Quantitative perfusion mapping with indu

1 |. INTRODUCTION

Gadolinium-based dynamic susceptibility contrast (DSC) imaging is the most commonly used methodology to assess cerebral perfusion and characterize tumor hemodynamics in clinical studies.¹ With the injection of a gadolinium exogenous contrast, DSC uses a series of echo-planar gradient-echo acquisitions to monitor the first pass of this contrast bolus through the capillary beds and derives an assortment of cerebral perfusion metrics, including cerebral blood flow (CBF), cerebral blood volume (CBV), and mean transit time (MTT).² The strength of this technique lies in its high temporal resolution to capture the bolus passage through brain tissue on the order of a few seconds and its ability to generate hemodynamic information within the same scan. In the past few decades, DSC has proven to be a valuable tool for evaluation of ischemic and infarcted tissues in stroke,³ differential diagnosis of intracranial masses,⁴ and assessment of blood-brain barrier permeability in traumatic brain injury.⁵

However, despite its utility in numerous clinical applications, DSC is limited by its reliance on various gadolinium-based contrast agents, each of which has a known or hypothetical association with nephrogenic systemic fibrosis in patients with kidney disease.^6,7 Even in patients with normal renal function, gadolinium agents have been demonstrated to accumulate in tissues in the brain, bone, and kidney.^8–10

In recent years, oxygen has been used as a safer alternative to exogenous contrast,¹¹ because the over-supply (hyperoxia) or the lack of oxygen (hypoxia) can change the blood oxygen content, modulate the deoxyhemoglobin concentration, and cause a T2-weighted change of the measured MR signal. MacDonald et al have previously demonstrated the use of hyperoxia-induced signal dynamics to estimate perfusion parameters.¹² However, Losert et al showed hyperoxia causes a relatively small increase in oxygen saturation and percent signal change in the white matter (WM) (less than 1%),¹¹ which prevents accurate flow measurement and examination of the role of compromised CBF in white matter disease.¹³ Therefore, we conjectured that briefly increasing the concentration of paramagnetic deoxyhemoglobin via desaturation would produce a larger and more robust signal change, especially in the white matter.

In this study, we explored the feasibility of using a transient hypoxia gas paradigm, in a similar manner to gadolinium injection, to deliver a bolus of paramagnetic material (deoxygenated hemoglobin) to the cerebral vasculature. The contrast passage through the microvasculature was monitored by a dynamic blood-oxygen-level-dependent (BOLD) MR sequence. Through tracer kinetics modeling, the BOLD signal change induced by this hypoxic bolus can be used to generate regional perfusion maps; we term this approach deoxygenation-based DSC (dDSC). In order to test the new technique over a wide range of CBF, we evaluated this technique on a total of 66 subjects, including healthy controls as well as anemic patients (sickle cell disease and non-sickle anemia) who have higher baseline blood flow.^14,15 Comparison with other popular CBF quantification methods, such as arterial spin labeling (ASL)¹⁶ and phase contrast (PC),¹⁷ can yield further insights into the feasibility of using transient hypoxia as a contrast mechanism for dDSC perfusion studies.

2 |. METHODS

2.1 |. Theory

In DSC imaging, the changes in relaxation rate $Δ R_{2}^{*}$ are directly proportional to the gadolinium concentration.^18,19 In parallel, this dDSC approach assumes that the changes of oxygen level in the brain can be approximated by a linear function of $Δ R_{2}^{*}$ during the gas paradigm:

where C_tissue(t) is the oxygen concentration time curve measured in the tissue and k is a proportionality constant that depends on tissue type, contrast agent, field strength, and pulse sequence.²⁰ In this experiment design, voxel-wise $Δ R_{2}^{*}$ is calculated with:

Δ R_{2}^{*} (t) = - \frac{1}{T E} \ln (\frac{S (t)}{S_{0}})

(2)

where S(t) is the gradient-echo time series, S₀ is the baseline signal computed from the first 25 dynamics prior to bolus administration and TE is the echo time. Arterial input function (AIF) and venous output function (VOF) are extracted from the middle cerebral artery (MCA) (Figure 3A) the superior sagittal sinus (Figure 3B) respectively. Due to difficulty extracting a pure intravascular blood signal from the MCA because of partial volume effect,²¹ the time integral of the AIF is scaled to match that of the VOF²² to yield an “effective AIF” signal eAIF(t) such that:

\int Δ R_{2, e A I F}^{*} (t) = \int Δ R_{2, V O F}^{*} (t) .

