|Home | About | Journals | Submit | Contact Us | Français|
While oxygen extraction fraction (OEF) and cerebral metabolic rate of oxygen (CMRO2) are fundamental parameters of brain health and function, a robust MRI-based mapping of OEF and CMRO2 amenable to functional MRI (fMRI) has not been established. To address this issue, a novel method called QUantitative Imaging of eXtraction of Oxygen and TIssue Consumption, or QUIXOTIC, is introduced. The key innovation in QUIXOTIC is the use of velocity-selective spin labeling to isolate MR signal exclusively from postcapillary venular blood on a voxel-by-voxel basis. Measuring the T2 of this venular-targeted blood allows calibration to venular oxygen saturation (Yv) via theoretical and experimental T2 versus blood oxygen saturation relationships. Yv is converted to OEF, and baseline CMRO2 is subsequently estimated from OEF and additional cerebral blood flow and hematocrit measurements. Theory behind the QUIXOTIC technique is presented, and implications of cutoff velocity (VCUTOFF) and outflow time parameters are discussed. Cortical gray matter values obtained with QUIXOTIC in 10 healthy volunteers are Yv = 0.73 ± 0.02, OEF = 0.26 ± 0.02, and CMRO2 = 125 ± 15 μmol/100 g min. Results are compared to global measures obtained with the T2 relaxation under spin tagging (TRUST) technique. The preliminary data presented suggest that QUIXOTIC will be useful for mapping Yv, OEF, and CMRO2, in both clinical and functional MRI settings.
The rate of oxygen consumption in brain cells is known as the cerebral metabolic rate of oxygen (CMRO2) and is an important indicator of brain health and function (1). As oxygen is not produced endogenously in the brain, it must be continuously supplied to meet metabolic demands. Cerebral blood flow (CBF) fulfills this requirement by delivering oxygen via hemoglobin, the principal oxygen carrier in the body. The ratio of oxygen consumption to oxygen delivery is called the oxygen extraction fraction (OEF). OEF can be expressed in terms of CMRO2 and CBF (2,3) in the following manner:
where [Hbtotal] is the total concentration of hemoglobin in blood and Ya is the arterial oxygen saturation.
One of the main difficulties in measuring OEF and CMRO2 with in vivo 1H MRI is the lack of a direct contrast mechanism. However, as hemoglobin modulates MR signal relaxivity based on degree of oxygen saturation (Y), the possibility to indirectly assess OEF and CMRO2 arises. Venous oxygen saturation (Yv) is extrapolated from the MRI signal and used first to compute OEF (3):
where arteriolar blood is assumed fully saturated (Ya = 1) or measured independently by pulse oximetry. Following Eq. 1, CMRO2 is then calculated from OEF and additional CBF and hematocrit measurements. As oxygen is typically extracted as blood traverses the capillary network, Yv measurements focused at the distal end of the network (i.e., in postcapillary venules) will yield spatially specific OEF and CMRO2 estimates.
One class of methods to measure Yv uses phase contrast imaging to examine magnetic susceptibility differences between veins and surrounding tissue (4-7) and uses simple, first-principle models to relate phase changes to Yv. These methods have the advantage of straightforward acquisitions via standard MRI sequences, but require manually visualizing vessels, limiting use in measuring regional or voxelwise Yv. A second class of methods (8,9) invokes an MRI signal model that quantifies Yv based on dHb-induced signal loss in extravascular tissue. One such approach developed by An et al. uses a single compartment tissue model to produce OEF (8) and CMRO2 (10) maps. He and Yablonskiy (9) have introduced “quantitative BOLD” which considers a multiple-compartment model and accounts for tissue, blood, and cerebrospinal fluid (CSF) signal and similarly produces OEF maps. This latter approach is especially promising and has been recently used in assessing change in OEF during functional stimulation (11).
A third class of methods to measure Yv implements an intravascular T2-based approach and has shown promise in functional MRI (fMRI) settings. This approach uses a theoretically and empirically derived relationship between blood transverse (T2) relaxation time and blood oxygen saturation (Y) (2). Specifically measuring Yv requires selective targeting of blood in the venous circulation, without partial voluming with tissue, CSF, or other brain constituents. Isolating this pure venous blood signal, however, has been a major challenge for existing techniques.
As a consequence, intravascular methods require strict selection criteria to identify voxels containing exclusively venous blood. Methods presented by Oja et al. (3) and Golay et al. (12), for example, require functional activation experiments to identify candidate voxels, and of these, only a subset can be used to calculate Yv. Moreover, because of resolution limitations, usable voxels are typically found only in larger venous vessels. As blood oxygenation in these larger vessels represents oxygen exchange of all supplying capillaries in a large surrounding region, these techniques have fundamental limitations on spatial specificity. Voxelwise Yv is unfortunately not feasible.
More recently, Lu and Ge have proposed T2 relaxation under spin tagging (TRUST) MRI and measure Yv in the sagittal sinus (13). TRUST-MRI delivers a pure blood signal, free from partial volume effects, by using a spin-labeling approach to eliminate static tissue and CSF via control–tag subtraction. While TRUST is an important advance for T2-based, Yv methodology, only global estimates are possible, as blood signal is isolated from the largest cerebral veins (i.e., the sagittal sinus). As such, TRUST Yv measurements are limited to the terminal draining veins; regional Yv information is consequently lost.
