Functional oxygen extraction fraction (OEF) imaging with turbo gradient spin echo QUIXOTIC (turbo QUIXOTIC)

Abstract

Purpose

QUantitative Imaging of eXtraction of Oxygen and TIssue Consumption (QUIXOTIC) is a recent technique that measures voxel-wise oxygen extraction fraction (OEF), but suffers from long scan times limiting application. We implement multi-echo QUIXOTIC dubbed turbo QUIXOTIC (tQUIXOTIC) that reduces scan time by 8× and then apply it in functional MRI.

Methods

tQUIXOTIC utilizes a novel turbo gradient spin echo readout enabling measurement of venular blood transverse relaxation rate in a single tag-control acquisition. Using tQUIXOTIC we estimate cortical gray matter (GM) OEF, create voxel-by-voxel GM OEF maps, and quantify changes in visual cortex OEF during a blocked design flashing checkerboard visual stimulus. Contamination from CSF partial volume averaging is estimated and corrected.

Results

Average (N=8) cortical GM OEF was estimated as 0.38±0.06 using a 3.4 minute acquisition. Average OEF in the visual cortex was estimated as 0.43±0.04 at baseline and 0.35±0.05 during activation, with average %ΔOEF of -20%. These values are consistent with past studies.

Conclusion

tQUIXOTIC is introduced and shown to successfully estimate cortical GM OEF in clinical scan times and to detect changes in OEF during blocked design visual stimulation. tQUIXOTIC will be useful to monitor regional OEF clinically and in blocked design or event-related fMRI experiments.

Introduction

Normal brain function depends on energy made available via cerebral oxygen metabolism. Oxygen extraction fraction (OEF) is directly related to the rate of cerebral oxygen metabolism (CMRO₂), and is a key quantitative indicator of brain function. As such, robust and reliable measurements of OEF have important clinical and basic neuroscience implications (1–3). Routine imaging of OEF, however, is not commonly performed in clinical or research settings, in part due to long acquisition times and complex protocols of typical MRI and PET approaches (4,5).

Clinically, OEF is relevant as an indicator of stroke risk (1) and a potential indicator of stroke ischemic penumbra, that is tissue at risk, but still viable (2). From a neuroscience standpoint, OEF offers a more direct assessment of metabolic changes than the traditional BOLD response, which is the mainstay of functional MRI (6,7). OEF reflects the balance of oxygen delivery and oxygen metabolism and is related to CMRO₂ by Fick's principle:

$$\text{OEF}=\frac{{\text{CMRO}}_2}{[\text{Hb}]\cdot \text{CBF}\cdot {\mathrm{S}}_{\mathrm{a}}{\mathrm{O}}_2}$$ [1]

where [Hb] is the blood concentration of hemoglobin, CBF is cerebral blood flow and S_aO₂ is the oxygen saturation of arterial blood (8,9). SaO₂ can be measured using a pulse oximeter, or in healthy adults assumed to be fully saturated.

There are three main classes of MRI-based techniques for quantifying regional cerebral OEF: magnetic susceptibility, T₂ relaxometry, and quantitative BOLD, reviewed in (3,4,10). These all measure OEF by quantifying effects of paramagnetic deoxygenated hemoglobin (dHb) in venous blood. T₂-relaxometry approaches are particularly appealing, as they typically offer less sensitivity to non-dHb sources of magnetic field inhomogeneity by eliminating extravascular signal. QUantitative Imaging of eXtraction of Oxygen and TIssue Consumption (QUIXOTIC) is among these intravascular T₂-relaxometry based approaches, in which T₂ of post-capillary venular blood is estimated and calibrated to venous oxygen saturation (SvO₂) and OEF (11–14).

The key innovation of QUIXOTIC is its ability to isolate signal from post-capillary venular blood on a voxel-wise basis, by using specialized velocity selective pulses that exploit differential blood velocities through the circulation (15). A two-step data acquisition paradigm is used where control and tag acquisitions are interleaved. In the control acquisition, a velocity selective module saturates all spins moving faster than some user defined cutoff velocity (V_CUTOFF). Blood spins are then allowed to decelerate into the capillary bed and accelerate into draining venules before being imaged some “outflow” time (TO) later. In the tag acquisition, after the first velocity selective module and outflow time, a second velocity selective module is used to saturate all blood spins that have accelerated into venules (faster than V_CUTOFF). Subtracting tag from control images generates an image where the signal primarily originates from blood in the post capillary venules.

A major limitation of QUIXOTIC is its long scan duration—only one echo time (TE) is acquired per repetition time (TR) and many repetitions are required to obtain enough SNR at the several echo times needed to determine T₂. The traditional QUIXOTIC approach therefore focused on acquiring one echo per scan, with each scan lasting several minutes. Standard single-run, blocked design functional MRI was not possible since T₂ (and thus OEF) could not be calculated every TR (16).

In this study, we have modified and improved the original QUIXOTIC approach (15) by implementing a turbo gradient spin echo (GRASE) (17) readout that acquires multiple echoes in a single TR, resulting in a several-fold decrease in imaging time, improved robustness to bulk motion and physiological noise, and providing the possibility to acquire a functional time series. We call this technique turbo QUIXOTIC (tQUIXOTIC). We demonstrate its effectiveness in estimating gray matter (GM) OEF in clinically feasible scan times, and then assess its ability to measure OEF changes in the visual cortex in response to a blocked design visual stimulus. We also estimate and correct for the effects of cerebrospinal fluid (CSF) diffusion attenuation on the tQUIXOTIC signal, a previously described confounder in the QUIXOTIC OEF measurement that is related to the application of velocity selective gradients intrinsic to the technique (15,18–20).

