial spin labelling (teASL), we performed two separate acquisitions (baseline and hypercapnia) as continuous acquisition is not possible with this sequence. To accommodate the time-encoded pCASL measurement, we used slightly longer blocks of 5 min. End-tidal CO2 (etCO2) was monitored by a gas analyzer (ML206; AD Instruments, New Zealand). Dyn-CBF–CVR was derived from single-post labelling delay (PLD) pseudo-continuous arterial spin labelling (pCASL) data following recommendations from the ISMRM Perfusion Study Group and the European Consortium for ASL in Dementia 23 using a 2D multi-slice single-shot echo planar imaging (EPI) readout: TE = 11 ms, TR = 4500 ms, flip angle = 90°, label duration = 1800 ms, PLD = 1800 ms, two background suppression pulses, SENSE factor = 2 (AP), EPI factor = 29, matrix size = 64 × 58, voxel size = 3.3 × 3.5 mm2, 20 slices (6.0 mm thickness, 0.6 mm gap); 85 dynamic images including one M0, total scan time = 12:54 min. Data was acquired for dataset 1. Dyn-CBF–CVR was derived from single-post labelling delay (PLD) pseudo-continuous arterial spin labelling (pCASL) data following recommendations from the ISMRM Perfusion Study Group and the European Consortium for ASL in Dementia 23 using a 2D multi-slice single-shot echo planar imaging (EPI) readout: TE = 11 ms, TR = 4500 ms, flip angle = 90°, label duration = 1800 ms, PLD = 1800 ms, two background suppression pulses, SENSE factor = 2 (AP), EPI factor = 29, matrix size = 64 × 58, voxel size = 3.3 × 3.5 mm2, 20 slices (6.0 mm thickness, 0.6 mm gap); 85 dynamic images including one M0, total scan time = 12:54 min. Data was acquired for dataset 1. BOLD–CVR was acquired using a 2D single-shot multi-band EPI (mbEPI) sequence: TE = 30 ms, TR = 1200 ms, flip angle = 70°, mb-SENSE factor = 2, SENSE factor = 1.5 (AP), EPI-factor = 43, matrix size = 64 × 64, voxel size = 3.0 × 3.0 mm2, 40 slices (thickness = 3.0 mm, gap = 0.2 mm ); dataset 1: 750 dynamics, total scan time = 15:04 min (two HC blocks); dataset 2: 450 dynamic images, total scan time = 09:00 min (one HC block). Data was acquired for both datasets. BOLD–CVR was acquired using a 2D single-shot multi-band EPI (mbEPI) sequence: TE = 30 ms, TR = 1200 ms, flip angle = 70°, mb-SENSE factor = 2, SENSE factor = 1.5 (AP), EPI-factor = 43, matrix size = 64 × 64, voxel size = 3.0 × 3.0 mm2, 40 slices (thickness = 3.0 mm, gap = 0.2 mm ); dataset 1: 750 dynamics, total scan time = 15:04 min (two HC blocks); dataset 2: 450 dynamic images, total scan time = 09:00 min (one HC block). Data was acquired for both datasets. Time encoded CBF–CVR (teCBF–CVR) was derived from teASL using a Hadamard-encoded pCASL (Hadamard matrix size = 8) sequence with segmented 3D gradient and spin echo (GraSE) readout: TE = 16 ms, TR = 4683 ms, bolus durations: BD1 = 1800 ms, BD2 = 600 ms, BD3 = 400 ms, BD4 = 300 ms, BD5 = 200 ms, BD6 = BD7 = 150 ms, PLD = 300 ms, two background suppression pulses, SENSE factor = 2.9 (AP), EPI factor = 23, matrix size = 64 × 64, voxel size = 3.0 × 3.0 × 3.75 mm3, 29 slices; three dynamic images, total scan time = 05:37 min. An M0 calibration scan was acquired with identical parameters without labelling or background suppression. Data was acquired for dataset 2. Time encoded CBF–CVR (teCBF–CVR) was derived from teASL using a Hadamard-encoded pCASL (Hadamard matrix size = 8) sequence with segmented 3D gradient and spin echo (GraSE) readout: TE = 16 ms, TR = 4683 ms, bolus durations: BD1 = 1800 ms, BD2 = 600 ms, BD3 = 400 ms, BD4 = 300 ms, BD5 = 200 ms, BD6 = BD7 = 150 ms, PLD = 300 ms, two background suppression pulses, SENSE factor = 2.9 (AP), EPI factor = 23, matrix size = 64 × 64, voxel size = 3.0 × 3.0 × 3.75 mm3, 29 slices; three dynamic images, total scan time = 05:37 min. An M0 calibration scan was acquired with identical parameters without labelling or background suppression. Data was acquired for dataset 2. T2*-mapping (for OEF) used a 12 echo multi-echo gradient echo (mGRE) sequence as previously described 24 : TE1 = ΔTE = 5 ms, TR = 2291 ms, flip angle = 30°, 36 slices (0.3 mm gap), matrix size = 112 × 92, voxel size = 2.0 × 2.0 × 3.0 mm3; total scan time = 06:25 min. Half- and quarter-resolution acquisitions of k-space were additionally acquired to facilitate motion correction. 