Our work uses the intravascular T2-based approach but addresses the largest obstacle for localized measurements: the isolation of postcapillary venular (PCV) blood signal. A novel excitation scheme is proposed that uses velocity-selective spin labeling (VSSL) to isolate PCV blood signal. In principle, the methodology uses velocity-sensitive pulses to exploit heterogeneous blood velocities in the vascular tree and subsequently create a flow-dependent venular blood component that persists after control-tag subtraction. This sequence incorporates T2 preparation-based technology, acquires PCV blood-weighted images at effective echo times, and exponentially fits these data to estimate PCV blood T2. T2,blood versus Y curves (with hematocrit as an independent parameter) are then generated from theoretical models and existing 3 T data and are used to calibrate PCV blood T2 to Yv. OEF is then estimated from Yv and combined with CBF (measured with arterial spin labeling MRI) and hematocrit to quantify CMRO2. This new approach is dubbed QUantitative Imaging of eXtraction of Oxygen and TIssue Consumption or “QUIXOTIC.”
In this study, we perform QUIXOTIC and TRUST-MRI on 10 healthy subjects and compare cortical gray matter (GM) values of T2, Yv, OEF, and CMRO2 estimated by QUIXOTIC, with global values estimated by TRUST. We further present representative Yv, OEF, and CMRO2 quantitative maps and proceed to explore the dependence of Yv measurements on key QUIXOTIC parameters.
The key innovation behind the QUIXOTIC approach is venular-targeted VSSL (VT-VSSL). VT-VSSL allows targeting of MR signal exclusively from the PCV blood compartment, from which T2 and Yv can be measured. VT-VSSL applies velocity-sensitive MR pulses to exploit differential velocities and accelerations of blood in the vascular tree, and relies on uniformly forward flow from arteries, through capillaries, to veins. The two-step data acquisition paradigm creates control and tag images with and without the desired venular blood component, respectively. Subsequent pairwise control–tag subtraction eliminates signal from static tissue, CSF, and nonvenular blood compartments, leaving an image exclusively containing blood in the postcapillary venules. The following section describes the details and timing behind this approach.
QUIXOTIC is adapted from velocity-selective (VS) arterial spin labeling (14) and uses similar VS modules to dephase blood spins above a given cutoff velocity. The pulse sequence is summarized schematically in Fig. 1 and executed for both tag and control acquisitions.
The principal component of the QUIXOTIC excitation scheme is the VS module. As a time derivative of position, velocity can be phase encoded by applying bipolar gradient pulses, or equivalently, unipolar pulses separated by a 180° refocusing pulse (15). Inserting this phase-encoding scheme between two excitation pulses with opposite phase (90° − 90°̅) results in the following radiofrequency (RF) pulse train: 90° − G − 180° − G − 90°̅. When applied to laminar flow systems found in noncapillary vasculature of the human circulatory system, the result is a filtering of the magnetization from moving blood spins flowing below a certain cutoff velocity (VCUTOFF) and elimination of magnetization from spins flowing above VCUTOFF, according to:
where M0 is the initial magnetization, v is spin velocity. β is related to the separation (Δ), duration (δ), and amplitude (G) of the bipolar gradient pulses in the following manner:
Figure 2 displays the sinc-filter profile and the ideal velocity filter for comparison.
Instead of using a 90° − G − 180° − G − 90°̅ pulse train, which can be implemented with simple hard RF pulses, a four RF and four gradient approach was used (as suggested in Ref. 14):
which incorporates adiabatic pulses as the refocusing 180°s. This train offers greater resistance to transmitted field (B1+) and static field (B0) inhomogeneity, water diffusion, and eddy current effects (14,17). Prior investigations using this eight-pulse sequence (14) verify its superior performance with reduced artifacts compared to data acquired with the simpler 90° − G − 180° − G − 90°̅ train (16). The final VS train is shown in Fig. 3; for perfect spin-echo refocusing, the adiabatic refocusing pulses are symmetrically spaced.
The difference between the QUIXOTIC tag and control acquisitions lies in the second VS module. While the tag acquisition applies a user-defined cutoff velocity (VCUTOFF) for both VS module I (VS1) and VS module II (VS2), the control uses VCUTOFF for VS1, but disables velocity selection for VS2 (i.e., flowing spins are unaffected, VCUTOFF = infinity). Incorporated into VS2 is a T2-preparation module (18,19), allowing T2 estimation via data acquisition at multiple echo times. An important feature of the sequence is an inversion pulse at TO1, which compensates for adverse effects of T1 relaxation that could potentially introduce unwanted signal from arterial blood.
