Hemodynamic imaging parameters in brain metastases patients – Agreement between multi-delay ASL and hypercapnic BOLD

Arterial spin labeling (ASL) MRI is a routine clinical imaging technique that provides quantitative cerebral blood flow (CBF) information. A related technique is blood oxygenation level-dependent (BOLD) MRI during hypercapnia, which can assess cerebrovascular reactivity (CVR). ASL is weighted towards arteries, whereas BOLD is weighted towards veins. Their associated parameters in heterogeneous tissue types or under different hemodynamic conditions remains unclear. Baseline multi-delay ASL MRI and BOLD MRI during hypercapnia were performed in fourteen patients with brain metastases. In the ROI analysis, the CBF and CVR values were positively correlated in regions showing sufficient reserve capacity (i.e. non-steal regions, rrm = 0.792). Additionally, longer hemodynamic lag times were related to lower baseline CBF (rrm = −0.822) and longer arterial arrival time (AAT; rrm = 0.712). In contrast, in regions exhibiting vascular steal an inverse relationship was found with higher baseline CBF related to more negative CVR (rrm = −0.273). These associations were confirmed in voxelwise analyses. The relationship between CBF, AAT and CVR measures seems to be dependent on the vascular status of the underlying tissue. Healthy tissue relationships do not hold in tissues experiencing impaired or exhausted autoregulation. CVR metrics can possibly identify at-risk areas before perfusion deficiencies become visible on ASL MRI, specifically within vascular steal regions.


Introduction
Advances in imaging techniques have made MRI a powerful tool for investigating not only brain function but also brain physiology and auto-regulatory status.One such method that now sees routine clinical application is Arterial Spin Labeling (ASL).2][3] Multi post label delay (multi-PLD) methods can even be used to infer the presence of collateral cerebral blood flow (CBF) pathways through their ability to estimate arterial arrival time (AAT) or characterize arterial transit artifacts (ATA).Together these ASL metrics can provide valuable physiological information in healthy subjects and can also identify disease-related auto-regulatory changes in patient populations.
A major cerebral auto-regulatory mechanism responsible for maintaining adequate CBF is cerebrovascular reactivity (CVR).5][6] The CVR response can be assessed and spatially mapped using Blood oxygenation level-dependent (BOLD) MRI in combination with controlled hypercapnic stimuli.This technique is distinct from resting-state and/or task-based BOLD MRI, where either spontaneous or evoked neuronal signals modulate the BOLD signal change, providing information related to brain function.CVR measurements can be interpreted as regional indicators of healthy or disease-impaired vasculature. 7Impairments often manifest as negative signal responses, where a vasodilatory stimulus can cause paradoxical decrease in CBF (known as vascular steal) to an area with exhausted dilatory reserve, due to a reduction in vascular resistance in neighboring, non-exhausted regions. 83][14] Growth of intracranial masses, like primary brain tumors or brain metastases, can cause local disruptions of the hemodynamic environment.Metastatic cells have to proceed through a range of developmental steps in order to form macrometastases; (1) cells arrest in the small microvessels, (2) cells extravasate to enter the brain tissue, (3) cells perpetuate into a strict perivascular position, and (4) vascularization is secured through either co-optive or angiogenic growth depending on the tumor type. 157][18] The vasogenic edema often surrounding these brain metastases, caused by the local blood-brain barrier disruption 19 , can lead to additional physiological changes.All these brain metastases related physiological changes are likely to be reflected in hemodynamic MRI measurements.ASL and CVR measure associated responses, both describing the status of the vascular network.Thereby they can provide complementary information, however, it remains unclear how ASL and CVR parameters relate in different tissue types or under different hemodynamic circumstances.
In the current exploratory study we test how ASL parameters measured during a physiological steady state relate to functional vascular parameters as measured using CVR in a population of patients with brain metastases.The expected tissue-heterogeneity caused by the brain metastases growth provides a testing ground to answer this question.CBF and AAT were compared to CVR magnitude and hemodynamic lag using two different approaches.Firstly, the intermodality agreement between ASL and CVR was tested by comparing group values for different tissue types (grey matter (GM), white matter (WM), edema and brain metastases) for both ASL and CVR metrics.Secondly, we assessed the spatial correlation between ASL and CVR metrics throughout the brain using both a ROI and voxelwise approach.Hereby a distinction was made between brain areas showing adequate vascular responses and those exhibiting vascular steal phenomena.Understanding how baseline vascular physiology relates to dynamic vascular processes in cases with brain metastases will aid in interpreting these measures in future research.