25 Data was acquired for dataset 2. T2*-mapping (for OEF) used a 12 echo multi-echo gradient echo (mGRE) sequence as previously described 24 : TE1 = ΔTE = 5 ms, TR = 2291 ms, flip angle = 30°, 36 slices (0.3 mm gap), matrix size = 112 × 92, voxel size = 2.0 × 2.0 × 3.0 mm3; total scan time = 06:25 min. Half- and quarter-resolution acquisitions of k-space were additionally acquired to facilitate motion correction. 25 Data was acquired for dataset 2. T2-mapping (for OEF) used an eight echo multi-echo GraSE (mGraSE) sequence as previously detailed 26 : TE1 = ΔTE = 16 ms, TR = 251 ms, EPI factor = 7, 36 slices, matrix size = 112 × 91, voxel size = 2.0 × 2.0 × 3.3 mm3; total scan time = 02:35 min. Data was acquired for dataset 2. T2-mapping (for OEF) used an eight echo multi-echo GraSE (mGraSE) sequence as previously detailed 26 : TE1 = ΔTE = 16 ms, TR = 251 ms, EPI factor = 7, 36 slices, matrix size = 112 × 91, voxel size = 2.0 × 2.0 × 3.3 mm3; total scan time = 02:35 min. Data was acquired for dataset 2. Relative CBV (rCBV) mapping (for OEF) used dynamic susceptibility contrast (DSC) MRI with intravenous contrast agent injection (15 ml Gd-DOTA, concentration 0.5 mmol/ml, flow rate 4 ml/s, injection ~ 6 s after DSC imaging onset, 45 ml saline flush) and a gradient echo single-shot mb-EPI readout: TE = 30 ms, TR = 1200 ms, flip angle = 50°, mb-SENSE factor = 2, SENSE factor = 1.5 (AP), EPI factor 75, matrix size = 112 × 112, voxel size = 2.0 × 2.0 mm2, 36 slices (3.0 mm thickness, 0.3 mm gap); 90 dynamic images, total scan time 02:01 min. Data was acquired for dataset 2. Relative CBV (rCBV) mapping (for OEF) used dynamic susceptibility contrast (DSC) MRI with intravenous contrast agent injection (15 ml Gd-DOTA, concentration 0.5 mmol/ml, flow rate 4 ml/s, injection ~ 6 s after DSC imaging onset, 45 ml saline flush) and a gradient echo single-shot mb-EPI readout: TE = 30 ms, TR = 1200 ms, flip angle = 50°, mb-SENSE factor = 2, SENSE factor = 1.5 (AP), EPI factor 75, matrix size = 112 × 112, voxel size = 2.0 × 2.0 mm2, 36 slices (3.0 mm thickness, 0.3 mm gap); 90 dynamic images, total scan time 02:01 min. Data was acquired for dataset 2. Structural imaging comprised a T1w magnetization prepared rapid gradient echo (MPRAGE) sequence: TE = 4 ms, TR = 9 ms, flip angle = 8°, TI = 761 ms, compressed SENSE (CS-SENSE) factor = 7.5, matrix size = 240 × 250, voxel size = 1 × 1 × 1 mm3; total scan time = 02:25 min. Data was acquired for both datasets. Structural imaging comprised a T1w magnetization prepared rapid gradient echo (MPRAGE) sequence: TE = 4 ms, TR = 9 ms, flip angle = 8°, TI = 761 ms, compressed SENSE (CS-SENSE) factor = 7.5, matrix size = 240 × 250, voxel size = 1 × 1 × 1 mm3; total scan time = 02:25 min. Data was acquired for both datasets. Data processing Processing of all data was performed with custom-built MATLAB programs (version 2024b; The Math Works, Inc., Natick, MA, USA) and Statistical Parametric Mapping 12 (SPM12; Wellcome Trust Centre for Neuroimaging, London, UK). Additionally, the FMRIB Software Library (FSL) was used for evaluation of teASL data. Spatial preprocessing generally involved rigid body motion correction using SPM12 and registration of the Atlas of Intrinsic Connectivity of Homotopic Areas (AICHA) 27 to the BOLD, CBF and rCBV image space. Structural data were segmented into tissue probability maps, which were used to generate brain masks and subsequently co-registered into individual image spaces. GM and WM masks were defined using a threshold of p > 0.7. Additionally, the three bottom and top slices from each image were excluded from analyses. Together with the tissue masks, this facilitated the calculation of parenchymal signal and reduced the contribution of venous signal, especially to BOLD–CVR. The detailed processing of each parameter is described below: For CVR-processing, we first extracted end-tidal CO2 time courses from the peaks of the recorded CO2 gas sampling curve. Subsequently, the etCO2 time curves were temporally shifted to achieve optimal alignment with the mean BOLD–MRI and dynamic CBF GM time courses. Then we extracted signals during baseline (~00:00–03:20 min (datasets 1 and 2) and ~07:30–09:15 min (for dataset 1)) and