To introduce properties of this sequence, we first neglect T1 relaxation and the TO1 inversion pulse. Figure 4 provides an idealized cartoon depicting blood spins in the circulation for both control and tag acquisitions, their history during the pulse sequence timing, and the final contribution in the subtracted image. At t = 0, before VS1, all blood (arterial, venous, capillary) is relaxed (Fig. 4a). Strong velocity weighting (low VCUTOFF) is then applied during VS1 for both tag and control, selecting for slow moving spins in small arterioles, capillaries, and small venules (V < VCUTOFF) but dephasing faster moving spins in larger vessels (V > VCUTOFF) (resulting in a loss of signal, Fig. 4b). Notably, this large-vessel signal is eliminated on both sides of the circulation. After VS1, the outflow time (TO) allows the targeted blood to flow out of the small vessel compartments and accelerate into larger venular vasculature (Fig. 4c). VS2 is then applied at TO. This time, however, the tag and control acquisition experience different velocity weighting: the tag sees velocity selection at VCUTOFF, but the control experiences no velocity weighting. Spins that have accelerated above VCUTOFF during TO are dephased by the tag acquisition but left unaltered in the control (Fig. 4d). As imaging starts immediately after VS2, subtraction of tag from control yields an image weighted to blood that has accelerated from below VCUTOFF to above VCUTOFF, during TO (Fig. 4e). Assuming flow always progresses from arteries to capillaries to veins, these spins are venous only. Signal from other sources (static, CSF, nonvenular blood) is eliminated via subtraction. If VCUTOFF and TO are chosen properly, signal from PCV blood is exclusively targeted.
Due to the tortuous nature of human vasculature, the spatial direction of flow naturally varies as blood migrates from arteries to capillaries to veins. This phenomenon, however, is not expected to appreciably affect the venular selectivity of the VT-VSSL experiment for the following reasons: first, as phase encoding does not occur during TO (the longest interval of the experiment), the directional changes during this interval have no impact on velocity selection. Conversely, directional changes during the VS module will affect selectivity. However, because the encoding time is so short (approximately 20 ms), directional changes are both minor and expected to average out given the broad distribution of vessel orientation in a parenchymal voxel. In this way, as long as there exists some flowing blood with a velocity component aligned with the encoding direction greater than VCUTOFF, VT-VSSL can occur. A more detailed discussion of the directional sensitivity of the VS module is discussed by Frank et al. (20).
T1 relaxation complicates the idealized model depicted in Fig. 4. Spins dephased by VS1 at t = 0 experience longitudinal magnetization recovery. Because velocity selection via VS2 occurs only for the tag but not control, spins from unwanted compartments will partially recover in the control, but dephase in the tag at TO+. Without compensation, these unwanted spins will not subtract completely, and QUIXOTIC loses venous selectivity. To address this effect, we place an inversion pulse at time TO1 to null recovering blood at TO; consequently, signal from spins in this unwanted population is eliminated in both control and tag at TO, leaving only desired PCV blood on subtraction.
To generate the PCV blood-weighted maps at different echo times, T2 preparation is integrated into the VS2 pulse train. For the shortest effective echo time (TEeff), the VS2 module is exactly as pictured in Fig. 2. For each successive TEeff, an additional pair of adiabatic 180° pulses is added between the flip-up 90° pulse and first VS gradient lobe. In this way, the duration of VS2 will increase by ΔTEeff for each successive TEeff. Importantly, this is done while keeping an identical readout onset time regardless of TEeff. Using T2 preparation and fixing the readout module timing makes the T2 encoding flow insensitive; in other words, images at successive TEeff should differ only by T2 weighting, not by volume of blood delivered during TO (18).
Signal intensity (SI) in the acquired blood-weighted images can be plotted versus TEeff and fit exponentially on a voxel-by-voxel basis to generate T2 maps of PCV blood. T2 values are then calibrated to Yv and Eqs. 1 and 2 can be used compute CMRO2 and OEF. The following sections explain these steps in greater detail.
Proper estimates of T2 from fitting SI versus TEeff data depend on accurate estimates of TEeff at each successive acquisition. Ideally, TEeff is equal to the duration of the T2-preparation module, assuming pure T2 relaxation and very short refocusing pulses. In practice, however, magnetization does not experience pure T2 decay during the hyperbolic secant refocusing pulses (18), and further, despite a long repetition time (TR; 4 s), a small amount of residual T1 contamination is present from spins that have not completely relaxed. To reduce the impact of such biases to the T2 estimate, we derived a small adjustment to the nominal echo spacing and applied the resulting effective echo spacing, ΔTEeff, for T2 estimation from the measured data. Specifically, we simulated the pulse sequence using the Bloch equation, assumed T1 and T2 decay during the hyperbolic secant refocusing pulses, and estimated the ΔTEeff appropriate for our pulse sequence. The T1 and T2 used for the simulation were 1660 and 70 ms, respectively, which are approximate T1 and T2 values for deoxygenated blood at 3 T (13,21).
The relationship between transverse relaxation time of blood and blood oxygen saturation (Y) was first examined in the early 1980s, by Thulborn et al. (22). In years following, a parametric model based on the Luz–Meiboom model of spin relaxation during two-compartment exchange (in this case, between plasma and the red blood cell) was derived to fit experimental data measuring variations of T2,blood and blood oxygen saturation. Such work enabled MRI-based measurements of oxygen saturation in vivo (23). In 1998, Van Zijl et al. proposed a fundamental theory for T2-based signal changes in perfused tissue and presented a quantitative model linking T2 relaxation of blood to physiologic parameters including Y (2,12). This expression can be written in terms of T2, blood, Y, and hematocrit (Hct) in the following form:
where A, B, C, D, and E represent lumped biophysical parameters related to susceptibility differences between the red blood cell and plasma, water relaxation rates in the red blood cell and plasma, and the spacing between 180° RF pulses in a CPMG spin-echo module (τCPMG).