Study set-up and population
For the current retrospective observational study, MRI datasets were included from the ongoing Assessing and Predicting Radiation Influence on Cognitive Outcome using the cerebrovascular stress Test (APRICOT) study.Participation consists of an elaborate neurocognitive exam and MRI scans, including a BOLD MRI scan during breathing challenges, before radiotherapy and approximately three months after radiotherapy.For the purpose of this study only the MRI data acquired before radiotherapy during the period between October 2020 and February 2022 was used.The study population consists of adult patients (!18 years) with either radiographic and/or histologic proof of metastatic brain disease that were referred to the radiotherapy department of the UMC Utrecht for radiation therapy of the brain metastases.Specific in-and exclusion criteria for participation in the APRICOT study are listed in Supplementary Methods.The study was performed in accordance with the Declaration of Helsinki 20 and the UMCU institutional ethical review (Medisch-Ethische Toetsingscommissie (METC) NedMec) approved the study (METC# 18-747).Written informed consent was obtained from all participants prior to participation.

Data acquisition
Imaging protocol.The participants were scanned on a 3 Tesla MRI scanner (Achieva, Philips Medical Systems, Best, The Netherlands) using a 32 channel receive coil.Whole-brain multi-slice single shot FE-EPI BOLD images (TR ¼ 1050 ms, TE ¼ 30 ms, flip angle 65 , resolution 2.292 Â 2.292 Â 2.5 mm 3 , acquisition matrix 96 Â 96 Â 51, 1000 dynamics, multi-band factor ¼ 3) were acquired throughout a computer controlled hypercapnic breathing protocol (described below).Clinical CT and MRI acquisition.CT and MRI scans were acquired as part of clinical care as usual 1 to 5 days before receiving radiotherapy.CT scans were acquired on a Brilliance Big bore 22 scanner (Philips Medical Systems, Best, The Netherlands) with a tube potential of 120 kVp, matrix size of 512 Â 512 and inplane slice thickness of 1 mm.The participants were scanned on a 1.5 Tesla MRI scanner (Inginia, Philips Medical Systems, Best, The Netherlands) using a 15 channel receive coil.A 3 D SPGR sequence after injection of 0.1 ml gadovist/kg was performed (TR ¼ 7.6 ms, TE ¼ 3.4 ms, flip angle 8 , isotropic resolution: 1 mm, acquisition matrix 232 Â 232 Â 170).The clinician uses this clinically acquired MRI registered to the CT, to plan the radiotherapy and delineate the so-called gross tumor volume (i.e.brain metastases).The CT and corresponding gross tumor volume were extracted for each patient.For the current analysis, a distinction was made between newly treated brain metastases, resection cavities of brain metastases and previously irradiated brain metastases.
Breathing protocol.Hypercapnic stimuli were delivered using a computer-controlled gas blender and sequential delivery system (RespirAct TM , Thornhill Research Institute, Toronto, Canada).The breathing mask was sealed to the patients' face using transparent dressings (Tegaderm, 3 M, St. Paul, MN, US) to acquire an air tight seal.Before starting the breathing challenges inside the MRI scanner, patients performed a test round with a CO 2 challenge which is similar to the CO 2 block given during the protocol.Only after successful completion of the test round, the breathing challenges inside the scanner were performed (see Figure 1).The breathing challenges started with a 5-minute baseline period, followed by a block-shaped increase of end-tidal CO 2 (PetCO 2 ) 10 mmHg relative to a patients' baseline for 90 seconds.After this socalled CO 2 -block, PetCO 2 values returned to baseline values for 120 seconds, followed by a PetCO 2 ramp increase of 12 mmHg relative to patients' baseline for 180 seconds after which patients returned to baseline for 90 seconds.