Methods

Construction of turbo QUIXOTIC sequence using a turbo gradient spin echo (GRASE) readout

To construct the tQUIXOTIC sequence, a custom GRASE readout utilizing adiabatic inversion pulses and spoiler gradients was added to the previously described QUIXOTIC pulse sequence (15) such that a complete set of control or tag images with different TEs is acquired in a single TR (Figure 1). This implementation of GRASE differs from the original (21) in that multiple images are generated per TR. We used a RF-pulse train with spoiler gradients for accurate and robust T₂ mapping (22). The spoiling gradients alternate in sign and linearly decrease in magnitude with every +/- pair. For improved robustness to B0 and B1 inhomogeneity, hyperbolic secant adiabatic pulses were implemented as the refocusing pulses. These pulses are nonselective, have a 3.142 ms duration, and are the same as those used in the velocity selective module (15,18). One complete echo planar imaging (EPI) readout is performed between each refocusing pulse and shortened to 12.6 ms using generalized autocalibrating partially parallel acquisitions (GRAPPA) parallel imaging technology (23), acceleration factor = 3. Notably, only even echoes from the GRASE readout were used for TE₁, TE₂, and TE₃ (as seen in Figure 1), as the odd echoes are contaminated by non-linear phase accrual from a single adiabatic inversion pulse, which is reversed after a pair (24). The accuracy of T₁ and T₂ quantification using the standalone GRASE module was verified using doped agar phantoms (3% Agar, 0.03 mM gadolinium) with similar T₁ and T₂ measurements to GM (15,25). The tQUIXOTIC sequence parameters used in the following experiments were V_CUTOFF = 2.3cm/s, TO₁ = 400 ms, outflow time (TO) = 725 ms, with GRASE readout parameters voxel size = 3.9×3.9×8 mm³, FOV = 250 mm, TE₁/TE₂/TE₃/TR =25.2/50.4/75.6/3000 ms, BW = 3256 Hz/pixel.

Figure 1.

Turbo QUIXOTIC Pulse Sequence. QUIXOTIC velocity selection followed by GRASE readout with one image acquired per TE.

Subject Selection and Data Acquisition

All scans of adult subjects took place at the Athinoula A. Martinos Imaging Center at the McGovern Institute for Brain Research, MIT. Eight healthy volunteers (aged 19 to 30, 4 males and 4 females) were scanned with IRB approval on a Siemens Tim Trio 3T scanner (Siemens Healthineers, Erlangen, Germany) using the Siemens 32-channel head coil.

The slice position for tQUIXOTIC imaging was chosen to intersect an area of the visual cortex that demonstrated a strong BOLD response to a visual stimulation paradigm. The calcarine fissure was located by visually inspecting a 1 mm³ isotropic resolution three dimensional T₁-weighted, gradient echo structural scan (FOV = 256×256×176 mm³ TE/TR = 3.44 / 2530 ms, GRAPPA acceleration factor = 3, total acquisition time = 4.3 min). An axial 2D slice selective EPI spin-echo (EPI-SE) scan that covered the primary visual cortex with seven slices was then acquired during a blocked design visual stimulus (8 Hz flashing radial checkerboard with central fixation point projected onto a translucent screen and viewed through a mirror mounted to the head coil) with one-minute periods of a central visual fixation point interleaved with one-minute periods of a central fixation point plus a radial checkerboard, for a total of five minutes (Figure 2). The EPI-SE scan parameters were: voxel size = 3.9×3.9×8 mm³, FOV = 250 mm, TE/TR = 60/3000 ms, TR delay = 2385 ms, BW = 3256 Hz/pixel, scan time = 5 min. These data were analyzed by fitting a linear signal model consisting of regressors representing the blocked design stimulus, a linear and quadratic drift term, and a constant (DC) term. The slice showing the highest t-statistic values and largest volume of activation during the EPI-SE scan was selected for tQUIXOTIC imaging.

Figure 2.

Experimental Design. Visual stimulus paradigm of interleaved central visual fixation point (1 min) with flashing radial checkerboard (1 min), with total duration of 5 minutes for EPI-SE and 7 minutes for tQUIXOTIC. Example tQUIXOTIC tag or control image sets acquired for every TE in one TR = 3s.

tQUIXOTIC imaging then took place in four 7-minute runs (4 minutes total at baseline and 3 minutes total during exposure to visual stimulus). During each run the aforementioned visual stimulus paradigm was displayed (Figure 2). After each tQUIXOTIC run, a double inversion recovery (DIR) image for GM-only imaging was acquired using the same slice position and spatial resolution. DIR imaging parameters were, TE/TR = 13/4340 ms, inversion times = 3700 ms and 4280 ms to select for GM only, 2232 Hz/Px bandwidth, and scan time = 6 s.