hypercapnia (~04:30–06:20 min (datasets 1 and 2) and ~10:20–12:00 min (for dataset 1)). Note that timings slightly differ from Figure 1, since additional delays due to the Gas Mixer and tubing were accounted for. CVR was calculated as the percentage signal change and normalized by the change in end-tidal CO2 in mmHg (∆etCO2) between baseline and the first hypercapnia block: For CVR-processing, we first extracted end-tidal CO2 time courses from the peaks of the recorded CO2 gas sampling curve. Subsequently, the etCO2 time curves were temporally shifted to achieve optimal alignment with the mean BOLD–MRI and dynamic CBF GM time courses. Then we extracted signals during baseline (~00:00–03:20 min (datasets 1 and 2) and ~07:30–09:15 min (for dataset 1)) and hypercapnia (~04:30–06:20 min (datasets 1 and 2) and ~10:20–12:00 min (for dataset 1)). Note that timings slightly differ from Figure 1, since additional delays due to the Gas Mixer and tubing were accounted for. CVR was calculated as the percentage signal change and normalized by the change in end-tidal CO2 in mmHg (∆etCO2) between baseline and the first hypercapnia block: (1) CVR=(SignalCO2−Signalbaseline)Signalbaseline⋅1ΔetCO2 CBF parameter maps from single PLD pCASL were calculated voxel-wisely for each dynamic image according to Alsop et al., 23 including normalization by M0. DynCBF–CVR was determined analogous to equation (1) with CBF under hypercapnia and in baseline. CBF parameter maps from single PLD pCASL were calculated voxel-wisely for each dynamic image according to Alsop et al., 23 including normalization by M0. DynCBF–CVR was determined analogous to equation (1) with CBF under hypercapnia and in baseline. teASL data were first decoded and subsequently evaluated using Bayesian inference for arterial spin labelling (BASIL) from FSL.28–30 CBF maps during baseline were registered to CBF maps acquired under HC, and subsequently teCBF–CVR was calculated as percentage CBF change/mmHg according to equation (1). Additionally, teASL facilitated the determination of arterial transit time (ATT) maps, both at baseline (medical air) and under hypercapnia, which we used to calculate percentage ATT changes: teASL data were first decoded and subsequently evaluated using Bayesian inference for arterial spin labelling (BASIL) from FSL.28–30 CBF maps during baseline were registered to CBF maps acquired under HC, and subsequently teCBF–CVR was calculated as percentage CBF change/mmHg according to equation (1). Additionally, teASL facilitated the determination of arterial transit time (ATT) maps, both at baseline (medical air) and under hypercapnia, which we used to calculate percentage ATT changes: (2) ΔATT=(ATTbaseline−ATTCO2)/ATTbaseline⋅100 For rCBV, we processed DSC MRI data as described previously, 31 including leakage correction using the method from Boxerman et al. 32 and normalization to NAWM CBV.24,26 For rCBV, we processed DSC MRI data as described previously, 31 including leakage correction using the method from Boxerman et al. 32 and normalization to NAWM CBV.24,26 A GM/WM contrast ratio was calculated for both BOLD- and dynCBF–CVR: A GM/WM contrast ratio was calculated for both BOLD- and dynCBF–CVR: (3) ΔCVRi=(CVRi,GM−CVRi,WM)/CVRi,WMi∈{CBF,BOLD} and for rCBV: (4) ΔrCBV=(rCBVGM−rCBVWM)/rCBVWM The calculation of relative OEF (rOEF) used the multi-parametric quantitative BOLD (mqBOLD) model, which is based on the BOLD signal model of Yablonskiy and Haacke 33 and uses three independent measurements of T2, T2* and rCBV to calculate rOEF.24,26,34 The relaxometry data (mGRE and mGraSE) were fitted mono-exponentially to obtain T2* and T2 parameter maps. Phase data of the mGRE acquisition were used to correct T2* for macroscopic B0 field inhomogeneities. 35 Additionally, the T2* parameter maps were corrected for subject motion. 36 T2* and T2 data were spatially co-registered using SPM12 before R2′ and rOEF calculation: The calculation of relative OEF (rOEF) used the multi-parametric quantitative BOLD (mqBOLD) model, which is based on the BOLD signal model of Yablonskiy and Haacke 33 and uses three independent measurements of T2, T2* and rCBV to calculate rOEF.24,26,34 The relaxometry data (mGRE and mGraSE) were fitted mono-exponentially to obtain T2* and T2 parameter maps. Phase data of the mGRE acquisition were used to correct T2* for macroscopic B0 field inhomogeneities. 