(3)

Regional CBV is computed using $Δ R_{2}^{*}$ from the tissue and effective AIF:

C B V = \frac{κ}{ρ} \frac{\int Δ R_{2, t i s s u e}^{*} (t)}{\int Δ R_{2, e A I F}^{*} (t)}

(4)

where ρ is the brain density 1.05 g/mL and κ is the hematocrit correction factor. The measure $K = \frac{1 - H c t}{1 - 0.69 H c t}$ accounts for the difference between the hematocrit in the large vessels (MCA for AIF determination) and in the microvasculature.²³ From tracer kinetics modeling, the measured C_tissue(t) is expressed as:

C_{t i s s u e} (t) = eAIF (t) \otimes (C B F \times R (t))

(5)

where R(t) is the residue function of the contrast fraction present in the vasculature at time t with R(0)=1 and R(∞)=0.¹⁸ To compute CBF, a deconvolution can be performed by an inverse Fourier transform approach²⁴ or a singular value decomposition (SVD) method proposed by Ostergaard et al²⁵ Stable solutions for inverse discrete Fourier transforms can only be obtained with effective experimental noise suppression, whereas SVD has been shown to reproduce flow and is independent of underlying vasculature and volume.²⁵ In this study, we used the SVD deconvolution approach, with the previous convolution discretized in the following form:

[\begin{matrix} C_{t i s s u e} (t_{1}) \\ C_{t i s s u e} (t_{2}) \\ ⋮ \\ C_{t i s s u e} (t_{N}) \end{matrix}] = Δ t [\begin{matrix} e A I F (t_{1}) & 0 & \dots & 0 \\ e A I F (t_{2}) & e A I F (t_{1}) & \dots & 0 \\ \dots & \dots & \dots & \dots \\ e A I F (t_{N}) & e A I F (t_{N - 1}) & \dots & e A I F (t_{1}) \end{matrix}] [\begin{matrix} C B F \times R (t_{1}) \\ C B F \times R (t_{2}) \\ ⋮ \\ C B F \times R (t_{N}) \end{matrix}] .

(6)

After fitting CBF × R(t) with SVD, CBF is determined as the initial height of this tissue response function. Finally, MTT is calculated as a direct ratio between regional CBV and CBF:

Localization of input functions from the difference image created by subtraction of baseline and hypoxic gradient-echo images. A, Representative individual AIF extracted from MCA. B, Representative individual VOF extracted from superior sagittal sinus. C, Representative AIF and VOF during hypoxia demonstrated the process of rescaling the AIF by the time integral of the VOF to minimize partial volume effects

2.2 |. Study protocol

The Committee on Clinical Investigation at Children’s Hospital Los Angeles (CHLA) approved the protocol; written informed consent and/or assent were obtained from all subjects (CCI#2011–0083). This study was performed in accordance with the Declaration of Helsinki.

A total of 66 subjects were tested between April 2012 and December 2017. This cohort was divided into two patient groups: 28 healthy controls and 38 chronically anemic patients (including sickle cell anemia, thalassemia. and other anemia disorders). Exclusion criteria were prior neurologic insult, pregnancy, acute chest or pain crisis hospitalization within 1 month, and major medical problems outside of their chronic anemia. Only subjects older than 12 years of age were included in the study. Imaging, vital signs, and blood samples were obtained on the same day for each subject. Complete blood count was analyzed in our clinical laboratory. Demographic and clinical variables of each patient group are summarized in Table 1.

TABLE 1.

Patient demographic and baseline flow data

PC analysis	Control (N = 28)	Anemia (N = 38)	P-value
Age (years)	26.1 ± 10.8	21.6 ± 8.4	.06
Sex	7M, 21F	17M, 21F	.10
Hemoglobin (g/dL)	13.4 ± 1.2	10.1 ± 1.8	<.01
Hematocrit (%)	39.7 ± 3.4	29.5 ± 5.0	<.01
CBF_PC (mL/100g/min)	60.7 ± 9.1	91.1 ± 22.3	<.01
ASL analysis	Control (N = 16)	Anemia (N = 25)
Age (Years)	23.4 ± 9.1	22.4 ± 8.9	.74
Sex	5M, 11F	10M, 15F	.58
Hemoglobin (g/dL)	13.4 ± 1.2	10.1 ± 2.2	<.01
Hematocrit (%)	40.1 ± 3.8	29.8 ± 5.9	<.01
CBF_ASL GM (mL/100 g/min)	60.9 ± 11.4	77.3 ± 30.0	.02
CBF_ASL WM (mL/100 g/min)	37.5 ± 10.9	42.1 ± 10.9	.20