Fit of available Y versus T2,blood data at 3 T obtained from in vitro bovine blood (3 T, Carr Purcell Meiboom Gill (CPMG) multiecho T2 sequence, τCPMG = 10 ms, courtesy of P. van Zijl and C. Clingman, Johns Hopkins University), yielded values for A, B, C, D, E, that we used in our work. Using these values, Y versus T2 curves (with Hct as an independent parameter) were generated. Bovine blood has the same properties as human blood in terms of hemoglobin content, erythrocyte size, and diffusional permeability and is expected to have similar MR relaxation properties (12). It has been successfully used in the past for similar calibration experiments (12,13,24).
QUIXOTIC was used to evaluate Yv, OEF, and CMRO2 in healthy volunteers. The protocol was approved by the University’s Committee of Using Humans as Experimental Subjects. Ten young (22–32 years) nonsmoking, healthy subjects (six females, four male) were scanned at 3 T (Siemens Tim Trio, Erlangen, Germany) with the following MRI protocol:
Automatic alignment routines (26) were used to ensure similar slice placement among all subjects. A slightly oblique-axial slice-of-interest was prescribed immediately superior to the corpus callosum for acquisitions two to four and contained a significant fraction of GM. A different oblique-axial slice intersecting the sagittal sinus was positioned for acquisition five (TRUST).
Venular blood-weighted imaging comprised the bulk of the scan session and the imaging time for the full protocol was less than an hour. Data were acquired at six TEeffs if possible, but in four subjects, specific absorption rate constraints imposed a limitation of only five TEeffs. To test the reproducibility, we performed the VT-VSSL experiment twice on two of the volunteers. These volunteers remained in the scanner for an additional 30 min for the second VT-VSSL acquisition. The two VT-VSSL trials were spaced approximately half an hour apart and used identical scan parameters.
After MRI scanning, hematocrit was measured via finger prick blood sample using the Ultracrit device (Separation Technologies, Altamonte Springs, FL). O2 saturation was measured with a pulse oximeter (8600FO Pulse Oximeter, Magmedix, Fitchburg, MA).
The data from each TEeff acquisition were corrected for bulk motion and then subtracted, control minus tag, in a pairwise fashion. The subtraction series was averaged to produce mean PCV-weighted images. The DIR image was used as a GM mask, within which venular blood SI from cortical GM tissue was measured. PCV blood signal intensities from the entire cortical GM region were plotted versus TEeff to provide whole-slice cortical GM measurements. The plots were then exponentially fit using a Levenberg–Marquardt least squares optimization method to measure the T2 relaxation parameter for whole-slice cortical GM, and the standard error of the estimated parameter (SEE) assuming a t-distribution was computed.
The TRUST acquisitions were similarly subtracted in a pairwise fraction and averaged to produce maps containing sagittal-sinus only blood at the five TEeffs. The mean SI from the six brightest sagittal sinus voxels blood was plotted versus TEeff and fit to estimate sagittal-sinus blood T2. The SEE assuming a t-distribution was computed.
T2 values from both VT-VSSL and TRUST analyses were then calibrated to Yv using curves generated with Eq. 7, incorporating the hematocrit measured from the volunteer. As VT-VSSL data derive mostly from blood in small vessels, the T2 versus Y calibration was calculated using the microvascular hematocrit (i.e., by multiplying the measured hematocrit by 0.85 to correct for hematocrit differences between small and large vessels; Ref. 27). This correction was not needed for the TRUST calibration curves, as the source of venous blood was exclusively from the large sagittal sinus. Ya was taken as the oxygen saturation measured with the pulse oximeter; Eq. 2 was then used to calculate OEF.
To obtain CMRO2 from Eq. 1, CBF was estimated by ASL MRI, with a VT-VSSL matching slice prescription. The ASL data were similarly analyzed to generate blood flow-weighted images, via subtraction, motion correction, and signal averaging. These images were calibrated to absolute CBF maps by using the local tissue proton density provided by the M0 scan (28). GM CBF was estimated from the region segmented by the DIR GM mask. White matter (WM) CBF was estimated from the remaining, nonsegmented brain region. The WM and GM CBF values were used to calculate whole brain CBF, by assuming a whole-brain WM:GM ratio of 0.675:1 (29). With these additional CBF measurements, and by using hematocrit to estimate [Hbtotal], both QUIXOTIC (GM) and TRUST (whole brain) CMRO2 were, respectively, calculated.
To demonstrate feasibility of using QUIXOTIC to create quantitative Yv, OEF, and CMRO2 maps, raw data from a representative subject were smoothed with a 10 mm full width at half maximum (FWHM) Gaussian kernel and fit for T2, but this time on a voxel-by-voxel basis (as opposed to using a GM mask). The resultant T2 map was subsequently calibrated to generate quantitative Yv and OEF maps, using aforementioned calibration curves. Finally a CMRO2 map was generated by multiplying the OEF map with the absolute ASL-CBF map on a voxel-by-voxel basis and incorporating [Hbtotal] and Ya (as described in Eq. 1).