Turbo QUIXOTIC Analysis for OEF measurement in cortical GM and during visual stimulation

Analyses were performed using custom scripts for Neurolens (www.neurolens.org) and MATLAB (The Math Works, Natick, MA). Each tQUIXOTIC run was individually analyzed to determine OEF of cortical GM at baseline. Simple pairwise subtraction of tag and control images for corresponding TEs generated a tQUIXOTIC time series for each TE (TE₁, TE₂, and TE₃, Figure 2). Each series was smoothed with a 6 mm Gaussian kernel (26). The acquisitions during visual fixation only (excluding the 12 seconds following each period of stimulation, to allow OEF to return to baseline) were averaged across time to get a single baseline tQUIXOTIC image per TE. The corresponding DIR-GM image was used as a mask to select GM voxels within the baseline tQUIXOTIC images and the average GM signal intensity was calculated. For one subject, DIR images were not acquired; in this case the GM mask was generated by segmenting the EPI-SE images using FSL tools (BET and FAST) (27). These resulting tQUIXOTIC GM signal intensities were plotted against echo time (TE₁, TE₂, and TE₃) and fit using the two-compartment signal model described below (Equation 3) using a non-linear least squares optimization method (MATLAB) to determine T₂ for venous blood in the cortical GM. T₂ values were then calibrated to venular blood oxygen saturation (SvO₂) based on empirical and biophysical models described in (11–14). This mapping depends on the hematocrit value for each subject, which we assumed to be the average standard values for hematocrit 0.45 for males and 0.42 for females (28), and on the inter-echo spacing of the GRASE train (12.6 ms). OEF for each state was then calculated as (SaO₂-SvO₂)/SaO₂, assuming fully saturated arterial blood in healthy adults (SaO₂ = 1).

Representative baseline OEF maps were generated for each tQUIXOTIC run for a single subject. As before, the baseline tQUIXOTIC data was pairwise subtracted, smoothed (10 mm Gaussian kernel) (15) and averaged across time, similar to the original QUIXOTIC mapping approach. On a voxel-wise basis, the tQUIXOTIC signal in cortical GM was again fit using the two-compartment signal model described below (Equation 3) to determine T₂ for venous blood in each voxel. SvO₂ and OEF were determined as above.

To determine OEF of a functionally activating region during both baseline and activation, the four tQUIXOTIC runs were averaged together. The average 7 min tQUIXOTIC run was then pairwise subtracted and smoothed (6 mm Gaussian kernel). For each TE, a linear signal model was fit to the 7-minute time series of each voxel. The linear signal model consisted of regressors representing the blocked design stimulus, a linear drift term, and a constant (DC). Maps of activation t-statistics were generated for each TE (Figure 3a). The t-statistic map for TE₃ was used to select a region of interest of the ten most significantly activating voxels (exceeding P < 6.5×10^-4, corresponding to t-statistic > 3.35) excluding the region within 10 mm of the center of the superior sagittal sinus to avoid contamination from sinus venous blood. Beta-coefficients (β_DC and β_EFFECT) from the GLM fit were averaged over the region of interest. Fits of β_DC (baseline) and β_DC + β_EFFECT (baseline plus effect size) for the three even echoes with the two-compartment signal model described below (Equations 2 and 3) determined T₂ values at baseline and during activation (Figure 3b).

Figure 3.

Representative data analysis for subject 7. A) t-statistic map (t > 3.4) for the GLM fit of the TE₃ = 76.6 ms venular blood weighted image series overlaid on the first tag image (TE₁ = 25.2). SSS contaminated voxels determined from structural imaging are overlaid in purple and removed from further analysis. GLM regressors were stimulus convolved with hemodynamic response function (EFFECT), DC offset (DC) and linear drift. B) Corresponding plot of beta-coefficients from GLM (β_DC and (β_DC + β_EFFECT) from within activation ROI (10 voxels) versus TE, with T₂ fit lines.

T₂ Estimation and correction of CSF contamination to the tQUIXOTIC signal

Applying velocity selective gradients immediately before the imaging readout for the QUIXOTIC tag acquisition, but not during the control acquisition, results in a very small amount of diffusion weighting that is not eliminated upon control-tag subtraction (18). While this is negligible for most voxel constituents, a small component of CSF will persist and slightly contaminate the pure venular blood signal at longer echo times since T_2,CSF is much longer than T_2,blood (20). Assuming a two compartment model for blood and CSF in the voxel, the measured signal would be:

$$\mathrm{S}(\text{TE})=k_{\mathit{\text{blood}}}{\mathrm{e}}^{-\frac{\text{TE}}{{\mathrm{T}}_{2,\text{blood}}}}+k_{\mathit{\text{CSF}}}{\mathrm{e}}^{-\frac{\text{TE}}{{\mathrm{T}}_{2,\text{CSF}}}}.$$ [2]

Where, k_CSF and k_blood are the y-intercepts and T_2,CSF and T_2,blood are the transverse relaxation time constants for each compartment. Consequently, CSF signal contamination creates an upward-bias in the estimated T_2,blood when a monoexponetnial fit is used, which results in a lower estimated OEF.

Corrected values for baseline T_2,blood and S₀ were determined by fitting a two compartment model, to the average tQUIXOTIC signal intensities, S(TE):

$$S(TE)=S_0[(1-X_{\mathit{\text{CSF}}})e^{-\frac{\mathit{\text{TE}}}{T_{2,\text{blood}}}}+(X_{\mathit{\text{CSF}}})e^{-\frac{\mathit{\text{TE}}}{T_{2,\mathit{\text{CSF}}}}}].$$ [3]

X_CSF is the magnetization fraction of diffusion attenuated CSF in a GM voxel acquired with the specific tQUIXOTIC protocol that we used. X_CSF and T_2,CSF were determined from calibration data from one subject through methods described in the appendix, and then used to correct the data for all other subjects. This assumes that that the volume fraction of CSF within GM voxels is relatively constant across humans, which is supported by several studies (29,30).

Since there is increased cerebral blood flow during neuronal activation X_CSF (fundamentally the ratio of diffusion attenuated CSF signal to blood signal) will change during activation. To account for this change, corrected values for T_2,blood and k_blood during activation were determined by fitting Equation 2 to the average tQUIXOTIC signal intensities during stimulation, using k_CSF = X_CSF*S₀ and T_2,CSF as determined from the corresponding baseline fitting. This approach reflects that CSF volume is assumed to remain constant during brain activation (29).