Pre-processing
Pre-processing steps were performed using the Oxford Centre for Functional MRI of the BRAIN (FMRIB) Software Library (FSL -version 6.0; see Supplementary Figure 1). 21First, both the T1 and T2FLAIR images were brain extracted using BET. 22Tissue segmentation into grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF) was performed on the T1 image using FSL Automated Segmentation Tool (FAST). 23dditionally, an edema mask was created based on the T1 and T2FLAIR images using the lesion growth algorithm as implemented in the Lesion Segmentation Tool (LST, https://www.statistical-modelling.de/lst.html)for SPM. 24Based on previous experience, the initial threshold was set at 0.14.The resulting edema mask was manually adapted to eliminate any false positives or false negatives from the LST edema mask.The CT image was registered to the T1 image using FMRIB's Linear Image Registration Tool (FLIRT). 22,25OLD data was motion corrected (MCFLIRT) 22 , corrected for geometric distortion using TOPUP 26,27 and linear spatial co-registered to the T1 image using the 'epi_reg' function. 22,25The T1w segmentation masks, edema masks and gross tumor volume masks were translated to the functional (i.e.BOLD) space using the inverse of the matrices output by epi_reg.For the ASL data, the reconstructed T1 (M0) was spatially registered to the pre-TOPUP mean BOLD image.The resulting transformation matrix as well as the distortion-correction fields were then applied to the ASL images in order to move them to functional patient space.
Data analysis MRI analysis.The pipeline to process and compare the BOLD and ASL data is visually depicted in Figure 2. CVR and hemodynamic lag maps were derived using the open-source seeVR toolbox (available at https:// www.seevr.nl/). 28In brief, in order to remove signal contribution from large veins that might overshadow tissue responses, a modified whole-brain mask was generated using the 'remLV.m'function of the seeVR toolbox. 28or this, a temporal noise-to-signal (tNSR) map was calculated by taking the inverse of the temporal signal to noise (tSNR) map.Next, voxels showing values higher than the 98th percentile tNSR value were removed from the original whole brain-mask.This modified mask was then used in subsequent analyses. 28ext, a manual bulk alignment was performed between the PetCO 2 and average GM time-series to minimize alignment errors that can occur when using automated correlation methods.Thereby, any bulk delays between end-tidal gas measurements at the lungs and BOLD signal responses in the brain were accounted for.Residual motion signals with a correlation higher than 0.3 with the GM time-series, along with a linear drift term were regressed out using a general linear model.BOLD data was then temporally de-noised using a wavelet-based approach 29 and was spatially smoothed using a 3 D Gaussian kernel (FWHM: 4 Â 4 Â 7 mm 3 ).This kernel was chosen in order to best match the effective spatial resolution of the ASL data.BOLD data and corresponding PetCO 2 traces were interpolated by a factor 4 (effective TR: 262.5 ms) to identify sub-TR signal displacements in subsequent hemodynamic lag analysis.A linear regression was performed between the bulk-aligned PetCO 2 trace with each processed BOLD voxel time-series.The slope of this linear regression was taken as the CVR (DBOLD/mmHg PetCO 2 ).Hemodynamic lag maps were generated using the Rapidtide 10,30 approach as implemented in the seeVR toolbox 28 and described previously. 31he multi-delay ASL was processed using the opensource ClinicalASL toolbox (available at https://github.com/JSIERO/ClinicalASL) and FSL BASIL for quantitative CBF maps. 32A T1-weighted image was reconstructed based on the M0 images using the 'ASLT1fromM0Compute.m' function.In short, as we used a Look-Locker read-out, the signal evolution over the multiple PLDs will show a T1-weighted signal response that was used to generate a surrogate T1 weighted image.This image had sufficient T1 contrast to be used for image registration (i.e.improved contrast compared to any single M0 or label/control image).Outlier removal was performed based on the standard deviation and tissue variance.Quantitative CBF and AAT maps were generated using the BASIL tool (Figure 2(2)). 32atistical analysis.The CSF mask and non-brain metastases mask (containing areas of previously resected or irradiated brain metastases) were excluded from the hemodynamic parameter maps (Figure 2(3)).Thereby, all analyzed tissue consists of either brain matter (GM and WM), edema or non-treated brain metastases.Next, the CVR values were used to divide brain into regions of vascular steal versus no vascular steal.In this study, steal was conceptualized as voxels containing negative CVR values (i.e.regions in which BOLD signal intensity decreased with increase in PetCO 2 ; Figure 2(4)).
The data provided by steps 1-4 were used in the subsequent comparative analyses.First the previously generated masks were used to subdivide the brain into GM, WM, edema and brain metastases regions (Figure 2(A)).Next, for steal and non-steal regions separately, the mean value per brain regions was calculated for each hemodynamic parameter map.Kruskal-Walis H-tests were performed to compare the mean hemodynamic values between tissue types.To reduce the false discovery rate due to multiple testing, alpha's were corrected according to the Benjamini-Hochberg method. 33dditionally, Spearman's correlation analyses were performed between the mean values of the different parameter maps within each tissue type.
For a second comparative analysis, the AAT data was sorted based on ascending values.The sorted AAT's were then subdivided into 20 bins, each containing 5 percent of the data.Using the boundaries of these '5%-bins', the AAT map was divided into ROIs.Thereby, each ROI contained 5 percent of the AAT data, where the lowest bin value represented the 5 percent lowest AAT values and the highest bin value the 5 percent highest AAT values (Figure 2(B)).For each hemodynamic parameter, the mean value for each of these binned ROIs was calculated and used in the subsequent correlation analysis to assess whether the hemodynamic parameters in each AAT bin correlate with other hemodynamic parameters.In order to compare CBF values to the CVR metrics as well, this process was repeated using the CBF maps to bin the data.As this binning resulted in 20 mean values per individual (i.e. for each binned ROI), a repeated-measures correlation was chosen to account for the withinindividual variance between these values.The 'rmcorr' package implemented in RStudio (Version 2021.9.1.372)was used to perform this repeated-measures correlation. 34This analysis estimates the common regression slope for repeated measures.
Lastly, a Pearson's correlation coefficient was calculated for the voxelwise association between the ASL and CVR metrics for each individual.This correlation was performed separately for non-vascular steal and vascular steal regions.Next, we performed a nonparametric Wilcoxon-signed rank-test to assess whether the correlation values significantly differed from zero on the group-level (See Supplementary Results).For all statistical tests a p-value of 0.05 was deemed significant.

