Acetazolamide-augmented BOLD MRI to Assess Whole-Brain Cerebrovascular Reactivity in Chronic Steno-occlusive Disease Using Principal Component Analysis

Abstract

Background

Exhaustion of cerebrovascular reactivity (CVR) portends increased stroke risk. Acetazolamide-augmented blood oxygenation level–dependent (BOLD) MRI has been used to estimate CVR, but low signal-to-noise conditions relegate its use to terminal CVR (CVR_end) measurements that neglect dynamic features of CVR.

Purpose

To demonstrate comprehensive characterization of acetazolamide-augmented BOLD MRI response in chronic steno-occlusive disease using a computational framework to precondition signal time courses for dynamic whole-brain CVR analysis.

Materials and Methods

This study focused on retrospective analysis of consecutive patients with unilateral chronic steno-occlusive disease who underwent acetazolamide-augmented BOLD imaging for recurrent minor stroke or transient ischemic attack at an academic medical center between May 2017 and October 2020. A custom principal component analysis-based denoising pipeline was used to correct spatially varying non–signal-bearing contributions obtained by a local principal component analysis of the MRI time series. Standard voxelwise CVR_end maps representing terminal responses were produced and compared with maximal CVR (CVR_max) as isolated from binned (per-repetition time) denoised BOLD time course. A linear mixed-effects model was used to compare CVR_max and CVR_end in healthy and diseased hemispheres.

Results

A total of 23 patients (median age, 51 years; IQR, 42–61, 13 men) who underwent 32 BOLD examinations were included. Processed MRI data showed twofold improvement in signal-to-noise ratio, allowing improved isolation of dynamic characteristics in signal time course for sliding window CVR_max analysis to the level of each BOLD repetition time (approximately 2 seconds). Mean CVR_max was significantly higher than mean CVR_end in diseased (5.2% vs 3.8%, P < .01) and healthy (5.5% vs 4.0%, P < .01) hemispheres. Several distinct time-signal signatures were observed, including nonresponsive; delayed/blunted; brisk; and occasionally nonmonotonic time courses with paradoxical features in normal and abnormal tissues (ie, steal and reverse-steal patterns).

Conclusion

A principal component analysis-based computational framework for analysis of acetazolamide-augmented BOLD imaging can be used to measure unsustained CVR_max through twofold improvements in signal-to-noise ratio.

© RSNA, 2023

Supplemental material is available for this article.

Summary

Application of principal component analysis to acetazolamide-augmented blood oxygenation level–dependent (BOLD) MRI in the setting of chronic steno-occlusive disease enabled dynamic whole-brain analysis of cerebrovascular reactivity, which yielded twofold improvement in signal-to-noise ratio.

Key Results

• ■ In a retrospective analysis of 23 patients with unilateral chronic steno-occlusive cerebrovascular disease who underwent acetazolamide-augmented blood oxygenation level–dependent (BOLD) MRI for recurrent minor stroke or transient ischemic attack, a data-driven method to measure whole-brain cerebrovascular reactivity (CVR) yielded twofold signal-to-noise ratio improvements.

• ■ The CVR response is often nonmonotonic and unsustained, with distinct maximal and terminal values.

• ■ Interrogation of dynamic signal time courses showed distinct kinetic signatures, including steal and reverse-steal patterns.

Introduction

Cerebrovascular reactivity (CVR) reflects the autoregulatory capacity of the cerebral circulation in response to falling perfusion pressures (1). CVR exhaustion portends elevated stroke risks in the setting of steno-occlusive disease (2,3).

Historically, CVR has been assessed through paired blood flow measurements obtained before and after hemodynamic stimuli, including hypercapnia or acetazolamide (ACZ), and acquired using lipid-soluble radiotracers, magnetically tagged blood water, or dynamic passage of nondiffusible gadolinium-containing or iodinated contrast material (1,4–10). Continuous ACZ-augmented blood oxygenation level–dependent (BOLD) MRI has been reported for this purpose (11); however, the low signal-to-noise ratio (SNR) of the signal time course has relegated the technique to characterization of terminal CVR (CVR_end) responses, measured at arbitrarily defined conclusions of the pharmacologic stimulus relative to pre-ACZ baselines (12).

Despite a precedent for performing imaging 10–20 minutes after ACZ administration, characterization of only one post-ACZ terminus at any arbitrarily selected time is anticipated to impose bias in the results, depending on local hemodynamics that may vary nontrivially between tissues. Implicitly, such paradigms presuppose the response to be monotonically increasing in all tissues, a condition which, to date, has not been amenable to rigorous experimental validation due to lack of pragmatic dynamic nonionizing measures of whole-brain CVR.