Results

Average cortical GM OEF at baseline for four runs from each subject are shown in Figure 4a, and values for T₂, SvO₂, and OEF are given in Table 1. The average T₂ for all subjects was 58±9 ms. Average SvO₂ was estimated as 62±6% and OEF as 0.38±0.06. Exponential fitting of the tQUIXOTIC signal averaged across GM voxels and over 3.4 minutes during baseline blocks, was high quality with R² > 0.99 for all fits. For OEF, the average coefficient of variation (COV) for all subjects across runs was 7%.

Figure 4.

(a) Average baseline GM OEF, corrected for CSF contamination, four runs for each subject, where a GM mask is used to average the tQUIXOTIC signal spatially before T₂ fitting. Baseline acquisition took 3.4 minutes per run. The average same session COV was 5.6%, not including the one motion degraded run for subject 5. (b) Average baseline GM OEF, baseline OEF in the visual cortex and activation OEF in the visual cortex, corrected for CSF contamination.

Table 1.

Baseline GM T₂, SvO₂, and OEF for all subjects.

	T2 [ms]						SvO2 [%]						OEF
Subject	Run 1	Run 2	Run 3	Run 4	Mean	STD	Run 1	Run 2	Run 3	Run 4	Mean	STD	Run 1	Run 2	Run 3	Run 4	Mean	STD
1	61	59	61	58	60	2	65	64	65	63	64	1	0.35	0.36	0.35	0.37	0.36	0.01
2	63	57	55	55	57	4	66	63	61	62	63	2	0.34	0.37	0.39	0.38	0.37	0.02
3	74	71	80	76	75	4	70	69	73	71	71	2	0.30	0.31	0.27	0.29	0.29	0.02
4	68	63	60	65	64	3	67	65	63	66	65	2	0.33	0.35	0.37	0.34	0.35	0.02
5	44	39	37	63	46	12	52	48	45	65	52	9	0.48	0.52	0.55	0.35	0.48	0.09
6	56	57	59	54	57	2	62	62	64	61	62	1	0.38	0.38	0.36	0.39	0.38	0.01
7	55	46	50	44	49	5	60	53	57	52	55	4	0.40	0.47	0.43	0.48	0.45	0.04
8	58	53	63	62	59	5	63	60	66	66	64	3	0.37	0.40	0.34	0.34	0.36	0.03
All Data					58±9						62±6						0.38±0.06

Representative OEF maps at baseline in GM are shown in Figure 5 for Subject 1. Average voxel-by-voxel OEF was 0.35±0.01 across the four maps, compared to 0.36±0.01 for the GM mask analysis.

Figure 5.

OEF maps in the GM (determined by a separate DIR GM-selective scan) for each run of subject 1, after correction for CSF contamination, generated by fitting the tQUIXOTIC signal from each voxel.

During visual stimulation, significant changes in T_2,blood, SvO₂, and OEF were observed in the visual cortex. Results corrected for CSF contamination are given in Table 2. Figure 4b presents average cortical GM OEF at baseline, and OEF in the visual cortex at baseline and activation for each subject. Average baseline T_2,blood was 50±6 ms and activation T_2,blood was 64±8 ms. OEF changed (paired t-test, P = 0.00001) from 0.43±0.04 at baseline to 0.35±0.05 during activation with an average relative change of -20%. Exponential fitting of the tQUIXOTIC signal at baseline and activation averaged across ten voxels in the visual cortex was high quality with R² > 0.99 for all fits.

Table 2.

Baseline and activation T₂, SvO₂, and OEF values from the visual cortex showing a significant decrease in OEF with activation (P = 0.00001).

Subject	T2 baseline [ms]	T2 activation [ms]	ΔT2 [ms]	SvO2 baseline [%]	SvO2 activation [%]	OEF baseline	OEF activation	%ΔOEF
1	52	71	18	60	70	0.40	0.30	-26
2	60	75	16	64	72	0.36	0.28	-23
3	53	68	16	58	67	0.42	0.33	-22
4	43	50	7	51	56	0.49	0.44	-12
5	48	58	10	55	62	0.45	0.38	-15
6	54	64	10	61	66	0.39	0.34	-15
7	47	61	14	54	63	0.46	0.37	-20
8	45	62	17	54	66	0.46	0.34	-25
Mean±STD	50±6	64±8	14±4.0	57±4	65±5	0.43±0.04	0.35±0.05	-20±5

Data above is reported after correction for CSF contamination using X_CSF = 0.112, determined from calibration scans in cortical GM at 3.9×3.9×8 mm³ resolution. Fit parameters for determining X_CSF as described in the appendix are given in Table 3. CSF correction did not alter the statistical significance of any reported comparisons. In cortical GM at baseline, uncorrected T₂ was on average 16±0.4 ms higher than corrected. Average uncorrected OEF was 0.29±0.05, or 0.09±0.01 lower than corrected OEF. X_CSF during activation—recalculated from k_CSF and k_blood after fitting—was 0.083±0.01 across all subjects.

Table 3.

Fit parameters for model given in Equation A3, with T_2,CSF = 1.2 s.