35 Additionally, the T2* parameter maps were corrected for subject motion. 36 T2* and T2 data were spatially co-registered using SPM12 before R2′ and rOEF calculation: (5) rOEF=R2′43⋅γ⋅π⋅Δχ⋅B0⋅rCBV with B0=3T and Δχ=Hct⋅Δχ0=0.35⋅0.264⋅10−6=0.925⋅10−7. 24 Deep GM and WM were excluded from evaluations, as well as physiologically unrealistic values >1. Statistical evaluations Statistical evaluations were performed in MATLAB. Prior to statistical analyses, we calculated subject-wise VOI-average values for all parameters and tested for normal distribution across subjects using the Anderson–Darling Test, assuming p > 0.05 indicating a normal distribution of the respective parameter values. Statistical testing employed paired Student’s t-tests for normally distributed parameter values and Wilcoxon signed-rank tests for non-normally distributed parameters. Correlation analyses used Pearson’s correlation and Spearman’s correlation, respectively, and corresponding linear regression lines are shown in all plots. Statistical significance is assumed at p < 0.05 for all analyses. Paired scatterplots show subject-wise VOI-average values as dots, which are connected with solid lines between parameter values, and means are indicated as red dashed lines. Whisker plot boxes indicate 25th and 75th percentiles, median values are shown as horizontal lines, and outliers are visualized by crosses. Data were analyzed both on an inter-subject level using subject-wise VOI-average values (Figures 3–6) as well as on an intra-subject level by evaluating values across AICHA VOIs within individual subjects (Figure S4). For evaluating GM–WM contrast, we compared ΔCVRBOLD and ΔCVRCBF (according to equation (3)) as well as ΔCRVBOLD and ΔrCBV (equation (4)). BOLD- and dynCBF–CVR signal characteristics were assessed by means of temporal signal-to-noise ratio (tSNR), which was calculated voxel-wisely across the first baseline periods of dynamic acquisitions as the quotient of the temporal signal mean and standard deviation. Correlation of the full time courses with etCO2 (for R and p) and CVR was calculated as described above. Correlation analyses were performed to investigate the agreement of both CVR methods (BOLD and CBF) and assess the potential impact of ΔATT and baseline CBF. Lastly, we correlated ΔSBOLD, as a proxy for BOLD–CVR, with rCBV and rOEF. Here we used ΔSBOLD to facilitate comparison with simulations. In addition to the above-described inter-subject investigations we also correlated rCBV and baseline CBF within single subjects (intra-subject analysis) and compared R and p values of those correlations. Simulations We simulated the dependence of the BOLD signal on hypercapnia-induced CBF and CMRO2 changes, as well as baseline CBV and OEF, using a previously published BOLD signal model 19 based on the balloon model from Buxton and co-workers.21,37 Using measured etCO2 and dynCBF time courses as driving inputs, we first varied baseline CBV and OEF to explore the dependency of ΔSBOLD on baseline physiology. In a second step, we varied the characteristic time constant (τm) and amplitude (m1) of ΔCMRO2 within physiologically plausible ranges to achieve an optimal fit of the measured ΔSBOLD. Further details and constant definitions can be found in the Supplementary Material. To facilitate comparability between simulations and experimental data in Figure 6, we normalized simulated ΔSBOLD to the mean difference in etCO2 of dataset 2 (∆etCO2 = 16 mmHg) to yield BOLD–CVR in %/mmHg. Results To compare BOLD and dynCBF signals as a function of etCO2, we inspected group-average time curves for both GM and WM (Figure 2). Under hypercapnia, both BOLD and dynCBF signal changes follow the rise in etCO2 levels (Figure 2(a)–(c)). Regarding BOLD–CVR (Figure 2(a) and (c)), mean signal time courses and parameter maps indicate a distinct contrast between GM and WM, which can be similarly observed in rCBV maps (Figure 2(d)). Compared to that, the dynCBF percent CBF changes are similar between GM and WM, and the group average parameter maps appear more homogenous (Figure 2(b)). Corresponding teCBF–CVR and