2.3 |. Transient hypoxia

Figure 1 illustrates the experimental setup for the respiratory challenge MRI. At the start of the image acquisition, patients were breathing through a custom, 2-L reservoir rebreathing circuit supplied by pressurized, non-humidified room air (21% oxygen, balanced nitrogen) at 12 L/min. This system included one-way valves to prevent partial gas mixtures and respiratory bellows (Invivo Corporation, Gainesville, FL) to display the breathing pattern and frequency. At 50 s into the data acquisition, the room air gas mixture was switched to 100% nitrogen until the patient had completed five breaths (approximately 25 s, counted through the respiratory bellows data display), then the circuit was changed back to room air. This previously developed protocol²⁶ caused a safe and short desaturation; typical desaturations had a minimum S_pO₂ of 75%−80%. The pulse oximetry measurements during transient hypoxia (Figure 2A) demonstrated similar depth and duration to spontaneous desaturations (Figure 2B). More details on this protocol are provided in Supporting Information Methods section, which is available online.

Experimental setup for transient hypoxia gas paradigm and concurrent peripheral saturation S_pO₂ and MRI BOLD acquisitions. During the experiment, the subjects breathed through the mouthpiece through a 2-L reservoir rebreathing circuit that included one-way valves to prevent partial gas mixtures. The subject also wore a respiratory bellows to display the breathing pattern and frequency

Transient hypoxia model. A, Representative recording of S_pO₂ signal during 100% nitrogen paradigm. B, S_pO₂ signal from the same patient prior to gas paradigm while patient was sleeping during anatomic scanning. C, Representative time series of global BOLD-MR signal and S_pO₂ signals during hypoxia paradigm. D, Representative $Δ R_{2}^{*}$ time curve and its time integral (area under curve)

2.4 |. MRI acquisition

Each participant underwent an MRI study using a 3T Philips Achieva using an eight-element phased-array coil. Anatomical 3D T1, BOLD, PC, and ASL scans were performed in all subjects prior to transient hypoxia.

For each subject, a 3D T1-weighted image was acquired with repetition time (TR) = 8.20 ms, TE = 3.8 ms, flip angle = 8°, resolution = 1 × 1 × 1 mm, and field of view (FOV) = 256 × 224 × 160 mm. BOLD-MR images were acquired with the following parameters: TR = 2000 ms, TE = 50 ms, flip angle = 90°, resolution = 2.5 × 2.5 × 5 mm, and FOV = 220 × 220 × 130 m. A total of 150 volumes were collected. PC images were obtained, positioned just above the carotid bifurcation with the following parameters: TR = 12.9 ms, TE = 7.7 ms, flip angle = 10°, resolution = 1.3 × 1.3 × 5 mm, FOV = 240 × 261 × 5 mm and velocity encoding gradient of 150 cm/s. Pseudo continuous ASL scans were performed using an unbalanced Hanning shaped RF pulse labeling train (mean gradient 1 G/cm, interpulse interval of 1 ms, pulse duration 0.5 ms) and a 3D GRASE two-shot readout scheme with the following parameters: TR = 3800 ms, TE = 9.8 ms, resolution = 3.7 × 3.7 × 10 mm, labeling duration = 2000 ms, post-labeling delay = 1600 ms, dynamics = 10, and EPI factor = 5. Two inversion pulses for background suppression were used, with an inversion efficiency of 95% each. All MR acquisition parameters have been detailed in prior publications.^13,27

2.5 |. Image pre-processing

The PC images were analyzed using an in-house MATLAB program (Mathworks, Natick, MA). Stationary tissue pixels were identified in the complex difference images by simple thresholding using a mean plus 2 SDs from a remote, nonvascular region. Phase differences of stationary pixels were fit to a second-order two-dimensional polynomial and used to remove the background phase from vessels. Vessel boundaries were identified semi-automatically from the complex difference image processed using a Canny edge-detector. A single-voxel dilation of vessel boundaries was then performed, and only voxels whose complex difference was greater than the stationary tissue threshold was retained. Finally, CBF (mL/100 g/min) was computed with $CBF = i T O T O \sum A_{i} \times V_{i}$ , where A_i is the area and V_i is the velocity defined by PC-MRI of each voxel. Total CBF, which was the sum of the flow in the left and right internal carotid arteries and vertebral arteries, was corrected for total gray and white matter volume assuming a brain density of 1.05 g/mL.

