GlucoCEST-based dynamic glucose enhanced (DGE) MRI at 3 T: Early experience in healthy volunteers and brain tumor patients

Abstract

Purpose:

DGE MRI has shown potential for imaging glucose delivery and blood brain barrier permeability at fields of 7T and higher. Here we evaluated issues involved with translating D-glucose weighted Chemical Exchange Saturation Transfer (glucoCEST) experiments to the clinical field strength of 3 T.

Methods:

Exchange rates of the different hydroxyl proton pools and the field-dependent T₂ relaxivity of water in D-glucose solution were used to simulate the water saturation spectra (Z-spectra) and DGE signal differences as a function of static field strength B_0, RF field strength B₁, and saturation time t_sat. Multi-slice DGE experiments were performed at 3 T on five healthy volunteers and three glioma patients.

Results:

Simulations showed that DGE signal decrease with B₀, due to decreased contributions of glucoCEST and transverse relaxivity, as well as the coalescence of the hydroxyl and water proton signals in the Z-spectrum. At 3 T, due to this coalescence and increased interference of direct water saturation and magnetization transfer contrast, the DGE effect can be assessed over a broad range of saturation frequencies. Multi-slice DGE experiments were performed in vivo using a B₁ of 1.6 μT and a t_sat of 1 s, leading to a small glucoCEST DGE effect at an offset frequency of 2 ppm from the water resonance. Motion correction was essential to detect DGE effects reliably.

Conclusion:

Multi-slice glucoCEST-based DGE experiments can be performed at 3 T with sufficient temporal resolution. However, the effects are small and prone to motion influence. Therefore motion correction should be used when performing DGE experiments at clinical field strengths.

Introduction

Recently, promising results have been published for the use of D-Glucose (1,2) and glucose analogues (3-7), as biodegradable contrast agents for MRI. Following infusion of D-glucose, dynamic glucose enhanced (DGE) imaging has allowed the detection of MRI signal changes in animals (8) and in humans (9-13), showing potential for the use of this readily available sugar to study tissue perfusion parameters such as blood volume and blood brain barrier (BBB) permeability. These DGE experiments have used either the Chemical Exchange Saturation Transfer (CEST) approach (14-18) called glucoCEST (1,2) or T_1ρ based MRI signal changes (11,12,19-22). To date glucoCEST work has been limited to higher field strengths (7T and up) and single-slice acquisitions. While a T_1ρ -based DGE approach has recently been demonstrated at 3 T (23), our preliminary work has been reported in abstracts (24-26). For clinical relevance it is necessary to develop DGE MRI at lower field strengths and to extend the protocol to multi-slice acquisition. The goal of this study is to evaluate the physical principles involved in moving to lower field and, based on this knowledge, design experiments and perform multi-slice DGE experiments on a clinical scanner at the field strength of 3 T.

In CEST experiments, radio-frequency (RF) saturation as a function of chemical shift offset is used to measure signal changes in the water saturation spectrum (the so-called Z-spectrum). The CEST effects are then judged at the resonance frequency of certain exchangeable proton groups. However, it is important to realize that the -OH protons in sugars, such as glucose, exchange very fast (exchange rates of 2000 Hz or larger, depending on pH, buffer concentration and temperature (22)), leading to coalescence of the -OH proton resonances with those from water, especially in vivo and at clinically relevant magnetic field strengths such as 3 T. This coalescence removes the frequency specificity for the hydroxyl protons, but fortunately still causes an asymmetry relative to the water resonance frequency in the Z-spectrum. Actually, when performing asymmetry analysis, there is the appearance of a distinct resonance for the combined -OH groups, but the intensity is reduced due to the presence of -OH signal at frequencies on the opposite site of the water resonance, an effect that becomes stronger at lower field (18). Furthermore, due to the spillover effect of direct water saturation and, in vivo, the competition with the large semisolid magnetization transfer contrast (MTC) effect, the apparent peak maximum of this residual coalesced -OH signal shifts to higher frequency with increasing B₁. In addition, it has to be realized that the water signal relaxation time T₂ (27,28) is also affected by the presence of the -OH groups in glucose. This relaxation effect on the water lineshape is symmetric, while the glucoCEST effect is asymmetric. An asymmetry analysis will remove the T₂ relaxation effect to a great extent, which may not be desirable when the goal is to maximize signal changes in a DGE experiment. Fortunately, since a baseline image can be acquired before glucose infusion, only a single irradiation frequency is in principle needed in a DGE experiment, allowing one to include both glucoCEST and T₂ effects. Thus, increased exchange-based relaxation effects contribute synergistically to the saturation-based signal changes when the glucose concentration is increased, as in DGE experiments.