ΔATT maps are presented in the Figure S3. Group results: Time courses and parameter maps. For BOLD–CVR (a, c) and dynCBF–CVR (b) group average signal time courses (∆SBOLD, ∆CBF) in GM (red), WM (green) and etCO2 (blue) are displayed as solid lines, together with standard deviations across the group indicated by the shaded areas. Time courses show distinct BOLD signal and CBF increases under hypercapnia. However, BOLD–CVR is clearly lower in WM than GM (a, c) for both datasets, while the dynCBF signal change is similar in GM and WM (b). This is in accordance with group-average parameter maps, where dynCBF–CVR maps appear homogenous (b) while BOLD–CVR exhibits a clear contrast between GM and WM (a, c). In dataset 2, we also compared maps of BOLD–CVR (c) to rCBV (d) and rOEF (e), and found a similar structural contrast, especially for rCBV. Note bright edges in the top and frontal regions of BOLD–CVR parameter maps, which are most likely due to venous signal contributions and excluded in quantitative VOI analyses. Also, the mqBOLD method for rOEF estimation is only applicable in GM, and, in this cohort, is also affected by an apparent fat-shift artefact at the posterior border of the brain. GM: grey matter; WM: white matter; etCO2: end-tidal CO2; rCBV: relative cerebral blood volume; rOEF: relative oxygen extraction fraction. To investigate this visually perceived GM–WM contrast more closely, we performed a quantitative group-level evaluation in Figure 3. Indeed, the GM–WM contrast ΔCVR (according to equation (3)) of BOLD–CVR was significantly higher compared to dynCBF–CVR (ΔCVRBOLD = 0.84 ± 0.29 vs ΔCVRdynCBF = 0.08 ± 0.11, p < 0.001; Figure 3(a), dataset 1) and teCBF–CVR (ΔCVRBOLD = 0.60 ± 0.24 vs ΔCVRteCBF = 0.29 ± 0.09, p < 0.001; Figure 3(b), dataset 2). Also, in line with the visual impression, ΔCVRBOLD closely resembled ΔrCBV (ΔCVRBOLD = 0.60 ± 0.24 vs ΔCVRrCBV = 0.62 ± 0.10, p = 0.63; Figure 3(c)). Normalized CVR contrast between GM and WM (ΔCVR). Compared to BOLD–CVR, GM–WM contrast was found to be significantly lower for dynCBF–CVR from dataset 1 (a) as well as teCBF–CVR from dataset 2 (b). In contrast, BOLD–CVR and rCBV had a similar GM–WM contrast (c). Whisker plot boxes indicate 25th and 75th percentiles, median values are shown with a horizontal line, and outliers are visualized by crosses. Dots indicate individual subject averages and are connected by dashed grey lines. Paired t-tests were used for (a, c) and Wilcoxon signed rank for (b). n.s.: not significant. ***p < 0.001. In order to compare signal quality between BOLD–CVR and dynCBF–CVR, we calculated several metrics, including tSNR, correlation with etCO2 and CVR, both in GM (Figure 4(a)–(c)) and WM (Figure 4(d)–(f)). BOLD–CVR demonstrates significantly higher temporal SNR (tSNRBOLD–CVR = 154.9 ± 19.5 vs tSNRdynCBF–CVR = 2.3 ± 0.3 in GM; Figure 4(a)), correlation coefficients (R), and lower p values for correlation of MRI measures with etCO2 (Figure 4(b,c) and (e,f). Statistical analyses BOLD–CVR versus dynCBF–CVR (dataset 1). BOLD–CVR yielded a significantly higher tSNR ((a) GM and (d) WM) and stronger correlation with the end-tidal CO2-ref curve, both in terms of correlation coefficient R ((b) GM and (e) WM) and p value ((c) GM and (f) WM). Dots indicate individual subject averages of GM and WM, respectively; solid lines connect values from the same subjects, and group means are indicated by red dashed lines. Asterisks indicate statistical significance (t-test). Note that for p (CO2-ref), significance was tested using Wilcoxon signed rank tests due to non-normal distribution. CO2-ref: CO2 reference; tSNR: temporal SNR. ***p < 0.001. While investigating the dependency of BOLD–CVR on different perfusion measures (Figure 5), we found a significant correlation between BOLD–CVR and dynCBF–CVR (Figure 5(a); r = 0.65) as well as teCBF–CVR (Figure 5(b); r = 0.60) but no correlation with ΔATT (r = 0.03; Figure 5(c)) or baseline CBF (r = 0.08; Figure 5(d)). Correlation of BOLD–CVR with perfusion parameters. BOLD–CVR is significantly correlated with dynCBF–CVR (r = 0.65 (a)) in dataset 1. In dataset 2, BOLD–CVR is similarly correlated with teCBF–CVR (r = 0.60 (b)). In contrast, no significant correlation was found between BOLD–CVR