The low SNR of BOLD signal time course typically obtained with T2*-weighted gradient-echo echo-planar imaging presents challenges to voxelwise sliding (ie, time resolved) windows, during which signal changes may be confounded by noise contributions, necessitating temporal averaging that undermines assessment of dynamic features. An ideal strategy would therefore precondition the BOLD time series to permit dynamic characterization of CVR to recover true maximal CVR (CVR_max) or other nuanced aspects of the signal time course, such as the response time to maximum (ie, time-to-CVR_max).

Thus, we aimed to demonstrate comprehensive characterization of acetazolamide-augmented BOLD MRI response in chronic steno-occlusive disease using a computational framework to precondition the BOLD signal time course for dynamic whole-brain CVR.

Materials and Methods

In our practice, BOLD MRI with acetazolamide is performed routinely for standard-of-care CVR estimations in patients with chronic steno-occlusive cerebrovascular disease through our neurologic, neurosurgical, and neurointerventional referral base. This Health Insurance Portability and Accountability Act–compliant retrospective study was approved by the institutional review board. Informed consent was waived for retrospective analysis.

Patient Selection

Patients with unilateral chronic steno-occlusive disease in any vascular territory who underwent acetazolamide-augmented BOLD MRI for recurrent minor stroke or transient ischemic attack at a single academic medical center (New York University Langone Health Center for Biomedical Imaging) between May 2017 and October 2020 were included. Examinations were prescreened to exclude prior surgery, prior infarctions (larger than lacunae), major anatomic deformities, and hemorrhage. Patients with unilateral chronic steno-occlusive disease were selected to facilitate internal normalization to the contralateral hemisphere in an exploratory study of dynamic CVR profiles vis-à-vis putatively normal hemispheres, defined by absence of any macrovascular steno-occlusive disease as determined with digital subtraction angiography, CT angiography, or both. Evaluation was initiated for suspected cerebrovascular events at the discretion of a vascular neurologist or cerebrovascular neurosurgeon (K.I.) and performed at least 30 days after an ischemic episode. All patients underwent CT angiography or digital subtraction angiography to document the presence, location, and severity of cerebrovascular stenoses. Angiographic findings were reviewed to assess for anterior communicating and posterior communicating arteries in all patients which, when present, were evaluated by a neuroradiologist with 12 years of experience (S. Dehkharghani) and blinded to CVR results, using a three-point visual rating scale as (a) small or absent, (b) moderate, or (c) large.

MRI and ACZ Challenge

Patients were scanned with a 3-T whole-body MRI system (Prisma; Siemens Healthineers) with a 64-channel head coil. The protocol included (a) a 1-mm isotropic T1-weighted magnetization-prepared rapid gradient-echo (MPRAGE) sequence (repetition time msec/echo time msec, 2300/2.9; flip angle, 9°; voxel size, 1 × 1 × 1 mm), (b) a 20-minute continuous BOLD sequence performed with a gradient-echo echo-planar imaging sequence (2000/40; one signal average; 90° flip angle, voxel size, 2.0 × 2.0 × 5.0 mm; whole-brain coverage), and (c) gradient-echo magnitude and phase field maps to correct for spatial distortions on BOLD images. ACZ (Diamox; Avet Pharmaceuticals; 1 g dissolved in 10 mL saline) was infused intravenously over 3–5 minutes after the initial 4–5-minute baseline BOLD sequence, without interruption of BOLD imaging upon or during injection. The total number of BOLD volumes and repetition times, and thus the effective post-BOLD duration, varied slightly between studies to ensure whole-brain coverage but exceeded 20 minutes in toto for all MRI examinations.

Image Processing

For each patient, images acquired with the T1-weighted MPRAGE sequence were analyzed using FreeSurfer software (version 7.1.1; Massachusetts General Hospital) for cortical and subcortical gray matter, white matter, and cerebellar segmentation (13). An in-house–developed denoising strategy was used for identification and removal of spatially varying non–signal-bearing noise-only eigenvalues obtained by a local principal component analysis of the MRI time series, as represented in the universal Marchenko-Pastur distribution or Marchenko-Pastur principal component analysis (MPPCA) (14–16). Denoised BOLD images were thereby preconditioned to allow fine (up to each repetition time or intervals thereof) dynamic analysis of sliding-window time-signal change across discrete epochs of the response. Denoised BOLD images and field maps were processed with FSL Melodic software (version 6.0.6; Analysis Group, FMRIB) for motion correction, unwarping, section-timing correction, and spatial smoothing (Appendix S1) (17). The initial and terminal 30 frames were discarded due to ringing artifact introduced by low-pass filtering (Fig 1).