The ASL images were rigidly registered to T1-weighted images and then to Montreal Neurological Institute (MNI) template using FMRIB Software Library (FSL).²⁸ Regional CBF quantification was performed using a two-compartment kinetic model¹³ and estimates of labeling efficiency, disease-specific blood T1, and tissue-specific arterial transit times. A comprehensive discussion of these parameters has been detailed in our previous publications.^13,29

The BOLD images were preprocessed with FSL using a standard spatial functional pipeline. Images were first slice-time corrected, realigned to remove physiological motion and then co-registered to the MNI template space. Finally, the registered volumes were normalized and smoothed using 8 × 8 × 8 mm Gaussian kernel similar to typical pre-processing for BOLD fMRI data sets.^30,31 More details of this preprocessing pipeline are described in previous publication.²⁷ Percentage change of the MR signal ΔBOLD is calculated as $Δ B O L D = \frac{B O L D_{h y p o x i a} - B O L D_{b a s e l i n e}}{B O L D_{b a s e l i n e}} \times 100 %$ (Figure 2C), where BOLD_hypoxia was the nadir of the hypoxic time curve and BOLD_baseline was the average of volumes [5, 25] out of 150 temporal volumes.

To minimize the partial volume effects, the AIF was rescaled by the time integral of the VOF measured at the superior sagittal sinus^22,32 (refer to the Methods Section 2.1), because the superior sagittal sinus is a large, long, and straight lumen parallel to the main magnetic field and less vulnerable to partial volume errors. The AIF and VOF were extracted from BOLD images that had been motion-corrected but neither smoothed nor registered to the MRI template. Localizations of the MCA and the superior sagittal sinus were done using a previously published technique on a simple difference data set of BOLD_difference =BOLD_baseline − BOLD_hypoxia (Figure 3).³³ The transient hypoxia model was approximated with a linear function during the rise time of the desaturation bolus and an exponential decay function during the signal recovery; the pre-hypoxia and post-hypoxia baselines were modeled as constants (Figure 2D). To avoid the effects of recirculation and hyperemia as confounders, the end-point of the exponential decay was determined by minimizing the least-square error, and the area under the curve was only computed under the linear and exponential portions using the Riemann sum integration approach.

From these data sets, perfusion maps for CBF, CBV, and MTT maps were computed using an in-house MATLAB program (Mathworks, Natick, MA). Gray matter and white matter perfusion values were computed as an average within tissue-specific masks derived from an average of 152 T1-weighted MRI scans in the common MNI coordinate system.³⁴ White matter masks were eroded by two voxels to minimize partial volume effects. Since cerebral perfusion measurements have a bimodal distribution, values were reported for gray matter and white matter separately. Masks of the anterior cerebral artery (ACA), MCA, and posterior cerebral artery (PCA) circulations in the grey matter were based on the published templates of vascular territories in both hemispheres.³⁵ The territories were created from anatomic studies of cerebral vascularization and evaluated on the bicommissural plane.

2.6 |. Statistical analysis

Statistical analysis on global desaturation values was performed in JMP (SAS, Cary, NC). Student’s t-test was used to examine the difference in clinical variables between the two study groups. Paired t-tests were used to compare CBF computed using dDSC, ASL, and PC. Correlation tests and Bland-Altman analyses were performed to assess global and regional agreement between different flow measures. Adherence to the a priori transient hypoxia model of linear rise and exponential decay was calculated with the root-mean-square percentage error (RMSPE) as $RMSPE = \sqrt{{[\sum_{i = 1}^{n} \frac{y_{i} - {\hat{y}}_{i}}{y_{i}}]}^{2} / n} \times 100 %$ , where n is the total number of sample points, y_i is the actual signal and ${\hat{y}}_{i}$ is the predicted signal.

3 |. RESULTS

All 66 subjects completed T1, BOLD, PC, and ASL imaging. When this study first started, we were initially unable to perform ASL compatible with the White Paper¹⁶ until October 2014, thus comparative ASL images were available in only 45 out of 66 subjects. Additionally, four additional ASL data sets were excluded due to motion, leaving 16 control and 25 anemic subjects in the ASL analysis (Table 1). There were no differences in the demographic and hematologic data between the complete cohort and the cohort after subject exclusion in the ASL analysis.

3.1 |. Quality of AIF and VOF vessels

Localizations of the AIF from the MCA and the VOF from the superior sagittal sinus are shown in a representative subject in Figure 3. Time integral of the normalized venous signal was 140% stronger in magnitude compared with the raw AIF signal in all subjects. Scaling of the AIF by the VOF yielded an average CBF of 42.4 ± 18.6 mL/100 g/min and an average CBV of 4.8 ± 1.3 mL/100 g in healthy controls within the acceptable range in the literature.

3.2 |. Perfusion measured by deoxygenation-based DSC

None of the patients consciously perceived the hypoxia episode and no complications were encountered. This gas paradigm produced an average drop of 16.5 ± 9.0% in S_pO₂ and of 4.3 ± 1.7% and 7.9 ± 2.8% in BOLD signal in the white matter and gray matter, respectively, compared with baseline. Table 2 summarizes average perfusion values in the gray matter and white matter separately for two patient groups, with anemic patients showing significantly higher CBF and CBV compared with healthy controls. Moreover, anemic patients also had slightly faster transit time compared with healthy patients.