Finally, to briefly investigate effects of the two key QUIXOTIC parameters, TO and VCUTOFF, on the measured Yv, three of the original 10 subjects returned for a separate, additional imaging session and were scanned at five different TOs (525, 625, 725, 825, and 925 ms) with a fixed VCUTOFF of 2.0 cm/s, and four different VCUTOFFs (1.5, 2.0, 3.0, and 4.0 cm/s) with a fixed TO of 725 ms. As using the standard QUIXOTIC approach to explore this parameter space would have resulted in impractical scan times (approximately 5 h), a faster variant (a so-called “turbo QUIXOTIC” approach) was developed and incorporated. Turbo QUIXOTIC uses a turbo spin echo EPI readout to generate images at multiple echo times per TR; this contrasts the standard T2 preparation approach, which acquires a single TEeff image per TR. Turbo QUIXOTIC thus obtains the full set of images for a particular TO-VCUTOFF combination in a single scan, allowing data for the above parameter exploration to be acquired in less than an hour. Specific imaging parameters were: TEeff = 22.6 ms, bandwidth = 3256 Hz/pixel, matrix size = 32 × 32, single slice, voxel size = 7.8 × 7.8 × 10 mm3, TR = 4 s, generalized autocalibrating partially parallel acquisition with 3× acceleration. Eighty measurements were acquired at four TEeffs, resulting in 40 control/40 tag per TEeff, and an imaging time of 5 min 30 s per TO/VCUTOFF combination. A resolution-matched double IR image was acquired for GM segmentation; subsequent data processing to calculate Yv was performed as described for the standard approach.
Data processing was done in Neurolens (www.neurolens.org) software and with custom Matlab (MathWorks, Natick, MA) routines.
Figure 5 shows the mean difference images for the VT-VSSL and TRUST acquisitions at five effective TEs for a representative subject. The signal in the VT-VSSL images is our estimate of the PCV blood, which decays with a T2 time constant that depends on the blood oxygenation level.
Figure 6a and b show the ASL-CBF map and the GM mask, respectively. Figure 6c displays the slice orientation of the preceding acquisitions, overlaid on a high resolution midsagittal section from the MP-RAGE scan.
Figure 7 shows the whole-slice cortical-GM SI versus TEeff (QUIXOTIC) and sagittal sinus SI versus TEeff (TRUST) on a semilog scale for all subjects, including the fitted lines for T2 estimates. All line fits had high R2 values: ≥0.95 for QUIXOTIC and ≥0.99 for TRUST. To improve the figure readability and highlight the relevant parameter (slope of line fits), the vertical offset of each line in each technique has been normalized. The y-intercept estimates the volume of tagged blood that remains after subtraction within the sagittal sinus and GM voxels for TRUST and QUIXOTIC, respectively, but does not enter the OEF or CMRO2 estimates.
Figure 8 shows calibration curves generated with Eq. 7 for the same subject. Values of A, B, C, D, and E for Eq. 7 were found as 1.09 s−1, 11.26 s−1, −7.96 s−1, 1.08 s−1/2, and 16.54 s−1/2, respectively. Two curves are depicted for both microvascular and macrovascular hematocrit, to be used for QUIXOTIC and TRUST analyses, respectively. Lower hematocrit shifts the T2,blood versus Y curve to the left, resulting in lower estimated oxygen saturations (Y) for a fixed measured T2,blood value.
Figure 9 shows the Yv, OEF, and CMRO2 maps for the same representative subject, created by applying QUIXOTIC theory on a voxel-by-voxel basis. Yv and OEF is seen to be relatively uniform across the entire brain, while the CMRO2 map shows substantially higher values in the GM cortex, compared to the interior WM.
Figure 10a and b show Yv versus TO and Yv versus VCUTOFF, respectively, for the parameter exploration data acquired with turbo QUIXOTIC in the three reimaged subjects. These data show the dependence of Yv on these parameters; error bars on the individual data points represent the SEE (roughly ± 1% for all data points). According to Fig. 9a, Yv is relatively stable across the chosen range of TO values (mean coefficient of variation across TO trials = 1.8 ± 0.4%), with a very slight decreasing trend (mean slope of linear fits, ΔYv/ΔTO = −0.007/100 ms). Fig. 9b, however, suggests substantially more variation in Yv across the VCUTOFF range (mean coefficient of variation across TO trials = 4.2 ± 2.4%); more importantly, there seems a more pronounced downward trend in Yv as VCUTOFF increases (mean slope of linear fits, ΔYv/ΔVCUTOFF = −0.03 s/cm), suggesting dependence of estimated Yv on the choice of VCUTOFF.
Table 1 summarizes the measured blood relaxation times and the estimated quantitative physiological parameters for the 10 subjects, i.e., T2, Yv, OEF, CBF, and CMRO2. These values are shown for QUIXOTIC (regional cortical GM and whole-slice cortical GM) and TRUST (whole brain). The QUIXOTIC estimates were Yv = 0.73 ± 0.02, OEF = 0.26 ± 0.02 and CMRO2 = 125 ± 15 μmol of O2/(100 g min), while TRUST yielded Yv = 0.63 ± 0.02, OEF = 0.36 ± 0.02, and CMRO2 = 125 ± 18 μmol of O2/(100 g min).
Table 2 shows the intrascan test–retest data for subjects 1 and 2, including the percent difference in the quantified parameters. The differences were low in both cases, less than 1% and 5% for subjects 1 and 2, respectively.