Model	Residual sum of squares	x1	x2	x3	x4	x5
(x1(ΔTE) + x2)e−TE/x3 + (x4(ΔTE) + x5)e−TE/1.2	0.0511	-57.1	6.85	0.061	-4.88	0.835

Discussion

We have introduced a GRASE implementation of QUIXOTIC that permits estimation of regional cerebral oxygenation in a fraction of the original imaging time, making the technique feasible for clinical and functional imaging. The average estimate of baseline cortical GM OEF falls in the expected physiological range and is comparable to results from other MRI (9,26,31) and PET studies (32–34). We were able to demonstrate measurement of OEF in GM in 3.4 minutes, making tQUIXOTIC one of the fastest approaches for determining baseline OEF (4). It permits GM OEF mapping and was validated by performing a blocked design functional activation experiment using a flashing visual stimulus. Our results demonstrate a significant decrease in OEF during full field visual stimulation (average %ΔOEF = -20%, P = 0.00001) in agreement with previously reported values for OEF in visual activation paradigms (Table 4). To our knowledge, these results are the first MRI based measurements of regional OEF change with brain activation, based on a single functional time course, and without the use of external gases like oxygen or carbon dioxide. This becomes possible since tQUIXOTIC permits estimation of T₂ every two TR, since all echoes are acquired per TR.

Table 4.

Baseline and activation OEF in published studies.

Study	Method	Baseline OEF	Activation OEF	Brain Region
Oja, et al. (1999). (9)	1.5T, T2 mapping (N = 7)	0.3±0.06*	0.2±0.07*	Visual cortex
An and Lin, (2000). (36)	1.5T, multi-echo gradient and spin echo (N = 8)	0.416±0.018		Whole brain
He, et al., (2007). (35)	3T, quantitative BOLD (N = 9)	0.383±0.053		Whole brain
Cheng, et al. (2015). (44)	7T, VASO (N = 6)		0.24±0.01*	Visual cortex
Fan, et al. (2014). (38)	3T, QSM (N = 10)	0.275±0.03*		Occipital pial veins
Liu, et al. (2013). (37)	3T, TRUST (N = 7)	0.375±0.05		Whole brain
Guo, et al. (2012). (19)	3T, VSEAN (N = 4)	0.395±0.03		Gray matter
Bolar, et al. (2011). (15)	3T, QUIXOTIC (N = 10)	0.26±0.02		Gray matter
Ito, et al., (2005). (34)	PET (N = 70, multi-study review)	0.44±0.06		Whole brain
Ibaraki, et al. (2008). (5)	PET (N = 8)	0.40±0.06		Occipital cortex
		0.35±0.06		Gray matter
Gauthier, et al. (2012). (26)	3T, calibrated BOLD (QUO2) (N = 7)	0.29±0.05*	0.23◊*	Visual cortex
Gauthier, et al. (2012). (46)	3T, calibrated BOLD (QUO2) (N = 8)	0.35±0.04		Gray matter
Fan, et al. (2016). (39)	7T, calibrated BOLD (QUO2) (N = 11)	0.435±0.14		Visual cortex
Wise, et al., (2013). (40)	3T, calibrated BOLD (N = 11)	0.48±0.09	0.45±0.14	Visual cortex
		0.42±0.12		Gray matter
This study	3T, tQUIXOTIC (N = 8)	0.43±0.04	0.35±0.05	Visual cortex
		0.38±0.06		Gray matter
		0.29±0.05		GM, Not corrected for CSF

^◊calculated from published results,

^*different from comparable tQUIXOTIC result (P < 0.05))

tQUIXOTIC results without CSF correction are similar to the previously published QUIXOTIC results (see values in Table 4). Though QUIXOTIC scans are lengthy, baseline QUIXOTIC and tQUIXOTIC measurements of T₂ and OEF in cortical GM from one subject were acquired as a sanity check. Both scans were acquired from the same slice location, and GM was selected using a DIR scan as a mask. QUIXOTIC sequence parameters were as described in (15). For QUIXOTIC GM T₂ and OEF were 83±19 ms (mean ± standard error of the estimated parameter) and 0.25, respectively, and for tQUIXOTIC 77±3 ms and 0.28. These values are indistinguishable considering the standard errors, and fall within one standard deviation of the mean values for all subjects scanned with each technique. The standard errors given for both T₂ estimates underscore that though raw SNR over an imaging run is 2× lower using the GRASE readout in tQUIXOTIC, we observe improved fit quality in tQUIXOTIC (R²>0.99) versus QUIXOTIC (R²>0.95), which we attribute to the lower physiological noise and less bulk motion captured with tQUIXOTIC. tQUIXOTIC 8× faster overall, meaning that in terms of SNR per unit time tQUIXOTIC outperforms QUIXOTIC by 4×, which reduces the physiological noise during acquisition. Furthermore, by acquiring images from all TEs in every control and tag repetition there are far fewer periods were bulk motion can disrupt the tag-control difference images, and any motion should affect one repetition's images similarly.

Numerous PET and MRI studies find baseline OEF around 0.4 (34–37). Table 4 summarizes the OEF findings from other studies and supports the following comparisons. Without correcting for CSF contamination our estimates of cortical GM OEF values are in the lower MRI range (9,26,38), however after correction they are better agreement with both PET and MRI estimates (5,34–36,39,40), suggesting the CSF correction improves QUIXOTIC accuracy. Our standard deviation across subjects for cortical GM OEF of 0.06 is similar to the values reported by PET and MRI studies. Since tQUIXOTIC scans to determine OEF in GM at baseline required just a few minutes of data, we were able to perform multiple OEF measurements for each subject to ascertain intrasubject, same session, variability (Figure 4 and Table 1). The average same session COV, a metric of measurement noise, was 5.6% (discounting the large variation for subject 5 caused by a motion degraded run), which is slightly higher than the 3.2% reported for the whole-brain T₂-relaxation under spin tagging (TRUST) technique (37), even though TRUST benefits from high SNR in comparison to tQUIXOTIC. The intersubject COV, which in addition to measurement noise reflects the variance due to different physiology between subjects, was 15%. This is similar to that reported for other MRI techniques and corroborates the idea that baseline OEF is quite variable between subjects (41).