TABLE 2.

Gray matter (GM), white matter (WM) and GM-WM ratio group average perfusion parameters

	Control	Anemia	P-value
CBF_dDSC GM (mL/100 g/min)	41.3 ± 17.5	70.8 ± 24.9	<.01
CBF_dDSC WM (mL/100 g/min)	28.1 ± 10.9	45.5 ± 17.9	<.01
CBF_dDSC GM/WM Ratio	1.5 ± 0.2	1.6 ± 0.1	.01
CBV_dDSC GM (mL/100 g)	4.6 ± 1.5	6.2 ± 1.4	<.01
CBV_dDSC WM (mL/100 g)	3.5 ± 1.5	4.7 ± 1.4	<.01
CBV_dDSC GM/WM ratio	1.4 ± 0.2	1.4 ± 0.2	.74
MTT_dDSC GM (seconds)	8.5 ± 1.9	7.5 ± 1.9	.04
MTT_dDSC WM (seconds)	8.4 ± 2.0	7.3 ± 1.7	.02
MTT_dDSC GM/WM ratio	1.0 ± 0.1	1.0 ± 0.1	.16

Figure 4 shows a trend of agreement between CBF values derived from dDSC, PC, and ASL. There is a systemic underestimation of 20.0 mL/100 g/min by dDSC compared with PC with large limits of agreement, and a smaller bias of 12.0 mL/100 g/min compared with ASL in the Bland-Altman analysis. MTT was also well-correlated with CBF in hypoxia-based and PC methods (r² = .25, P < .01 and r² = .13, P = .03, respectively) but only trended for ASL (r² = .08, P = .08). Higher perfusion values were negatively correlated with age (r² = .11, P = .01) but not with sex (P = .77). Additionally, lower hemoglobin levels were associated with higher CBF and CBV (r² = .29, P < .01, and r² = .17, P < .01, respectively) and with lower MTT (r² = .11, P < .01).

Agreement between CBF_dDSC and alternative flow methods. Correlation (A) and Bland-Altman (B) analyses between CBF_dDSC and PC flow. Correlation (C) and Bland-Altman (D) analyses between CBF_dDSC and ASL blood flow in the gray matter (GM). Correlation (E) and Bland-Altman (F) analyses between CBF_dDSC and ASL blood flow in the white matter (WM)

In addition to global analysis, voxel-wise perfusion maps were also generated for the hypoxia-based perfusion method (Figure 5). RMSPE maps in Figure 5D demonstrated strong adherence to the a priori hypoxia model of linear rise and exponential decay, demonstrating robust desaturation observable in voxels across the brain. In addition to group average perfusion maps, examples of individual perfusion maps were also illustrated for two control and two anemic subjects in Supporting Information Figure S1. Test-retest was performed in two healthy control subjects, with individual values shown in Supporting Information Table S1 and individual maps in Supporting Information Figure S2.

Group average perfusion and fit evaluation maps. A, CBF. B, CBV. MTT (C) and root-mean-square percentage error (RMSPE) (D) maps derived from the dDSC protocol. RMSPE maps showed adherence to the a priori hypoxia model of linear rise and exponential decay

Regional analysis of perfusion within three arterial territories is detailed in Table 3. Even though the middle cerebral arteries’ territory showed a trend of higher blood flow and faster transit time compared with the anterior and posterior cerebral arteries, one-way analysis if variance (ANOVA) tests demonstrated no significant difference between perfusion in the three territories (P = .54 for CBF, P = .07 for CBV, P = .83 for MTT). CBF and CBV maps also showed significantly higher values in the gray matter compared with the white matter (P < .01), with the average gray–white matter ratio of 1.4 for CBV and 1.5 for CBF lower than the gray–white ratio of 1.9 in ASL maps. Similarly, MTT map demonstrated gray–white matter differentiation (P = .01), but the distinction was not as strong and more confined to the deep white matter tissue.

TABLE 3.