and ΔATT (c) or baseline CBF (d). Dots indicate individual subject averages of parameter values in GM and WM, and regression lines are plotted to visualize potential linear trends. Pearson correlation was used for normally distributed data (a, c), while Spearman correlation was used to assess correlations of BOLD–CVR with non-normally distributed teCBF–CVR (b) and baseline CBF (d) from teASL (p < 0.05). Comparing the measured and simulated relationship between BOLD–CVR and baseline physiology (rCBV and rOEF; Figure 6), we found good concordance between experimental results and simulations: data indicate a significant correlation of BOLD–CVR with rCBV (Figure 6(a); r = 0.67), which agrees with simulations of BOLD–CVR for varying baseline CBV0, which was performed at various baseline OEF0 values (Figure 6(b)). Note that correlation of BOLD–CVR and rCBV is also significant in GM only (r = 0.65, p < 0.01, not shown). Similarly, we observed a weak trend towards a correlation between BOLD–CVR and baseline rOEF (Figure 6(c); r = 0.12, p = 0.67), with simulations showing a clear dependence of BOLD–CVR on OEF0 (Figure 6(d)). Note that rOEF results are only presented in GM, since mqBOLD does not allow for reliable oxygenation mapping in WM. Dependency of BOLD–CVR on baseline physiology. Experimental results (a, c) are compared to results of dynamic BOLD-signal simulations yielding BOLD–CVR for physiologically plausible ranges of baseline CBV0 and OEF0 (b, d). Simulated plateau values of BOLD–CVR are plotted in dependence on CBV0 for a range of colour-coded OEF0 (b) or in dependence on OEF0 for a range of colour-coded CBV0 (d), respectively. In experimental data (a, c), BOLD–CVR was found to be significantly correlated with rCBV (a), essentially resembling the simulation results (b). Also, BOLD–CVR was weakly correlated with rOEF (r = 0.12, p = 0.62 (c)), and simulations showed strong dependency on OEF0 (d). Note that for simulations, venous CBV0 is assumed (see Supplementary Material), while rCBV from DSC represents total CBV. Beyond its dependency on baseline CBV0 and OEF0 we investigated the dependency of the BOLD-signal shape on the amplitude m1 and time constant τm of CMRO2 changes using our BOLD signal simulations with measured etCO2 and CBF-change as inputs (Figure 7(a)–(c)). Interestingly, the best fit of the measured ΔSBOLD (Figure 7(d)–(f), red curves) and the simulated BOLD signal (green curves) was found when a slow (τm ~15 s) decrease in CMRO2 of ~20% (m1 = 0.8) was allowed in response to 5% CO2 (Figure 7(c) and (f) and see Figure S2 for more details regarding the determination of optimal fit parameters for temporal characteristics of the CMRO2 change). In particular, the slow decrease in CMRO2 contributed to achieving a steady increase in the BOLD signal throughout the hypercapnia plateau phase, which was not observed in the etCO2 or CBF time curves. Investigating the impact of hypercapnia-induced CMRO2 changes on simulated BOLD signal responses. Top panels ((a)–(c), input): simulated CBF (fin(t), cyan), CBV (v(t), green) and CMRO2 (m(t), blue) input curves that were used for simulating the BOLD signal responses presented in the bottom panels. Bottom panels ((d)–(f), simulated BOLD-signal change ∆SBOLD): BOLD signal responses to 5% hypercapnia were simulated for a baseline CBV value of CBV0 = 1.7% and exemplary CMRO2 responses (see blue curves and parameter values in respective top panels), using the measurement-informed, adapted CBF input (f1 = 1.4, i.e. ΔCBF = 40%, τ f = 1 s; cyan curves in (a)–(c)) and assuming a slow CBV response (α = 0.2, τ V = 20 s; green curves in A-C; see Archila-Meléndez et al. 19 ). Left: isometabolic conditions (a) with constant CMRO2 (m1 = 1), elicit the lowest simulated BOLD-signal change ΔSBOLD (d), which also appears flat compared to the measured ΔSBOLD in GM (red). Note that for m1 = 1, τ m does not have any effect on simulated BOLD signals. Middle: decreasing CMRO2 during hypercapnia (m1 = 0.8, τ m = 1 s (b)) causes higher ΔSBOLD (e). Right: Longer CMRO2 time constants (τ m = 15 s (c)) achieve simulated BOLD signals that better fit the measured ΔSBOLD in GM (red (f)). See Table S1 for the definition of all constants. Lastly, we also investigated the