Figure 1:

Flowchart of the study pipeline. For each patient, the T1-weighted magnetization-prepared rapid gradient-echo (MPRAGE) sequence was reconstructed using FreeSurfer software (version 7.1.1; Massachusetts General Hospital) for automatic segmentation of the cortical gray matter, white matter, subcortical gray matter, and cerebellum. A custom principal component analysis-based denoising strategy coupled to random matrix theory (Marchenko-Pastur principal component analysis) was used for noise removal from blood oxygenation level–dependent (BOLD) MRI. Denoised BOLD MRI scans were processed with FSL Melodic software (version 6.0.6; Analysis Group, FMRIB) for motion correction, unwarping, section timing correction, spatial smoothing, and temporal low-pass filtering. The initial and final 30 frames were discarded due to ringing artifact introduced by low-pass filtering. Labeled color bars refer to the range of cerebrovascular reactivity (CVR) values but are merely representative in this figure. CVR_end = terminal CVR, CVR_max = maximal CVR.

The quantitative impact of the pipeline on SNR characteristics in the signal time course was assessed. The number of preserved principal components was estimated using MPPCA, requiring no further processing to analytically compute the temporal gain (16). Empirical temporal SNR was calculated by estimating the noise level for each image using the high-frequency band of the wavelet decomposition, as reported elsewhere (18). To estimate the SNR improvements gained specifically by denoising with MPPCA and dissociated from other SNR gains occurring elsewhere in our processing pipeline, BOLD signal time course was generated using raw post-MPPCA only and separately with fully preprocessed voxel-level data across all examinations. Further details are provided in Appendix S2.

In preliminary analyses, varying low-frequency bands were filtered to assess the influence on observed CVR estimations, with consideration given to past studies exploring frequency ranges spanning 0.0 to 0.1 Hz (19,20). Across this range, greatest stability in CVR was observed at 0.01 Hz, which we elected to use in all subsequent analyses. Further detrending was not performed to preserve the linear component of CVR within the low-frequency band of interest.

Conventional terminal CVR_end maps were produced from initial and final 1-minute BOLD imaging and were averaged respectively as BOLD_pre and BOLD_post, as reported previously (12,21,22). CVR was calculated as follows: CVR = 100 × (BOLD_post – BOLD_pre)/BOLD_pre. CVR_max was extracted from the maximal prebinned (ie, per each repetition time) denoised BOLD signal time course per voxel. A linear boundary-based co-registration was used to align CVR maps with FreeSurfer-processed T1-weighted MPRAGE images (23).

Cerebral white matter CVR was calculated directly using the cortical white matter segmentation, while cerebral gray matter CVR was calculated as a volume-weighted average of cortical gray matter and subcortical gray matter structures. Hemispheric CVR was calculated as a weighted average of the white and gray matter.

Atlas-defined vascular territories of anterior, middle, and posterior cerebral arteries were segmented and applied to a common coordinate frame of reference (Montreal Neurologic Institute) (24). Flow-based territorial masks were inverse warped from the Montreal Neurologic Institute space to the individual T1 space using Diffeomorphic Anatomic Registration Through Exponentiated Lie Algebra, or DARTEL, (SPM12; Functional Imaging Laboratory Methods Group, University College London) and then linearly transformed to BOLD space (25). Average BOLD signal time course was demeaned for uniform scaling.

Windowed CVR maps were generated by averaging postprocessed CVR signals over 1-minute periods. Resulting maps were generated in FSLeyes (version 6.0.6) (26–28). Density charts were generated showing relative frequencies of different hemispheric averages for diseased and healthy tissue using ggplot2 in the R software package (version 4.0.5; R Core Team) (29).

Volumetric analysis in segmentations was performed with FreeSurfer aparc+aseg stats. Median heathy hemispheric CVR_max and CVR_end were calculated for each patient. The 1-mm³ voxels with CVR_max and CVR_end below 25%, 50%, 75%, and 100% of these respective medians were thresholded. At each threshold, percentage of the diseased hemispheric volume comprising the thresholded value was calculated.