Group average dDSC perfusion parameters in different flow territories

	Control			Anemia
	CBF (mL/100 g/min)	CBV (mL/100 g)	MTT (s)	CBF (mL/100 g/min)	CBV (mL/100 g)	MTT (s)
ACA	46.5 ± 19.9	4.6 ± 1.3	8.6 ± 2.4	75.3 ± 24.6	6.3 ± 1.5	6.8 ± 1.9
MCA	47.4 ± 20.0	5.4 ± 1.5	8.0 ± 2.2	82.6 ± 29.0	7.0 ± 1.6	6.9 ± 2.2
PCA	42.9 ± 20.1	4.8 ± 1.7	8.4 ± 2.7	77.1 ± 26.2	6.6 ± 1.6	7.0 ± 2.3

Furthermore, regional correlation analysis between CBF_dDSC and CBF_ASL (Figure 4C,E) demonstrated a good association within the gray matter but not in the white matter. Even though imprecise registration prevented voxel-wise comparison between the two approaches, the ACA, MCA, and PCA vascular territories within the gray matter all demonstrated reasonable agreement between CBF estimates by dDSC and ASL techniques (Figure 6A,C,E). There is no significant bias in perfusion measurements by dDSC compared with ASL in different vascular territories of the brain in the Bland-Altman analyses (Figure 6B,D,F).

Regional agreement between gray matter dDSC and ASL flow methods. Correlation and Bland-Altman analyses between gray matter CBF_dDSC and ASL blood flow in the ACA (A,B), MCA (C,D), and PCA (E,F) territories

4 |. DISCUSSION

Our hypoxia protocol induced a reproducible drop in MR signal in both control and anemic patients similar to the mechanism of gadolinium-based DSC. However, unlike DSC, the signal loss in dDSC was smaller in magnitude and induced by increased paramagnetic deoxygenated hemoglobin, leading to shorter T2* and lower gradient-echo signal. Whereas other gas paradigms, such as 100% oxygen, had been used to investigate baseline perfusion in healthy subjects,¹² this work represented the first study to use transient hypoxia as a contrast method for CBF quantification. It is also the first to show a reasonable trend of agreement between dDSC perfusion values and independent flow techniques, such as ASL and PC. Additionally, the physiologic associations between dDSC with age and hemoglobin were expected and provided an extra layer of reassurance for our data quality.

This novel dDSC technique offers many advantages over conventional gadolinium-based DSC. In eliminating the need for an exogenous contrast medium, this method offers an alternative contrast that is well-tolerated, repeatable, does not lead to anaphylaxis, and is convenient in patients whose venous access is more challenging, such as in pediatric or anemic subjects. This method is significantly cheaper compared with gadolinium as well as iron-based contrast, such as ferumoxytol. It is also potentially easily translatable to clinical protocols, especially for perfusion quantification in patients under anesthesia. Apart from brain imaging, since previous work has demonstrated good performance of BOLD imaging and different respiratory challenges in the liver, kidney, and heart,^36–39 this dDSC technique can potentially be applied to assess in these organs where non-contrast imaging modalities, such as ASL, have proven less effective due to low signal-to-noise ratio.⁴⁰

Most importantly, this hypoxia gas paradigm is very safe and well-tolerated; our laboratory has performed this experiment on over 200 subjects in a larger study on sickle cell anemia without any major adverse events, such as stroke or transient ischemic attack.^26,41 In a review on the safety of hypoxia challenges, Bickler et al showed that hypoxia experiments have been performed at many different research centers across the world with arterial saturations dropping as low as 45%, which is considerably lower than the typical hypoxia levels experienced in our study.⁴² Overall, these brief exposures to hypoxia have been demonstrated to be well-tolerated without any evidence of cardiovascular compromise, systemic acidosis or lasting cognitive impairment.⁴²

Absolute quantification of CBF is difficult due to the large partial volume effects in the AIF related to small vessel sizes and limited spatial resolution.⁴³ While many multiplicative rescaling methods for AIF correction have been explored,⁴⁴ rescaling with a VOF from the superior sagittal sinus has proven robust due to the large vessel caliber of the superior sagittal sinus as well as its position parallel to the magnetic field and orthogonal to the imaging plane.^32,45 However, even with the use of venous rescaling, our CBV measurement was still overestimated compared with normal literature values. Additionally, even though PC has been shown to overestimate total blood flow,⁴⁶ the large bias between CBF measured by dDSC and PC might be due to the low quality of the AIF. AIF that was contaminated with nearby tissue voxels was lower and broader in shape, leading to lower CBF values.⁴⁷ Partial volume effects were still present in the superior sagittal sinus, leading to lower area-under-curve of the VOF and AIF and consequently to overestimation of tissue CBV. Heterogeneous and low quality of the AIF were the potential causes for large limits of agreement of dDSC compared with PC and ASL. An overestimation of CBV and underestimation of CBF led to a systematic overestimation of MTT in this study. Future work to use alternative and more effective methods to extract the AIF and VOF will be necessary to improve the accuracy of CBV and CBF measurements in dDSC.