dependency of BOLD–CVR on rCBV and CBF within individual subjects (see Figure S4). Both data of an exemplary subject and comparing R and p values across subjects confirmed significantly stronger correlation of BOLD–CVR with rCBV than CBF. Discussion In this study, we investigated the influence of several physiological factors, including perfusion and metabolism, on BOLD-based CVR measurements. As hypothesized, both our experimental and simulation results revealed a significant dependence of BOLD–CVR on (baseline) cerebral blood volume and OEF, while no correlation was found with baseline CBF. Regarding practical applicability, we first evaluated how BOLD- and ASL-based CVR compare in terms of data quality. As expected, BOLD–CVR yielded significantly higher tSNR and correlated better with etCO2—a fact that seems plausible with regard to previous investigations.12,38 Likewise, absolute values of quantitative BOLD–CVR in GM of 0.22% ± 0.06%/mmHg and dynCBF–CVR = 3.5% ± 1.0%/mmHg in young HC are in excellent agreement with literature (BOLD–CVR ~0.17%–0.31%/mmHg,18,39–41 CBF–CVR ~2.8%–4.5%/mmHg12,40–42). Examining subject- and VOI-average signal time courses and parameter maps (Figure 2), a clear GM–WM contrast was observed for BOLD–CVR, which resembles contrast in rCBV parameter maps. In comparison, CBF–CVR is similar in GM and WM, both facts that we confirmed with a statistical analysis of ΔCVR and ΔrCBV (Figure 3). This finding also agrees well with previous studies, where similar CBF–CVR was found in GM and WM, both from single PLD 17 and teASL, 43 while BOLD–CVR differences between GM and WM have also been reported.16–18 However, there was a good correlation of BOLD–CVR and both dynCBF–CVR and teCBF–CVR, findings that likewise agree with previous work.17,41,44 In addition to GM–WM contrast, BOLD–CVR maps showed bright edges along the borders of the brain, which were largely excluded from analyses because they most likely arise from superficial draining veins and large vessels such as the superior sagittal sinus. This concords with venous architecture 45 and previous BOLD–CVR studies41,44 and has been described as especially prominent for GE-EPI acquisitions. 46 Regarding the dependence of BOLD–CVR on baseline physiology, we examined several factors. First, we investigated whether baseline perfusion, i.e., CBF, and perfusion delays, i.e., ATT, had an impact on BOLD–CVR. Interestingly, BOLD–CVR appears to be independent of baseline CBF, despite CBF being significantly larger in GM compared to WM, due to differences in vessel density.47–49 This finding is supported by both our inter- (Figure 5) and intrasubject analyses (Figure S4), and has implications for clinical imaging, as BOLD–CVR appears to add additional diagnostic value to a baseline CBF scan. In the context of this work, this result indicates that differences in baseline CBF do not explain the apparent GM–WM difference in BOLD–CVR. Furthermore, we did not observe any dependence on ΔATT, thus indicating stability across perfusion delays in BOLD–CVR. Considering blood volume, we confirmed the hypothesized dependence of BOLD–CVR on baseline rCBV (Figure 6). This also very well explains the strong GM–WM contrast of BOLD–CVR and agrees with previous findings from PET 50 and MRI studies,51–53 demonstrating substantially larger CBV in GM than WM. Our findings also align with the theoretical considerations by Schulman et al. 9 who hypothesized a similar GM–WM contrast for BOLD–CVR without including experimental data. It also agrees with the analytical model of the BOLD signal by Yablonskiy and Haacke that describes the BOLD signal as dependent on (venous) CBV. 33 To some extent, it would also appear plausible that GM–WM differences of BOLD–CVR may also depend on differences of hypercapnia-induced CBV changes between GM and WM, considering that PET studies reported larger CBV reactivity in GM than WM. 54 However, we could not investigate this because we lack CBV measurements during hypercapnia. With respect to oxygen metabolism, a weak relationship between BOLD–CVR and baseline rOEF appears plausible given theoretical considerations. 