tQUIXOTIC is a voxel-by-voxel technique capable of producing cortical OEF maps in 3.4 minutes, as demonstrated in Figure 5. We observe that variation between maps from different runs is small, and that these maps appear qualitatively similar to those generated by qBOLD (35), and T₂ calibrated BOLD (36). It is reassuring that average OEF calculated for the voxel-by-voxel analysis and whole GM analysis were in excellent agreement. However, fits were generally not possible in white matter due to low SNR. Combining tQUIXOTIC runs was attempted to increase SNR, but suspected out-of-plane motion between runs and physiological noise (42) (over the course of 28 minutes of total scanning) diminished SNR gains despite the use of 2D-motion correction, and OEF mapping of white matter was not improved. Since tQUIXOTIC, as implemented, is a single slice technique with fairly low in-plane resolution (64×64 voxels in plane), out-of-plane motions are not well addressed by standard motion correction approaches. Future approaches will consider prospective motion correction strategies to improve SNR with longer acquisitions (43).

A robust response in the visual cortex to a flashing radial checkerboard stimulus makes it a popular paradigm for characterizing changes in cerebral oxygenation upon stimulation (9,11,26,40,44). We examined changes in OEF in the visual cortex due to visual stimulation to further validate tQUIXOTIC GM results and demonstrate the feasibility of tQUIXOTIC for detecting functional changes in brain state. The GRASE implementation is an important advance for functional imaging, as the required echoes for T₂ fitting and SvO₂ or OEF estimation are acquired in single tag-control acquisitions, giving a time point every two repetitions, which is similar to standard arterial spin labeling approaches that have been commonly used for functional challenges (6). This advance permits use of standard single-run, blocked and event-related functional paradigms, not possible with the original implementation. Whereas functional imaging with BOLD depends on blood volume, blood flow and tissue metabolism (7), and arterial spin labeling is only a marker of vascular changes (6), OEF-based functional imaging is related to the underlying balance between tissue metabolic demand and oxygen delivery. Ultimately, changes in CMRO_2, which could be estimated by combining tQUIXOTIC derived OEF with a measure of local CBF (15), would permit direct quantitative assessment of cerebral metabolic demand. Further, given that baseline SvO₂ or OEF has been shown to modulate BOLD and CBF signal changes (39,41), fast approaches to OEF mapping may also provide useful calibration in multi-subject fMRI studies.

Our results for changes in OEF with visual stimulation are comparable to two previous calibrated BOLD studies at 3T (26) and (40), see Table 4. Our value for %ΔOEF is in excellent agreement with (26), -20% versus -21%, which used a similar visual stimulus (off 1 min, on 3 min, off 2 min). Notably our result for absolute OEF of 0.43 at baseline in the visual cortex is higher than that found in (26) of 0.29. Calibrated BOLD uses the information provided by hyperoxia and hypercapnia calibration experiments to determine baseline properties of the BOLD signal, which are then used to estimate absolute OEF and CMRO_2. In contrast, tQUIXOTIC is used to directly measure the T₂ from the venular blood compartment, which is empirically and theoretically related to SvO₂ (14). Thus, despite different modeling and dominant signal source (intravascular versus extravascular) the agreement between (26) and tQUIXOTIC for assessing relative change in OEF during activation in the visual cortex suggests both methods are sensitive to the same physiological change with stimulation when a similar stimulation paradigm is used. However, the different absolute OEF may be attributable to these underlying differences in methodology. Interestingly, the combined hyperoxia-hypercapnia calibrated BOLD technique found a higher baseline OEF of 0.435 at 7T (39), nearly identical to our value.

The other calibrated BOLD study, (40), found an absolute OEF at baseline in the visual cortex of 0.48, which is even higher than our own, but a much smaller %ΔOEF of -6.25% than observed in (26) or our own study. The authors of (40) attribute the small %ΔOEF to a decline in stimulus induced CMRO₂ over the long 20 min stimulation paradigm, which was necessary to present a graded gas challenge to calibrate the BOLD signal model used to calculate OEF and CMRO₂. These differences draw attention to the fact that the metabolic response to visual stimuli is still being investigated (45) and variations in the stimulus presentation such as visual field coverage, presentation duration and subject attention may also influence the OEF response.

We compared OEF in the visual cortex and in cortical GM for eight subjects (Figure 4b). In 6 subjects visual cortex OEF is higher than GM OEF (Figure 4b), which is a trend supported by MRI and PET studies (5,40). In two of the subjects the trend is reversed which is similar to that found in (26,46). Differences in the apparent distribution of OEF across the cortex could be explored with an upgraded tQUIXOTIC with improved SNR and multi-slice acquisition as we discus below, while a greater number of subjects would be needed to address the overall tendencies in OEF within the primary visual areas.

Our correction for signal contamination due to diffusion-attenuated CSF resulted in a greater change in T_2,blood (-15.5 ms) than a previous study (-10.2 ms) that explored the contamination in the QUIXOTIC technique (20). Differences in sequence implementation that affect the degree of diffusion attenuation or the volume of the selected blood pool would directly impact the magnetization fraction, X_CSF, which could explain these different changes in T₂. The approach of biexponential modeling to determine the signal contamination of CSF has been used previously (47). The signal contamination we determined is small given that the CSF volume fraction in a typical voxel (10.7% at baseline) is about double the blood volume (29). Most of the CSF signal is being eliminated during control-tag subtraction. In the healthy young adults who were subjects in our study, we are confident that our approach to CSF correction requires calibration scans only once per voxel size and QUIXOTIC parameter set, and the relevant correction factor can be applied to all similar scans. Differences in the CSF volume fractions between subjects has only a small effect on corrected OEF estimates (COV = ±3.4% and no bias, based on a Monte Carlo simulation using the 13% intersubject variation in CSF volume fraction reported in (30) as the expected variation in X_CSF) in comparison to the average same session variation we found (COV = 5.6%).