In general, CBF_dDSC showed reasonable inter-modality agreement with PC and ASL using correlation and Bland-Altman analyses. The stronger correlation between CBF_dDSC with PC compared with ASL was understandable, because PC is independent of T1 and T2 relaxations and has relatively high signal-to-noise ratios. In addition to the agreement in whole-brain blood flow, our dDSC-based CBF measures also showed regional agreement with ASL perfusion maps. Gray matter as well as the three vascular territories all displayed more robust inter-modality association between CBF_dDSC and CBF_ASL compared with deep white matter regions, suggesting a divergence in the sensitivities of these two methods to microvascular perfusion characteristics in the white matter.

Our observation of significantly higher blood flow and volume in anemic patients compared with healthy controls was consistent with previous works showing increased baseline perfusion in compensation for compromised oxygen carrying capacity in chronic anemia.^14,15 The faster MTT in anemic patients was also in accordance with the shorter vascular transit times measured with multi-post-labeling delay ASL.⁴⁸ Shorter MTT in anemic subjects was apparent in the gray matter as well as regions of normally perfused white matter; however, these subjects also displayed abnormally slower contrast dynamics in the deep white matter at the end of the perfusion branches. These borderzone regions coincided with typical watershed areas of flow limitation and peak oxygen extraction^13,49 and were consistent with previous studies that showed reduced flow reserve in areas vulnerable to silent strokes,⁵⁰ suggesting a connection between hemodynamic impairment and the development of strokes in anemic patients.

Even though ASL is the prominent non-invasive perfusion technique and has been the focus of much research and development in the MRI community, ASL is still hampered by some disadvantages compared with dDSC. Most notably ASL has low signal-to-noise ratio in the white matter, possibly contribution to the lack of agreement between white matter CBF measured by the two techniques. Additionally, ASL is confounded by model assumptions and physical parameters, including blood T1 and arterial transit time, many of which are different in anemic patients compared with healthy controls.²⁹ Blood T1 can be measured with T1-mapping,^51,52 but measurement of arterial T1 remains challenging and not universally available. Arterial transit time confounds can also be avoided by using multiple post-labeling delay ASL,¹⁶ but there is currently too much heterogeneity in the acquisition techniques. Additionally, simple multi-delay ASL can take significantly longer than single-delay ASL. Acceleration methods, such as Hadamard encoding or Look-Locker readouts, can be used to shorten acquisition time, but these are typically unavailable on routine clinical scanners. Therefore, even though ASL will undoubtedly have an important place in the future of perfusion imaging, the development of this novel dDSC technique provides an alternative and generalizable method for measuring tissue blood flow and builds upon the previous work that has already been pioneered by the perfusion MRI community.

An important assumption in our dDSC model was the linearity between cerebral oxygen level and the changes in MRI signal $Δ R_{2}^{*}$ during the hypoxia gas paradigm in Equation (1). Calibrated BOLD model typically presents a non-linear relationship between deoxygenated hemoglobin and changes in the BOLD signal with a vessel size-dependent parameter β = 1.3 at 3T.⁵³ However, under hypoxic conditions, in addition to capillaries and venules, larger deoxygenated arterioles also contribute to the BOLD signal, in which case β is likely decreased closer to 1, and the relationship between the BOLD signal and oxygen saturation could be approximated with a linear function. Furthermore, previous work by Rostrup et al has demonstrated an approximately linear relationship between $Δ R_{2}^{*}$ and oxygen saturation during hypoxic exposure, thus confirming that the linearity assumption was reasonable.⁵⁴ Future work to validate this linear assumption in normal tissue as well as diseased tissue is warranted to confirm and improve the dDSC diagnostic capability. In addition to validation of the linear assumption, other confounds to the BOLD signal, including the effects of vessel size, orientation, water diffusivity, and the influence of hematocrit on BOLD signal variations,⁵⁵ need to be explored. The practicality of assessing the impact of these parameters in vivo is limited, but in silico validation using Monte Carlo simulation can explore the relationship between these parameters and the dDSC signal in future work.

This study has some notable limitations. First, several confounds in the DSC model potentially affected the perfusion measurements in our dDSC technique, including delay and dispersion of bolus in the AIF. These confounds typically introduced errors in cases of severe vasculopathy, so they likely did not affect our cohort; however, application of dDSC to subjects with vascular disease will require technical improvement to address these issues, such as by modeling the vascular bed, extracting local AIF, or using a delay-insensitive SVD-variant for perfusion quantification.^56–58 Second, these perfusion measurements showed large variability compared with ASL. However, previous comparison studies between ASL and gadolinium-based DSC have also demonstrated large variability.^59–61 And since this is a proof-of-concept study, our results demonstrated that the novel dDSC technique has the potential to measure blood flow and that these measurements do vary proportionally with other conventional flow measurements, such as PC and ASL.