Participants
In total, 20 patients were included in the APRICOT study in the given time period.Data of 14 of these participants was used in the current analysis and 6 patients were excluded due to various, mostly technical reasons (see Supplementary Table 1).The average age of the included patients was 65.1 years and four were female.On average patients had 5.4 brain metastases of which the average volume of the newly treated, nonresected brain metastases was 10.7 cc.Edema encompassed on average 23.93 cc of the brain.Two patients had previous resection of at least one BM and three patients had received previous brain radiotherapy.On average, 62.30 cc of the brain volume (5.7% of the total brain volume) showed steal phenomena as indicated by negative CVR values (see Table 1).On average 13.4% (range: 0-43.6%) of the brain metastases volume and in 21.9% (range: 0-52.3%) of the edema volume showed steal.

Mean group comparison
For each hemodynamic parameter map, the group mean values were compared between steal versus non-steal regions and between different tissue types (GM, WM, edema and brain metastases) using Kruskal-Walis tests (see Figure 3 and Supplementary Table 2).For the AAT, steal regions showed longer AAT than non-steal regions, but only in WM regions (p ¼ .015).Additionally, for both non-steal and steal regions AAT values in GM were lower than in WM (non-steal: p ¼ 0.022; steal: p ¼ .004),edema (non-steal, p ¼ .006;steal, p ¼ .001) or brain metastases (non-steal, p ¼ .018;steal, p ¼ .007).AAT values in WM, were lower than in edema in steal regions (p ¼ .031),but did not differ from edema in non-steal regions (p ¼ 0.37) or brain metastases (non-steal: p ¼ .111;steal p ¼ .143).
As steal regions were defined using the negative CVR values, all steal and non-steal CVR values significantly differed from each other (p < .001).In non-steal regions, GM showed a higher reactivity than WM (p ¼ .002),edema (p < .001)and brain metastases (p ¼ .009).WM also had a higher reactivity than edema (p < .001).Within steal regions, GM had a larger negative response than both WM (p ¼ .003)and edema (p < .001),but not brain metastases (p ¼ 0.61).Within these steal regions, WM had a more negative response than edema regions (p ¼ .027).
Next, Spearman's Rho correlations were performed to test for an association between the different MRI metrics within the tissue types.For the non-steal regions, in brain metastases only hemodynamic lag significantly negatively correlated with CVR values (r ¼ À0.73, p ¼ 0.009).For steal regions, longer AAT was significantly related to lower perfusion (r ¼ À0.534, p ¼ .048),but only within GM.See Supplementary Figure 2 for scatterplots of these associations.

Repeated measures correlation
Repeated measures correlations were performed between each MRI metric using either the AAT binned or CBF binned ROIs (see Figure 4 and Supplementary Figure 3).In non-steal regions, longer hemodynamic lag time was associated with both longer AAT (r rm ¼ 0.712, p < .001) as well as lower CBF (r rm ¼ À0.822, p < .001).No relationship between either AAT and hemodynamic lag (r rm ¼ 0.063, p ¼ .308)or CBF and hemodynamic lag (r rm ¼ À0.002, p ¼ 0.976) was found for steal  regions.Higher CVR was related to shorter AAT (r rm ¼ À0.833, p < .001) as well as higher CBF (r rm ¼ 0.792, p < .001) in non-steal regions.Within the steal regions this association flipped with stronger negative CVR (i.e.stronger vascular steal effect) related to shorter AAT (r rm ¼ 0.320, p < .001)and higher CBF (r rm ¼ À0.273, p < .001).Additionally, similar associations can be visually observed in the representative example patient (see Figure 5).