Many central nervous system pathologies that cause focal or regional oxygenation changes will not have significant fluctuations in CSF content, particularly in their early stages (e.g. early ischemic stroke). When detecting changes in OEF, non-CSF corrected tQUIXOTIC will certainly be useful in some cases, as the lesional OEF values can be compared to OEF in regions of normal brain. However, several central nervous system pathologies (e.g. Alzheimer's disease, certain brain tumors) and old age, can lead to different CSF to tissue volume ratios, in which case X_CSF might be a source of error in corrected OEF estimates. The proposed CSF correction addresses the volume fraction effect arising from single voxels (0.12 ml) deemed GM based on DIR scans. It is unclear how an overall change in CSF volume and brain volume affects the volume fraction of CSF in a nominally GM voxel of the size we use, since most studies investigating CSF volume fraction changes have been undertaken to correct the metabolite concentrations obtained from much larger (∼8 ml) MR spectroscopy voxels, with one study suggesting no detectable change (48) while another shows larger intersubject variation that accompanies disease (49). Though higher resolution may minimize these fluctuations in CSF volume fraction, in specific settings like Alzheimer's disease, brain tumor, or studies of aging, we suggest that measuring X_CSF is important to ensure that estimates of OEF are not biased.

A two compartment exchange model was used to perform the T₂ to SvO₂ calibration (14,50), which depends on the subject's hematocrit and inter-echo spacing of the T₂ preparation module, with model coefficients empirically derived from T₂-relaxation under spin tagging (TRUST) data (14). TRUST varies TE based on pre-readout T₂ preparation, which is inherently insensitive to flow, in contrast to our GRASE approach, which will be slightly sensitive to flow effects. However, we assume that flow effects are minimal between the short 12.6 ms echoes, particularly given that most of the venular blood signal within the cortex is originating from low-velocity microvasculature. We acknowledge a small bias may exist since later echoes will have their corresponding signal intensity biased upward due to increased signal accumulation in venules. Experiments varying the inter-echo spacing will further characterize the magnitude of this error, although preliminary work in (15) suggests that this is a small source of error.

Additional work is also needed to characterize the tQUIXOTIC signal originating from larger vessels, such as the superior sagittal sinus, which limited the choice of activated visual cortex voxels in this particular study. Theoretically, though tQUIXOTIC velocity selection should eliminate signal in large veins, turbulence and changes in flow direction when vessels emerge from the cortex and combine with larger veins may mean that some apparently slowly moving blood arrives in the superior sagittal sinus during TO. Similarly, signal appearing in the superior sagittal sinus in velocity selective arterial spin labeling experiments has been attributed to the nonmonotonic velocity of blood as it drains from the microvasculature into large vessels (51).

Future work on tQUIXOTIC will focus on improving the SNR of the technique and implementing a multi-slice readout. One approach to improve SNR will be to optimize the radio frequency pulse train in the GRASE module to reduce the signal intensity difference between odd and even echoes (52,53). Optimization will also reduce specific absorption ratio (SAR). Currently we operate tQUIXOTIC scans with the SAR limit set to IEC Level 1, and we were SAR limited to acquire only 6 echoes for the 7 minute tQUIXOTIC runs. We could further minimize SAR by empirically determining the lowest power necessary to meet the adiabatic threshold for the inversion pulses. tQUIXOTIC was implemented as a single slice technique such that one complete image could be acquired per TE and 6 second temporal resolution could be achieved overall. However, a multi-slice acquisition would significantly increase the value of tQUIXOTIC, in both clinical and functional settings, where regions of interest may not be known a priori and more regional measurements are desired. Specifically, improved volume coverage would permit exploration of how OEF is tied to neural activity, perhaps using retinotopic mapping to examine the specificity of BOLD in comparison to tQUIXOTIC based OEF. When a lower image-to-image temporal resolution is acceptable (for example in clinical and non-functional MRI applications), but multiple slice coverage is desired, the GRASE module could be substituted with a 3D segmented readout (e.g. 3D spiral or 3D GRASE) (54,55). For fMRI applications that still require high image-to-image temporal resolution, simultaneous multi-slice acquisition technology will allow acquisition of several slices per TR (56).

Conclusion

Turbo QUIXOTIC marks a 8× decrease in acquisition time over the original approach, enabling clinical and functional OEF imaging including cortical GM OEF mapping in 3.4 minutes and generation of an OEF time course for functional MRI. Future experiments will use tQUIXOTIC to quantify OEF changes during different challenges (e.g. gas inhalation and non-visual functional tasks), demonstrate CMRO₂ quantification with additional cerebral blood flow measurements, and evaluate OEF and CMRO₂ in pathology such as stroke and brain tumor.

Appendix

In the original QUIXOTIC technique (15), T₂ of blood is determined by fitting a monoexponetnial decay to the venular-blood weighted signal:

$$\mathrm{S}(\text{TE})=k{\mathrm{e}}^{-\frac{\text{TE}}{{\mathrm{T}}_2}}.$$ [A1]

However, due to diffusion weighting from the velocity selective module applied only for tag acquisitions, some CSF signal remains after control-tag subtraction (15,18). Thus, the T₂ of blood is overestimated by monoexponetnial fitting since the signal includes some portion of CSF with its much longer T₂.