Statistical Analyses

Statistical analyses were performed in R (version 4.0.5; R Core Team). Because the study sample was not preselected for known severity or distribution of hemodynamic compromise, no a priori assumptions were made regarding likelihood of significant differences between healthy and diseased hemispheres. Because some patients underwent multiple MRI examinations, a linear mixed-effects model was used to compare mean CVR_end and CVR_max in healthy and diseased hemispheres.

Vascular territory statistical analysis was performed only for middle cerebral artery (MCA) territories due to insufficient numbers of patients with anterior cerebral or posterior cerebral artery stenotic disease. Healthy MCA territory CVR was defined as mean CVR of normal MCA territories (healthy territory in patients with unilateral MCA disease and bilateral MCA territories in patients who had neither MCA nor ICA disease). Here again, a linear mixed-effects model was used to compare mean CVR_end and CVR_max in healthy and diseased vascular territories.

Mean time-to-CVR_max, expressed as the temporal volume (ie, repetition time) at which CVR_max occurred, was extracted and converted to minutes for comparison in a linear mixed-effects model.

For volumetric analysis, mean percentages of diseased hemispheric volume at CVR thresholds below the 25%, 50%, 75%, and 100% quantiles of the healthy hemisphere CVR_end and CVR_max were compared at each threshold using a linear mixed-effects model.

All statistical comparisons were two tailed, with statistical significance set at P < .05.

Results

Patient Characteristics

A total of 23 patients with unilateral chronic steno-occlusive disease (median age, 51 years; IQR, 42–61; 13 male) underwent 32 BOLD MRI examinations with ACZ. Among these examinations, 31 were deemed to have severe stenoses of the affected vessels. No patients or examinations necessitated exclusion. Patient demographics and clinical characteristics are shown in Table 1.

Table 1:

Patient Demographic and Clinical Characteristics, Including Presence of Anterior and Posterior Communicating Arteries

Processed MRI Signal Time Series

Figure 2 shows voxel-level native BOLD signal time course superimposed on the postprocessed signal time course. After undergoing the processing steps outlined in Figure 1, processed MRI data revealed improved demonstration of dynamic features, facilitating sliding window CVR_max. Comparable frame-by-frame analysis was not attainable from native BOLD images due to severe baseline fluctuations, as evidenced by Figure 3A, comparing the time-varying CVR_end (ie, terminal CVR produced from arbitrary premature conclusions) with true CVR_max (Fig 3B). Attenuation plots of hemispheric CVR are shown in Figure 3C, which shows that CVR_max plots yielded two distinct peaks and reduced overlap compared with CVR_end plots.

Figure 2:

Sample blood oxygenation level–dependent (BOLD) signal time course from two patients with right hemispheric disease. BOLD signal changes from signal course mean of a right hemispheric voxel from raw (gray) and processed (orange) signals show nonmonotonic post-acetazolamide effects. (A) Signal course with maximal cerebrovascular reactivity (CVR_max) greater than terminal cerebrovascular reactivity (CVR_end). (B) Signal course where CVR_max is equal to CVR_end. Note that a 0% change represents the voxel mean for its entire respective time course, hence the initial signal may begin above or below the voxel mean, for illustrative purposes.

Figure 3:

(A) Time evolution of terminal cerebrovascular reactivity (CVR_end) maps throughout the course of the scan of a patient with left steno-occlusive disease compared with (B) the maximal cerebrovascular reactivity (CVR_max) map of the same patient. Note that the sequential CVR_end maps in A represent arbitrarily defined prematurely terminated values. Despite their quasidynamic representation, no CVR_end terminus entirely shows that the computed CVR_end equates to the CVR_max, even during the including epochs (ie, for nonmonotonic responses). (C) Density plots of the population hemispheric averages of CVR_end (top) and CVR_max (bottom) show varying overlap in distinguishing diseased and healthy hemispheric distributions. The y-axis is a histogram of the voxels normalized such that the area under the receiver operating characteristics curve is 1. Dotted lines indicate medians of the respective distribution. Color bar indicates cerebrovascular reactivity ranging from -10% (light blue) to 10% (red).

SNR analysis indicated that, in excess of the base SNR increases inherent to our routine preprocessing steps, an additional twofold SNR boost was achieved in BOLD signal time course using the Marchenko-Pastur principal component analysis for denoising (Fig 4). In particular, the temporal signal-to-noise ratio (the ratio between the time-averaged MRI signal and the thermal noise level before and after denoising), computed voxelwise across all examinations, increased from a median of 109 before denoising to 205 after denoising.