We have introduced a novel MR technique, QUIXOTIC, to quantify regional OEF and regional CMRO2 and have demonstrated its use in normal volunteers. Our estimates of cortical GM OEF and CMRO2 fall within the expected physiological range and are comparable with those reported by other MRI and positron emission tomography methods (10,30-33). Some studies that measure OEF specifically report values of 35%, 42.6%, 41%, and 40% (34-37), which are slightly higher than the 26% reported here. WM values were not calculated in this investigation, due to low WM SNR in VT-VSSL maps for several subjects.
QUIXOTIC produces venular blood maps by creating flow-derived signal from within the imaging voxel. In principle, this property allows generation of venular blood maps on a voxel-by-voxel basis. By obtaining maps at several effective echo times (Teff), it is possible to fit for T2 and subsequently obtain voxelwise measures of Yv, OEF, and CMRO2 (as shown in Fig. 9). Of note, we assumed T1 and T2 values of 1660 and 70 ms, respectively, to estimate ΔTEeff via Bloch simulation. Deviations in these assumed values will result in a small bias in the QUIXOTIC measured T2s of venular blood. To examine this effect more closely, we tested: (a) the range of T1s from 1560 and 1760 and found a resulting effect of approximately 0.5 ms on estimated T2, and (b) the range of T2s between 50 and 90 ms, resulting in an effect of approximately 2.5 ms on measured T2. Even at the extremes, these small deviations in measured T2 will result in even smaller deviations on measured Yv, OEF, and CMRO2 (between 1 and 2%); well within the noise of our measurement.
Two key parameters define the nature of the QUIXOTIC signal and directly control its spatial specificity, signal-to-noise ratio (SNR), and maximum SNR: cutoff velocity (VCUTOFF) and TO. When taken together, VCUTOFF and TO determine the extent of the velocity-selected bolus at imaging time by defining its trailing and leading edge, respectively, and subsequently the venular blood component remaining in the subtracted image.
The TO of the QUIXOTIC experiment determines the SNR and spatial specificity of the subtracted venular blood component. During TO, blood spins accelerate above the VCUTOFF threshold and into receiving venular vasculature within the imaging voxel. Imaging signal is created, and intensity increases as these spins begin to fill the imaging voxel. Neglecting T1 decay, the final SI is directly proportional to the volume of blood that has accelerated above VCUTOFF during TO, while still remaining within the voxel (hereafter termed as “imaging blood volume”). Realistically, however, T1 decay will have a competing effect on overall SI; thus in practice, an SNR-optimized TO will consider both imaging blood volume and T1 decay.
As flowing spins in the velocity-selected bolus disseminate through the tortuous venular microvasculature during TO, they flow deeper into the venous circulation, and the leading edge of the bolus travels further away from capillary sites of gas exchange. As long as these tagged spins do not leave the voxel, oxygen saturation within these pools properly represents oxygen exchange of their capillary sources. Spatial specificity is preserved, and imaging blood volume is exactly the product of venular outflow and TO.
VCUTOFF determines the maximum possible SNR of the venular blood component in the final image, by defining the maximum available vascular volume that can be occupied by the venular bolus. As blood velocity is closely related to vessel caliber, VCUTOFF not only defines the cutoff blood velocity, but also determines how far into the venous vascular tree the trailing edge of imaged bolus is located. A low VCUTOFF means that trailing edge originates closer to the distal end of the capillary bed, while a high VCUTOFF means that the trailing edge originates closer to terminal draining veins. As VCUTOFF increases, the trailing edge moves farther along the venous vascular tree, effectively reducing the remaining vascular voxel volume that can be filled by tagged blood. In this way, VCUTOFF sets a hard limit to the maximum imaging blood volume that can occupy a voxel, with lower VCUTOFFs offering higher SI at the optimal TO. A VCUTOFF just above capillary blood velocities is the ideal choice. A VCUTOFF lower than capillary velocities is undesirable for three reasons: (1) VS theory assumes laminar flow, not capillary plug flow (14), for proper velocity filtering, (2) as VCUTOFF decreases, the relaxed blood pool available to flow above VCUTOFF will at some point become small, also resulting in SNR reduction, and (3) when large velocity weighting is applied, tissue and CSF water diffusion becomes a contaminating source of signal (14). Notably, prior VS-ASL studies have indicated that a VCUTOFF ≥ 2 cm/s will result in subtractions free from diffusion contamination (14,38).
Based on the above considerations and pilot data (39), we chose VCUTOFF = 2 cm/s and TO = 725 ms. These choices yield sufficient SNR in the final venular blood images for cortical analysis at 3 T and maintain high spatial specificity. As seen in Fig. 5, venular blood signal appears well-matched to cortical gyri, with minimal blurring, which can be appreciated by comparing the venular blood-weighted images to the CBF and DIR images in Fig. 6.