Discussion
In the current study, baseline perfusion measurements were compared with functional vascular parameters acquired using CVR in a population of patients with brain metastases.These parameters were compared both spatially, and across different tissue types (GM, WM, edema and brain metastases), and also tissues exhibiting different hemodynamic characteristics (steal versus non-steal regions).When visually inspecting the MRI data, all vascularly compromised regions visible in the ASL data were also reflected in the CVR metrics.Results indicate a strong, positive relationship between regional baseline CBF and CVR.Additionally, there was a strong positive relationship between the temporal metrics AAT and hemodynamic lag.In both instances, however, these relationships do not hold in brain regions with exhausted cerebral autoregulation (i.e.steal regions).Thus, while ASL may be able to inform on some functional vascular aspects, CVR provides additional information regarding brain tissue at risk as indicated by either vascular steal or increased hemodynamic lag.Tissue-specific hemodynamic characteristics were observed in our results.WM areas were characterized by lower CBF and longer AAT, which has been previously reported in healthy subjects. 35Additionally, WM areas showed lower reactivity as well as longer hemodynamic lag values than GM.This is consistent with existing literature, and reflects lower WM blood volume, different venous draining architecture and/or potential flow redistribution resulting from the strong GM CO 2 response. 29,31,36Our findings suggest that the hemodynamic characteristics of healthy tissue are preserved in the normal appearing brain tissue of patients with brain metastases .On the other hand, large variability in the hemodynamic vascular parameters was seen within tissue containing untreated brain metastases.This was reflected in both the ASL and the CVR parameters.High variability in CBF measurements has previously been shown both between patients with brain metastases as well as between multiple brain metastases within patients. 37The primary tumor (i.e.origin of metastatic cells), might cause some of this heterogeneity as it affects the brain metastases growth pattern. 15,38,39Our findings additionally show that edema regions are characterized by lower CBF, lower CVR and both longer AAT and hemdoynamic lag times.Previous literature has also found impaired CVR within perifocal edema surrounding diffuse gliomas. 13It could be speculated that the local pressure in edematous regions restricts the ability of vessels to dilate and thus cannot maintain adequate perfusion.Additionally, this could limit the vasodilatory reserve capacity, as reflected by the CVR parameters.
GM CBF values varied between 30 and 50 ml/ 100 ml/min in our patient sample, which is similar to the GM CBF values as measured using PET. 40CBF and CVR values were strongly correlated within adequately functioning regions, both when performing ROI and voxelwise statistics.That is, regions with lower blood supply also showed lower vascular reactivity.This was to be expected since CVR measurements are heavily dependent on changes in CBF.Additionally, longer hemodynamic lag times were seen within regions with lower CBF.2][43][44] In the current study, however, in regions exhibiting vascular steal the negative CVR decreased further with increasing CBF.Even though the effect size of this relationship is considered small, 45 this result is still striking.Previous research has likewise postulated that the relationship between CBF and CVR may be dependent on the staging of vascular reserve. 8This could be explained by the following: within non-steal regions (i.e.adequately responsive regions), the process of cerebral autoregulation ensures CBF is maintained through varying the vasodilation of the vessels.However, in vascular steal regions, vessels are likely to be maximally dilated.When this maximum vascular reserve capacity is able to compensate, adequate CBF could be maintained while CVR is negative.In even more vascularly comprised regions, this maximum vasodilation fails to compensate, leading to both low CBF and CVR values.These different associations could be reflective of the underlying vascular reserve clasifications as proposed by both Derdeyn and colleagues 46 or Kuroda and colleagues 47 .