Figure 4:

(A) Raw and (B) fully preprocessed image sections from a representative patient (51-year-old man) are shown. (C) Raw (green), post–Marchenko-Pastur principal component analysis (blue), and fully preprocessed cerebrovascular reactivity (CVR) time-signal curves of a representative voxel from this image section.

CVR Measurements

Mean CVR_max was significantly higher than mean CVR_end in diseased (5.2% vs 3.8%, P < .01) and healthy (5.5% vs 4.0%, P < .01) hemispheres. There was no significant difference between healthy and diseased hemispheres for mean CVR_end (4.0% vs 3.8%, P = .53); mean CVR_max (5.5% vs 5.2%, P = .16); middle cerebral artery territory mean CVR_end (3.2% vs 2.5%, P = .10) or mean CVR_max (4.4% vs 4.3%, P = .73); or time to CVR_max (16.90 minutes vs 16.73, P = .59) and with an inconsistent relationship to CVR (discussed later).

Brain Volume Calculations

Percentages of diseased hemispheric volume at CVR_end and CVR_max thresholds of less than 25%, 50%, 75%, or 100% quantile of healthy hemispheric CVR_end and CVR_max are shown in Table 2. CVR_end maps produced volumetrically larger abnormalities at lower thresholds (13% vs 6% at the <25% threshold with P < .001; 25% vs 22% at the 50% threshold with P < .001), suggesting that CVR_end may be particularly inaccurate in increasingly diseased tissues, as compared with CVR_max.

Table 2:

Mean Percentage of Diseased Hemisphere Volume Composed of Voxels with CVR_end and CVR_max below 25%, 50%, 75%, or 100% Quantile of Healthy Hemispheric CVR_end and CVR_max

Kinetic Signatures

Several representative patterns are demonstrated in Figure 5, showing demeaned MCA territory signal time course for representative patients with right-sided ICA disease. Figure 5A shows signal time courses with monotonic features, albeit with visibly delayed and dampened features in the diseased hemisphere. Signal time courses in Figure 5B reveal paradoxical elevations in the diseased hemisphere in this patient with a large anterior communicating artery, angiographically suggesting paradoxical siphoning of flow, and suggesting brisk ACZ effect to the diseased hemisphere (ie, Robinhood effect) on BOLD MRI scans obtained with acetazolamide. Signal time courses in Figure 5C show prolonged Robinhood effects in a patient without a large anterior communicating artery but who possessed a large fetal posterior cerebral artery in the normal hemisphere and an absent posterior communicating artery on the right side, in combination with well-preserved anterograde flow through an ICA stenosis on digital subtraction angiograms. In aggregate, the features were suspected to impose larger baseline demands on the left MCA in its support of multiple territories. The effect in this case persists to the terminus, even as the normal hemisphere exhibits the earlier and brisk pharmacologic response, yielding an initial but transient steal phenomenon in the diseased hemisphere, prior to the exuberant augmentation occurring later in the diseased territory.

Figure 5:

(A–C) Demeaned signal time courses from three different patients with angiographically proven right middle cerebral artery (MCA) disease from healthy (blue) and diseased (red) MCA territories. Time-signal courses are shown as demeaned values, where the temporal mean was removed from the data prior to plotting, to show both time curves in a comparable y-axis range to allow for comparison of the kinetics. The y-axis represents blood oxygenation level–dependent signal normalized to the mean value.

Discussion

While cerebrovascular reactivity (CVR) is an increasingly studied hemodynamic parameter with implications for cerebrovascular ischemia, poor signal-to-noise conditions relegate quantification to that of terminal CVR measurements (CVR_end). We report development of a custom computational pipeline for enhanced blood oxygen level–dependent (BOLD) signal time course analysis in response to global hemodynamic stimuli, enabling calculation of examination-level maximal CVR (CVR_max) in a real-world clinical cohort of treatment-naïve, angiographically confirmed unilateral steno-occlusive disease. We show distinct CVR_max exists and significantly exceeds CVR_end in both diseased (0.052 vs 0.038, P < .01) and healthy (0.055 vs 0.040, P < .01) hemispheres, contradicting previous assumptions of sustained monotonic CVR responses.