In this initial demonstration of QUIXOTIC, we performed limited exploration of the dependence of Yv on TO and VCUTOFF parameters. This required development of the turbo QUIXOTIC variant, which allowed the considerably shorter scan times needed to accommodate a multiparameter exploration. Turbo QUIXOTIC differs from the standard approach by using a turbo spin echo readout to acquire data at multiple effective TEs per scan, instead of only at a single TEeff. Consequently, turbo QUIXOTIC generates all necessary images for T2 fitting and subsequent analysis in a single scan, thereby reducing the required imaging time by several fold. In this specific implementation, the total acquisition time for a fixed TO and VCUTOFF combination decreased from 25–30 min to 5.5 min. This dramatic reduction in imaging time enabled us to explore several TO and VCUTOFF combinations in a single scan session. We used this approach in three subjects to see how calculated Yv varied across (1) a range of TOs at a fixed VCUTOFF = 2 cm/s, and (2) a range of VCUTOFFs at a fixed TO = 725 ms. Figure 10 shows the calculated Yv across these parameter ranges. The average slope of −0.007 per 100 ms for the Yv versus TO data (Fig. 10a) suggests that the measured Yv is robust to variations in TO; across a range of 400 ms, this translates to a less than 3% variation in Yv. On the other hand, Fig. 10b indicates a more significant dependence of Yv on the VCUTOFF parameter; an average slope of −0.03 per cm/s suggests that as VCUTOFF increases over a range of 4 cm/s, the measured Yv will decrease by 12%. Whether this effect is an artifact of increased diffusion weighting at lower VCUTOFFs, or a consequence of the physiologic properties of the isolated venular pool that survives after velocity encoding, is the topic of detailed future QUIXOTIC studies.
An additional positive consequence of using turbo QUIXOTIC is that R2 for all T2 fits was >0.99, with an SEE of roughly 3 ms; this marks a substantial improvement over the standard QUIXOTIC variant. We postulate that this improvement is due the fact that turbo QUIXOTIC is considerably more resistant to both bulk and physiological motion for the following reasons: (1) motion affects all TEeff points equally in turbo QUIXOTIC due to parallel TEeff acquisition, whereas it will affect each TEeff point differently in standard QUIXOTIC (due to serial TEeff acquisition) leading to noiser SI versus TEeff data and (2) total scan time is significantly shorter in turbo QUIXOTIC, resulting in less motion artifact across the experiment. Due to higher quality data and reduced scan time, we expect turbo QUIXOTIC to eventually be the variant of choice for Yv, OEF, and CMRO2 voxel-by-voxel mapping.
Values of OEF and CMRO2 estimated with QUIXOTIC fall within the physiologic range and are in good agreement with other PET and MR measures in the literature. The key physiologic parameter measured by QUIXOTIC is local venular oxygen saturation Yv, which shows low measurement variability across 10 subjects (coefficient of variation = 0.03). Of note, however, is that the whole-slice cortical GM Yv measurements by QUIXOTIC were significantly higher than sagittal sinus Yv measured with TRUST (0.73 vs. 0.63 for QUIXOTIC and TRUST, statistically different at P < 0.00001, paired t-test). Moreover, the 95% CI for the true difference in Yv between the techniques lies between 0.08 and 0.11, suggesting a systematic bias between the two techniques. This Yv discrepancy translates into a lower OEF for QUIXOTIC, compared to TRUST (0.26 vs. 0.36). Table 1 highlights these differences, where whole-brain Yv measured with TRUST is on average 14% less than the single-slice cortical GM Yv measured with QUIXOTIC.
Sources of bias may originate from the QUIXOTIC technique itself. Among these is contamination of the venular-derived VS bolus by arterial blood, which could manifest in at least two different ways. As mentioned in Theory section, initially dephased arterial blood will experience T1 recovery after VS1. In an effort to prevent this blood from giving rise to signal in the final subtraction, we inserted a single inversion pulse at a time TO1, which in theory should null this arterial component. TO1 is based on an assumed T1,blood (21); if the true T1,blood differs significantly from this assumed value, the inversion null may not be effective, leading to an arterial contribution in the final subtracted signal. As the T2 of fully oxygenated arterial blood is greater than 150 ms (as empirically determined by ASL-based experiments using an identical T2-preparation module), a small fractional contribution of arterial signal could bias the measured T2, and subsequently Yv. To explore this potential contamination source, we implemented and tested a double-inversion approach in QUIXOTIC (data not shown), which dramatically improves nulling robustness and reduces signal across a much wider T1 range (but imposes an additional specific absorption rate penalty). Comparing the single- and double-inversion variants of QUIXOTIC did not yield a substantial change in measured T2, leading us to conclude that arterial contamination by incomplete T1 nulling is not a dominant source of error.
Another possibility for arterial contamination could be attributed to the nonideal velocity selection profile. While the ideal velocity-selective profile perfectly preserves signal from all spins below VCUTOFF and dephases all spins above VCUTOFF, the actual VS profile follows a sinc envelope as a function of spin velocity (Fig. 2). Consequently, the final VS bolus will also have a sinc-like shape, including side lobes that extend into the arterial side of the circulation. As TO increases, the VS bolus will move further into the venous circulation, and the side lobes on the arterial side become smaller, resulting in less contamination. This potential contamination source was explored with turbo QUIXOTIC in three subjects, and as previously mentioned, there was little variation in measured Yv across a TO range of 400 ms. These turbo QUIXOTIC data suggest that this source of arterial contamination is not a large source of bias in QUIXOTIC.