One proposal for how to empirically correct this CSF signal contamination in QUIXOTIC (20), assumes that any QUIXOTIC signal remaining after a long TE (∼400 ms) originates from CSF alone, since the T₂ of blood is considerably shorter than the T₂ of CSF and will have decayed into the noise. This CSF signal can then be subtracted from the QUIXOTIC signal to compensate for the CSF contamination at short TEs where the blood signal predominates:

$$S_{\mathit{\text{corrected}}}(\mathit{\text{TE}})=S_{\mathit{\text{uncorrected}}}(\mathit{\text{TE}})-k_{\mathit{\text{CSF}}}{e}^{-\frac{\mathit{\text{TE}}}{T_{2,\mathit{\text{CSF}}}}}$$ [A2]

Since k_CSF is determined empirically by measuring the signal intensity from a second long TE image, this provides a simple empirical method to correct for CSF contamination in QUIXOTIC.

For tQUIXOTIC, we first attempted the approach described in (20), but increasing the number of refocusing pulses (from 6 to 32) at the nominal delta TE of 12.6 ms to reach TE = 400 ms was not possible due to SAR limits. We thus increased the ΔTE to acquire data at longer TEs, which utilized less refocusing pulses and was less SAR intensive.

We acquired calibration data from one subject to determine the appropriate CSF correction for a typical grey matter voxel and then applied this result to all our data acquired with the same resolution and acquisition parameters. The calibration data consisted of tQUIXOTIC acquisitions with different ΔTE values from one subject. ΔTE values were 12.6, 50, 60, 70, 80, 90 and 100 ms. Five echoes were acquired with each ΔTE, except for the standard three echoes for ΔTE = 12.6. The total scan duration was 317s per acquisition. A DIR image for GM-only imaging was collected after each acquisition using the parameters described previously.

Since the signal from CSF is purposefully suppressed by subtraction, we first determined the value of T_2,CSF within GM using the pre-subtraction images. Images were smoothed and averaged over all grey matter. Each acquisition with different ΔTE was analyzed separately. Those acquisitions with two or more average signal intensities with TE ≥ 500 ms were then fit assuming monoexponetnial decay (Equation A1), since all signal remaining at these long TEs should only be from CSF. These fits resulted in average T_2,CSF = 1.2 s, which is similar to that computed in (20).

The calibration data was then analyzed using the same methods described for the tQUIXOTIC experiments. In short, the data was pairwise subtracted, smoothed (6 mm Gaussian kernel), and averaged across time. The corresponding DIR-GM image for each acquisition was used as a mask to select GM voxels and the average GM signal intensity was calculated.

We fit Equation 2, with T_2,CSF = 1.2 s, to the average GM signal intensity and observed a large residual due to signal discontinuities when similar TEs were acquired with different ΔTE (Figure 6), suggesting that k_CSF and k_blood had some dependence on the inter-echo spacing (ΔTE) of the GRASE readout. We altered the Equation 2 model for the average signal intensity measured at each echo time (S(TE)) by introducing a linear dependence on ΔTE for both y-intercept terms:

Figure 6.

Average GM signal intensity of the tQUIXOTIC calibration data for different ΔTE fitted with 3-parameter (Equation 2) and 5-parameter (Equation A3) models with T_2,CSF = 1.2 s.

$$S(\mathit{\text{TE}},\Delta \mathit{\text{TE}})=(x_1(\Delta \mathit{\text{TE}})+x_2)e^{-\mathit{\text{TE}}/x_3}+(x_4(\Delta \mathit{\text{TE}})+x_5)e^{-\mathit{\text{TE}}/1.2}.$$ [A3]

This data-driven alteration of the model dramatically improved fitting with an F-statistic = 133, P < 5×10^-15. We suspect the variation in y-intercept with ΔTE is biophysical and may be due to exchange of water between voxel compartments or diffusion around high susceptibility veins in the CSF (8,57).

We then determined X_CSF, the magnetization fraction of diffusion attenuated CSF within a GM voxel, which is specific to the tQUIXOTIC parameters used in our experiments. Specifically, for ΔTE = 12.6 ms, X_CSF = k_CSF/(k_blood+k_CSF) = 0.112, where k_CSF = x₄(0.0126)+x₅ and k_blood = x₁(0.0126)+x₂ with x_i given in Table 3. The magnetization fraction (X_CSF) is valid for ΔTE = 12.6 - 100 ms, but X_CSF for ΔTEs > 0.5∙(T_2,blood) are likely not relevant for tQUIXOTIC in practice, as the blood signal will have significantly decayed at these longer echo times. Variation in X_CSF for different ΔTE is due to the same biophysical properties hypothesized to affect the y-intercepts in Equation A3. To be clear, X_CSF is not the CSF volume fraction. It is the fraction of diffusion attenuated CSF that is not eliminated by control-tag subtraction in comparison to the venular-blood weighted signal resulting from QUIXOTIC velocity selection (i.e. the magnetization fraction). As long as the parameters of the velocity selective modules (i.e. the gradient, gradient pulse separation, and gradient pulse duration) are the same as what we used here, we would expect the same relative contribution of CSF contamination so this result could be applied to other studies. However, the approach of (20) is subject specific when used to correct QUIXOTIC acquisitions.

Corrected T_2.blood values were then determined as described above, by fitting tQUIXOTIC data with the two-compartment model given in Equations 2 and 3.