The timing of the response as measured with CVR was related to the temporal metrics of the baseline CBF (i.e.AAT).In other words, regions in which the arrival time of labeled blood was longer also showed a longer vascular response delay to a hypercapnic stimulus.Since the timing of the CVR is related to the traveling duration of the blood, this positive relationship was to be expected.However, both the correlation analyses as well as visual inspection of the data suggest these two temporal measures are not identical.Within the same region, the hemodynamic lag time is longer than the AAT.This probably reflects the difference in the two temporal metrics.AAT solely reflects the arrival time of labeled blood in a physiological steady state on the arterial side.Hemodynamic lag, on the other hand, is a time measure in reaction to a hypercapnic stimulus, measured at the venous side and influenced by multiple factors like blood redistribution and vascular response speed. 29While prolonged AAT might be able to indicate areas with possible increased vascular collateralization 48 and thereby also longer hemodynamic lag values, AAT cannot be used to identify regions at risk for decreased vascular response speed.
When comparing ASL with BOLD MRI data, some technical limitations of both techniques should be considered.The main limitation of ASL is the short halflife of the endogenous tracer ($1-2 seconds), leading to possible underestimation of perfusion in areas with long transit delays. 49Additionally, the specific ASL Van Grinsven et al.
technique may influence the resulting perfusion maps.An example is that too short post-labeling delays can lead to arterial transit artefacts or less accurate CBF measures. 3,50,51In the current study, we used a multi post-labeling delay ASL pcASL sequence and did not perform partial volume correction leading to CBF values that are close to the ground truth as measured using PET.A disadvantage of BOLD-metrics, is that the BOLD response is dependent on multiple factors, like changes in cerebral blood volume, cerebral metabolic rate of oxygen consumption, arterial partial pressure of oxygen and baseline parameters like hematocrit, OEF, CMRO 2 , and blood volume. 52,53This makes it difficult to pinpoint the exact underlying mechanisms and could be viewed as a less pure measure of brain hemodynamics when compared to ASL.Nevertheless, on average the change in BOLD to changes in PetCO 2 seem to be dominated by CBF changes. 42,54,55Thereby it could be speculated that the BOLD response is mainly affected by changes in CBF and provides a good comparison with ASL MRI.
Additionally, specific analyses choices could have influenced results.In areas with vascular steal, low variability in CVR values in combination with highly variable hemodynamic lag values were observed.As steal regions were defined by any CVR value below zero, these regions could thereby include voxels containing noisy, just below threshold CVR values, thereby influencing the results.Moreover, as the current analysis pipeline is influenced by the time to peak for the hemodynamic lag calculation, it may be difficult to discern the actual lag time of the response.Thereby hemodynamic lag values will not always reflect true underlying temporal characteristics of tissue reactivity.Therefore, we advise not to perform lag analysis within regions indicated by vascular steal as it will possibly lead to spurious lag quantifications.
To further understand the exact hemodynamic mechanisms, oxygen extraction fraction (OEF) would be a valuable additional metric.To illustrate, in regions with exhausted cerebral autoregulation, a further reduction in cerebral perfusion pressure will cause CBF to drop.In order to maintain tissue function, OEF can be increased.The functional consequences of reduced regional CBF will therefore only become apparent when OEF is maximal. 4Thus, future studies should add OEF measurements to further understand whether tissue at risk as indicated by either ASL or CVR is also reflective of tissue with maximized OEF.To fully understand the functional consequences of these hemodynamic measure, the MRI metrics should be related to behavioral measures, like cognitive performance.
In the current study, we investigated how ASL parameters measured during a physiological steady state relate to functional vascular parameters as measured using CVR in a patient population with brain metastases.When visually inspecting the MRI data, all vascularly compromised regions visible in the ASL data were also reflected in the CVR metrics.This was confirmed by both the regional and voxelwise relationship between on the one hand CBF and CVR measurements and the temporal metrics of ASL and CVR on the other hand.However, the relationship between ASL and CVR measures seems to be dependent on the vascular status of the underlying tissue.That is, relationships do not hold in tissues exhibiting vascular steal.Thus, CVR metrics may be able to flag at-risk areas before they become visible on ASL MRI.However, the downside of using BOLD-metrics is that they are influenced by multiple variables, making it difficult to pinpoint the exact mechanisms underlying this vascular risk.Consequently, to fully understand vascular changes within patients with pathology, combining ASL and CVR will provide a more complete picture.