A second potential source of bias in measuring Yv with QUIXOTIC is the inevitable diffusion weighting in the experiment. The combination of gradient and RF pulses that allow velocity selection will at the same time lead to (mild) diffusion weighting. As the control image acquisition involves turning off the gradients during VS2, the control image will not be diffusion weighted, while the tag image will be slightly diffusion weighted. The subsequent subtraction map will have some degree of diffusion weighting, thus introducing both tissue and CSF components into the otherwise pure venular blood-weighted images. This problem was explored by Wong and colleagues for VS-ASL (14,40). Both studies concluded that diffusion effects were minimal, due to the weak gradients (G), short gradient durations (δ), and large gradient separation (Δ) used in the VS module. In our specific experiments, b value was calculated to be approximately 1 s/mm2, leading to less than 0.07% attenuation in tissue (Dtissue = 0.0008 mm2/s) and less than 0.020% attenuation in CSF (DCSF = 0.0024 mm2/s), given CSF and tissue fractions of up to 95% (41) and up to 10% (42), respectively. Given that the subtracted blood-weighted signal is roughly 2% of the original signal, we expect less than 4% error from diffusion effects at TEeff = 0. However, because the T2s of CSF and tissue are longer than the T2 of deoxygenated blood, diffusion-based effects could become more prominent at longer effective echo times. Results from the turbo QUIXOTIC parameter exploration are consistent with this idea. As seen in (Fig. 10b). Yv (and thus T2) increases with decreasing VCUTOFF. One explanation for this phenomenon is that lower VCUTOFFs result in images containing non-negligible signal from diffusing tissue/CSF water. As water from these sources has a larger T2 than venular blood water, the measured T2 (and subsequent Yv) could be biased upward. In this limited turbo QUIXOTIC dataset, it is also interesting to note that as VCUTOFF increases (and diffusion effects abate), parenchymal Yv values trend towards global Yv values measured by TRUST. Further studies are needed to investigate these effects in greater detail.
Finally, a physiological source to the TRUST-QUIXOTIC Yv discrepancy cannot be ruled out, especially based on the dramatic difference in the location and size of the targeted vessels in each technique. TRUST measures Yv from the final draining vein of the cerebral circulation, the sagittal sinus. At this stage of the circulation, oxygen exchange is complete, and blood is expected to be at its lowest level of oxygenation. In contrast, QUIXOTIC by design measures oxygenation of blood in venules just distal to draining capillary beds. If extraction is not complete by the time blood reaches the distal end of the capillaries, exchange will continue to occur until equilibrium is reached. Such a phenomenon would result in early venular blood to remain more oxygenated than blood in terminal draining veins, resulting in QUIXOTIC Yv values greater than TRUST sagittal sinus Yv values. A more detailed investigation is required to properly determine if such a phenomenon could exist and play role in the TRUST-QUIXOTIC Yv discrepancy.
An important practical difference between TRUST and QUIXOTIC is SNR. SNR is directly related to the imaging blood volume, which is represented by the vertical intercepts in Fig. 7a. Because of the high fractional blood volume in the sagittal sinus (approximately 1), the imaging blood volume is between one and two orders of magnitude greater in TRUST, compared to QUIXOTIC. In this way, TRUST has a major SNR advantage, affording shorter imaging times and higher precision fits.
While beyond the scope of this article, a brief comment can be made on the feasibility of using QUIXOTIC (and more specifically the turbo QUIXOTIC variant) in functional MRI. The primary advantage offered by QUIXOTIC is the ability to generate a OEF and CMRO2 map every two TR (analogous to how CBF maps are generated in ASL-fMRI). While the TR used in these studies was 4 s, shorter TRs (on the order of 2 ms) are possible, with similar temporal SNR. In this way, similar to functional ASL-MRI, we anticipate useful application of QUIXOTIC in block design and event-related fMRI experiments. One potential confounder is the increase of local CBF during activation; however, we do not anticipate this to cause a large error. While the trailing edge of the VT-VSSL bolus will be closer to the capillary bed during activation, venular blood will still be considered and oxygenation properly represented (substantial venular oxygenation heterogeneity, however, could complicate this picture).
QUIXOTIC MRI introduces a novel approach to isolate PCV blood signal and subsequently measure cerebral Yv, OEF, and CMRO2. Values reported for Yv, OEF, and CMRO2 are comparable with those acquired by other PET and MR studies, and fall within a normal physiological range. Advantages of QUIXOTIC include: (1) QUIXOTIC maps blood in the venous circulation only, with CSF, static tissue, and capillary/arterial blood eliminated; (2) subject to SNR constraints, QUIXOTIC analysis can be performed on a voxel-by-voxel basis, allowing creation of Yv, OEF, and CMRO2 maps; and (3) QUIXOTIC generates images every two TR, making the technique amenable to functional imaging of Yv and OEF during block-design and event-related fMRI. To our knowledge, no currently available technique offers all three features. Future studies will explore optimal parameter settings, employ turbo QUIXOTIC for rapid data acquisition, and deploy QUIXOTIC at higher fields for increased SNR.
The authors would like to thank Tim Reese and Thomas Benner for their useful discussion about the velocity selective pulse train.
Grant sponsor: Siemens Medical Solutions; Grant sponsor: Division of Health Sciences and Technology Martinos Catalyst Fund; Grant sponsor: Advanced Multimodal Neuroimaging Training Program Grant; Grant sponsor: NIH; Grant number: RO1 EB007942; Grant sponsor: NIH Medical Scientist Training Program Fellowship; Grant number: T32-GM07753; Grant sponsor: NIH Neuroimaging Training Program Grant; Grant numbers: 5-T32-EB001680, 5-R01-EB002066-20; Grant sponsor: Center for Functional Neuroimaging Technologies; Grant number: P41RR14075S10RR023401.