CVR measurements have been an intriguing probe of cerebrovascular integrity for decades and have been used to predict future ischemic stroke in patients with hemodynamic stress. Radioisotopes and single-photon emission CT provide neither sufficient anatomic detail nor sufficient signal to enable rapidly resolved profiles. Serial dynamic bolus passage cannot be performed within reasonable limits of radiation or contrast material dose to provide fully temporally resolved responses. Endogenous contrast from spin tagging of diffusible blood water (ie, arterial spin labeling) offers flexibility for repeated measurements with neither dose nor radiation penalty, but delay dependency and SNR limitations leave conventional approaches vulnerable to confounding biases in slow-flow and diseased territories (12,22,30–32). We have previously reported on the use of continuous BOLD imaging coupled to ACZ for such purposes.

All ACZ-based strategies share the limitation that CVR measurements are derived from arbitrarily selected post-ACZ phases and there is often a 10–20-minute delay (4,33). By comparison with terminal CVR maps, important distinctions in CVR_max bear emphasis. Specifically, the augmentation plateau in CVR_end enforces an arbitrary examination terminus and then assesses CVR_end at that end point but neglects to address if that arbitrary terminus is indeed the maximal. Further, because of SNR constraints, lengthy signal averages are required to compute any such terminal levels of augmentation, neglecting frame-by-frame signal evolutions that may conceal valuable hemodynamic features. CVR_max, by comparison, yields a temporally unbiased view of augmentation and its specific timing.

CVR_max analysis produces more nuanced characterization of pharmacologic response, free of assumptions regarding optimal post-ACZ timing. Our results present a challenge to the paradigm for CVR as a monotonically increasing process. Consequently, studies using only terminal CVR_end cannot be assumed to equally weigh optimal timing for the response across tissues, undermining inferences regarding hemodynamic stress or compensation and yielding spurious volumetric abnormalities across CVR thresholds. Our findings suggest that bias imparted through CVR_end may operate in either direction, with unpredictable magnitude, and potentially on either normal or abnormal tissues.

Our pipeline incorporates denoising using the MPPCA, improving voxel-level average temporal SNR twofold. As the ratio of the number of preserved and total principal components is directly proportional to the reduction in noise variance, temporal SNR gain can be computed without impact by physiologic noise, patient motion, or other intrinsic sources of variability. Prior investigation suggests that motion correction of two-dimensional echo-planar images is rendered more accurate and precise by MPPCA prior to such preprocessing steps, and thus using MPPCA may allow synergistic improvements to temporal SNR (34).

Our study had several limitations, including its retrospective nature. Second, while our study sample was comparable with that of similar studies (11,12,35), it may have been underpowered to identify certain interactions between parameters. Importantly, patients were not pre-selected based on any known hemodynamic decompensation; thus, they were not known a priori to harbor major, if any, CVR deficits prior to examination. Third, the prognostic value of CVR_max remains indeterminate, as sufficient longitudinal follow-up data were unavailable, and the performance of CVR_max as a prognostic marker is consequently not fully established. Additionally, ACZ pharmacokinetics, while producing generally sustained pharmacologic responses, are unsuited to ensure uniform delivery after intravenous infusion. Lastly, lack of pragmatic reference standards impeded quantitative calibration of CVR.

In conclusion, application of principal component analysis to acetazolamide-augmented blood oxygenation level–dependent MRI in chronic steno-occlusive disease enabled dynamic whole-brain cerebrovascular reactivity (CVR) analysis, yielding twofold signal-to-noise ratio improvement. Although this work comprises a preliminary exploration, we anticipate that the described technique presents a valuable path forward in cerebrovascular studies. Maximal CVR (CVR_max) reduces volumetric abnormalities relative to terminal CVR (CVR_end) in diseased tissue and enables the characterization of previously unattainable dynamic features in the hemodynamic response. Kinetic parameters introduced in this study, such as time to CVR_max, may also offer paths to further study acetazolamide pharmacokinetics. Additionally, while we elected to use a frequency of 0.01 Hz to filter signal time courses, the potential influence of this frequency band on the results deserves dedicated study. Conditions necessary for observed reverse or counter-steal (ie, Robinhood) phenomena are currently unclear; robust cross-hemispheric collaterals (eg, a large anterior communicating artery) have been proposed as supporting efficient hemodynamic equilibria (36), and while commonly present in our experience, were not immutably the case. Alternative mechanisms should be sought. Unique hemodynamic signatures in the dynamic CVR response require further longitudinal evaluation to assess their prognostic significance and characterization to codify such profiles more formally.