Funding
The APRICOT study described in the current manuscript, as well as the first author (Eva van Grinsven), were supported by research funding from the Dutch Cancer Society "Koningin Wilhelmina Fonds (KWF)" (#11110).Alex Bhogal is supported by a Dutch research council talent grant (NWO VENI: The ischemic fingerprint, [VI.VENI.194.056].The funding sources had no role in conducting this research or preparation of this manuscript. Perfusion data was acquired using a multi-PLD ASL sequence.A whole volume was acquired at each of the 4 post-labeling delays(660, 1325, 1989, 2654 ms), using a pCASL Look-Locker multi-slice EPI read-out (total scan time ¼ 240 s, labelling train duration ¼ 1650 ms, TR ¼ 5 s, TE ¼ 12 ms, flip angle 25 , acquired resolution: 3 Â 3 Â 7 mm 3 , acquisition matrix: 80 Â 80 Â 17, 23 averages, SENSE factor ¼ 2, 2 background suppression pulses).A total of 23 label-control pairs were acquired.The first acquired ASL dataset pair here has the labelling, saturation pulses and background suppression turned off and are the M0 images for each post-label delay.An additional dataset pair was acquired for EPI phase (distortion) correction.No breathing challenges were performed during ASL imaging.The ASL was planned using a phase contrast angiography scan, with the labeling plane carefully placed perpendicular to the internal carotid arteries and vertebral arteries.Additionally, a 3 D spoiled gradient echo (SPGR) sequence (TR ¼ 8 ms, TE ¼ 3.25 ms, flip angle 10 , isotropic resolution 1 mm, acquisition matrix 240 Â 240 Â 180), a 3 D T2-weighted FLAIR sequence (TR 4800 ms, TE 240 ms, TI ¼ 1650 ms, flip angle 90 , isotropic resolution: 1 mm, acquisition matrix 256 Â 256 Â 182), as well as a SWI (TR 50 ms, flip angle 17 , inplane slice thickness 2 mm, acquisition matrix 384 Â 383 Â 63) were acquired as an anatomical reference.A minimum-intensity-projection was constructed from the SWI data, using an in-house developed Matlab script (Matlab R2020a, The MathWorks, Inc., Natick, Massachusetts, United States).