9 Here, we speculate that this relation might be stronger than presented in our data, since, compared to baseline rCBV and CBF, rOEF is particularly noisy because it is derived from a total of three independent measurements, namely T2, T2* and rCBV. 26 From those, T2* is particularly susceptible to motion and susceptibility artefacts,35,36 and, in our cohort, T2 was additionally affected by an apparent fat shift artefact. Reassuringly, both dependencies on baseline rCBV and rOEF are supported by the dynamic BOLD signal simulations when using physiologically realistic model parameters as informed by our measurements of (baseline) physiology and employed in previous studies. 19 Our simulations predict a relationship between BOLD–CVR and both baseline CBV and OEF (see Figure 6). Interestingly, our simulations also revealed that an optimal fit of the measured and simulated BOLD data appears to be aided by allowing for a slow reduction in CMRO2 by ~20% (see Figure 7 and Figure S2). In particular, this explains a steady increase of the BOLD signal across the etCO2 plateau, that is, not found for dynamic CBF measurements. This simulation finding contradicts the frequently used isometabolic assumption in calibrated fMRI experiments.14,22,55,56 And, it at least qualitatively agrees with a number of previous studies reporting a CMRO2 decrease of ~13% using MRI by means of T2-relaxation under spin-tagging (TRUST) and PC–CBF 57 as well as TRUST and ASL–CBF, 58 but also applying near infrared spectroscopy and Doppler ultrasound. 59 Limitations Our study is subject to a few limitations. Due to technical reasons (i.e. the use of a static protocol file for controlling the gas mixer), we used a fixed scan order, while randomizing scans would have been beneficial to balance potential saturation effects of the BOLD signal and CBF responses to CO2. However, we did not find any significant differences in etCO2 values between scans of dataset 1 (Table S2). In addition, BOLD signal as well as CBF changes were comparable in the first and second CO2 blocks for both BOLD–CVR and dynCBF–CVR time courses (see Figure 2(a) and (c)). Thus, we are confident that a slightly extended period of 5 min CO2, as used for teCBF–CVR to accommodate the time-encoded acquisition, does not compromise teCBF–CVR results. With regard to OEF, the applied mqBOLD method is restricted to cortical GM, since T2, T2* and DSC–MRI-based CBV measurements exhibit directional dependencies in WM, 60 and iron deposition in deep GM also compromises R2′ estimates.26,61 Regarding our two different data sets, the subject groups exhibit a considerable age difference. Nevertheless, despite the fact that CBF 62 and CVR decrease with age,40,63 we assume that the underlying relationships are generally valid. Notably, since all conducted analyses yielded similar results in both datasets (i.e. Figures 2(a) and (b), 3(a) and (b) and 5(a) and (c)), we assume that our assumption is reasonably valid. Conclusion In conclusion, our results indicate that BOLD–CVR is strongly dependent on baseline CBV and may also be influenced by baseline OEF in pathologies that considerably affect brain oxygen metabolism. In healthy subjects, the dependence on baseline CBV mainly manifests as a clear GM–WM contrast, which is not present in CBF–CVR. These findings also agree with our dynamic BOLD-signal simulations in response to a hypercapnic challenge. Interestingly, our simulation results are consistent with a clear decrease of CMRO2 under hypercapnia. Thus, our results do not just provide a deeper understanding of BOLD-based CVR and calibrated fMRI experiments, but may also have clinical implications, especially in patients with altered CBV and OEF, such as in cerebrovascular diseases or brain tumors.64–68 Supplemental Material sj-pdf-1-jcb-10.1177_0271678X261459813 – Supplemental material for BOLD-based cerebrovascular reactivity is influenced by baseline cerebral blood volume and oxygen metabolism in humans Supplemental material, sj-pdf-1-jcb-10.1177_0271678X261459813 for BOLD-based cerebrovascular reactivity is influenced by baseline cerebral blood volume and oxygen metabolism in humans by Gabriel Hoffmann, Lena Schmitzer, Jan Kufer, Matthias JP van Osch, Lena Václavů, Jan Kirschke, Christian Sorg, Stephan Kaczmarz and Christine Preibisch in Journal of Cerebral Blood Flow & Metabolism