Further optimization will be necessary to improve the quantitation and control for potential confounders, such as breathing rate, minute ventilation, and intra-subject variability in hypoxic exposure. Gaussian smoothing for noise reduction in the BOLD data sets led to underestimated gray–white matter ratio in both CBF and CBV maps. Future work to optimize the hypoxic bolus depth and duration and use repeated hypoxic stimuli to improve signal-to-noise can require less spatial smoothing and increase gray–white matter perfusion differentiation. Even though repeatability tests showed reasonable performance on two healthy subjects (Supporting Information Figure S2), test–retest on a larger subset of subjects needs to be conducted for this methodology to better validate the diagnostic value of dDSC.

Another important limitation to our study is that a hypoxic stimulus may impact CBF in an unpredictable manner. Prolonged hypoxic exposure increases CBF,⁶² but also causes hyperventilation that tends to counteract the hypoxic vasodilation.^62,63 Whether a challenge this brief raises or lowers CBF during the observation interval has not been determined since the transient duration of the hypoxia paradigm prevented measurement of CBF at steady-state hypoxia. A more in-depth and thorough investigation into stimulus-mediated changes in CBF is required to assess their contribution to our measurement uncertainty. Dual-acquisition of ASL and BOLD can be used to capture the dynamic BOLD signal as well as regional CBF changes across the brain during the hypoxia challenge; 4D-flow MRI can also be used to assess vasodilation in of the major arteries in the Circle of Willis in response to the hypoxic stimulus. Additionally, this experiment was originally designed with end-tidal O₂ and CO₂ acquisitions (Biopac Systems, Goleta, CA). However, during the course of the study, the gas sampling sensors broke, and we were not able to retrieve the end-tidal measurements. For future gas experiments, the use of a controlled respirator with prospective targeting of pO₂ and end-tidal CO₂ will be helpful to eliminate hypocapnia as a potential confounder.⁶⁴

In conclusion, in this proof-of-concept study, we have demonstrated the feasibility of using transient hypoxia to generate a contrast bolus that mimics the effect of gadolinium and yields reasonable blood flow measurements. Even though our current dDSC implementation suffers from biases in perfusion estimates and requires further validation of its assumptions, dDSC still offers a novel gadolinium-free approach, which will be attractive in pediatric and end-stage renal disease patients. Additionally, dDSC can be potentially beneficial for patients in whom serial DSC studies are required (such as in brain tumors), but this application awaits further investigation into the feasibility of dDSC in measuring impaired blood-brain barrier function. To improve the diagnostic utility of this method, block-design gas paradigms⁶⁵ and transfer function analysis⁶⁶ can be applied, and hypoxic stimuli can also be used in an interleaved manner with CO₂ inhalation to jointly estimate resting perfusion and cerebrovascular flow reserve. Since the AIF can easily be calculated in the heart and aorta, future studies can also extend this cerebral perfusion technique to blood flow measurements in other thoracic and abdominal organs.

ACKNOWLEDGMENTS

The authors acknowledge Mr Bertin Valdez for his efforts coordinating the patient study visits and Dr Thomas Coates, Dr Tom Hofstra, Dr Jackie Bascom, Susan Carson, Trish Peterson, and Debbie Harris from the CHLA Hematology Division for their assistance with patient recruitment.

Funding information

This work was supported by National Heart, Lung, and Blood Institute (grant 1U01-HL-117718-01, 1R01-HL136484-01A1 and a Minority Supplement to grant 1U01-HL-117718-01), the National Center for Research (5UL1-TR000130-05) through the Clinical Translational Science Institute at Children’s Hospital Los Angeles, the National Institutes of Health (grant R01-NS074980), the National Institutes of Health Predoctoral Training in Interdisciplinary Neurosciences (1-T32-MH-111360-1-A1) and National Institute of Health grant (R01-ES024936). Philips Healthcare provided support for protocol development and applications engineering on a support-in-kind basis

CONFLICT OF INTEREST

Chau Vu, Yaqiong Chai, Julie Coloigner, Aart Nederveen, Matthew Borzage, Adam Bush: none. John C. Wood: Research Funding NHLBI and NIDDK of the National Institutes of Health, Research Support-in-Kind from Philips Healthcare, Consultant for BluebirdBio, Celgene, Apopharma, WorldcareClinical, Imago Biosciences, Novartis, and BiomedInformatics.

Quantitative perfusion mapping with induced transient hypoxia using BOLD MRI - PMC