Figure 1 .
Figure 1.CO 2 breathing trace of a representative subject during the breathing challenge.The blue line indicates the actual measured PetCO 2 values, while the red line indicates the desired values based on the protocol.PetCO 2 : end-tidal CO 2 .

Figure 2 .
Figure2.Data analysis steps for the BOLD and ASL MRI data. 1. BOLD time series and PetCO 2 traces were used to generate CVR and hemodynamic lag maps.2. Quantitative AAT and CBF maps were generated based on Multi-delay ASL data 3. CSF and previously resected or irradiated brain metastases were excluded from all the maps.4. Based on the CVR values, the brain was segmented into non-steal and steal regions of interest (ROI).Parameter maps were analyzed comparatively (shown in grey).For analysis A, the brain was divided into GM, WM, edema and brain metastases ROIs for each individual and the mean and standard deviation were calculated for each hemodynamic parameter map (CBF, AAT, CVR and hemodynamic lag) for each tissue-ROI.For analysis B, the AAT data were sorted and the AAT maps were divided into ROIs each containing 5% of the sorted data ('5%-bins' -visualized by the colored histogram) for each individual.For each bin-ROI the mean value of each hemodynamic parameter map (CBF, AAT, CVR and hemodynamic lag) was calculated and used in subsequent repeated-measures correlation analysis.AAT: arterial arrival time; ASL: arterial spin labeling; BM: brain metastases; BOLD: blood oxygenation level-dependent; CBF: cerebral blood flow; CSF: cerebrospinal fluid; CVR: cerebrovascular reactivity; GM: grey matter; PetCO 2 : end-tidal CO 2 ; ROI: region of interest; WM: white matter.

20 Figure 4 .
Figure 4. Repeated measures correlation plots between ASL metrics (AAT, CBF) and CVR metrics (CVR, hemodynamic lag) using AAT-binned or CBF-binned ROIs.Each color represents a different patient and the lines represent the common linear relationship between the MRI metrics when all participants are taken into account.Statistics for each relationship are provided for each plot separately.Correlations between ASL metrics (AAT and CBF) and CVR metrics (CVR and hemodynamic lag) were also performed and are shown in Supplementary Figure3.AAT: arterial arrival time; CBF: cerebral blood flow; CVR: cerebrovascular reactivity; ROI: region of interest.

Figure 5 .
Figure 5. Visual comparison of hemodynamic parameter maps of a representative subject.Red voxels ona the T2 FLAIR represent the location of the brain metastases, and yellow voxels are regions with edema.The minimum intensity projection (SWI-mIP) indicates the locations of larger cerebral venous vessels.AAT: arterial arrival time; CBF: cerebral blood flow; CVR: cerebrovascular reactivity.

Table 1 .
Patient, treatment and volume information of the included subjects in this study.