Here, in order to establish initial guidelines for DGE at 3 T, we used the Bloch-McConnell equations to simulate the contributions of glucoCEST, direct saturation, MTC and transverse relaxation to the coalesced line shape in the Z-spectrum. The simulations provide insight into how the DGE difference spectrum after D-glucose infusion changes as a function of magnetic field strength B₀, radiofrequency field strength B₁, saturation time (t_sat), and irradiation frequency offset (Δω). We then tested these parameters for DGE experiments at 3 T in a phantom and in healthy volunteers and patients with high grade malignant glioma.

Methods

Determination of exchange rates for glucose hydroxyl groups:

Glucose was dissolved in phosphate buffered saline (PBS, 10 mM H₂PO₄⁻/HPO₄⁻², pH = 7.2) at a concentration of 10 mM and placed in an NMR tube. In vitro experiments were performed at 17.6T on a vertical bore scanner (Bruker, Ettlingen, Germany) at 37°C using a 5 mm Broadband inverse (BBI) high resolution probe. Continuous wave (CW) CEST signal intensity (S_sat) data (t_sat = 7 s; B₁ = 1, 3, 5 μT) were acquired as a function of frequency (steps of 0.2 ppm, ranging from 6 to −6 ppm with respect to the water proton frequency) and a reference acquisition without saturation was acquired to determine S₀. Glucose hydroxyl proton exchange rates were quantified by globally fitting the entire Z-spectra at multiple saturation pulse powers (QUESP (29)) using the Bloch-McConnell equations in Python using custom written scripts.

Simulations:

Simulations of Z-spectra were performed using a custom written script in Python by numerically solving the Bloch-McConnell equations. It was assumed that four types of protons exchange with bulk water, corresponding to chemical exchange of D-glucose hydroxyl protons at 1.2, 2.2 and 2.8 ppm and MTC originating from protons on semisolid macromolecules. Three different magnetic field strengths (3 T, 7 T and 11.7 T) and four tissue compartments (arteries, veins, gray matter, and white matter) were simulated. The parameters used in the simulations are summarized in Table 1. To simulate the effect of glucose infusion, we assumed the concentration of blood glucose was changed from a baseline of 5 mM (90 mg/dL) to a typical level of 15 mM (270 mg/dL) after infusion of 50 mL, 50 % dextrose, as based on blood sampling data reported in the literature (9,13,30). The exchange-based transverse relaxivity contribution from glucose at physiological pH and 37°C was estimated according to literature, namely 0.012 s⁻¹ mM⁻¹ at 3 T, 0.053 s⁻¹ mM⁻¹ at 7 T and 0.102 s⁻¹ mM⁻¹ at 11.7 T (28). The glucose concentration in the brain tissue compartments was assumed to be ¼ of the plasma concentration (31,32).

Table 1:

Baseline parameters used in the simulations for Figure 1.

component	conc (M)	# of 1H	Δω (ppm)	kex (s−1)	3.0 Tesla		7.0 Tesla		11.7 Tesla
					T1 (s)	T2 (s)	T1 (s)	T2 (s)	T1 (s)	T2 (s)
D-glucose	0.005
OH(1)	(Blood)	3	1.2	4,000
OH(2)	0.00125	1	2.2	8,000	1.2	0.1	1.2	0.1	1.2	0.1
OH(3)	(tissue) ‡	1	2.8	10,000
H2O	55.5	2	0.0
Arterial					1.86*	0.158*	2.27*	0.050*	2.72*	0.044*
Venous					1.70*	0.066*	2.06*	0.012*	2.40*	0.008*
GM					1.30@	0.090@	1.70@	0.055$	2.10#	0.037#
WM					0.800@	0.070@	1.22@	0.046$	1.80#	0.030#
MTC
GM	5.5&	1	−3.0	40&	0.500¶	9×10−6 &	0.500¶	9×10−6 &	0.500¶	9×10−6 &
WM	15.4&	1	−3.0	23&	0.500¶	10×10−6 &	0.500¶	10×10−6 &	0.500¶	10×10−6 &

Glucose Δω and k_ex are from the phantom experiments in the current paper; relaxation times were assumed.

^*Li et al. (43,44), Grgac et al. (45), Qin et al. (46);

^#de Graaf et al, rat brain (47);

^@Rooney et al (48), Lu et al.(49), Wansapura et al. (50). Notice that sometimes GM T₂ is too long in cortex (CSF mix-in), so we used deep GM values.

^$Yacoub/Rydhog, human brain (51);

^**Li et al. (43) , Table 2, saline values;

^‡(52);

^¶Stanisz et al. (53), Sled et al. (54), Henkelman et al. (55), and van Gelderen et al. (56)

The relaxation rate R₂ = 1/T₂ is a factor in the line width (LW) of the saturation water resonance (33,34) in the Z-spectrum:

$$LW(Hz)=(1∕\pi )\sqrt{\frac{R_1{R}_2^2+{\omega }_1^2{R}_2}{R_1}}\approx (1∕\pi )\sqrt{\frac{{\omega }_1^2{R}_2}{R_1}}$$ (1)

The approximation in Eq. (1) is allowed because ω₁ in radians (B₁ = 1 μT corresponds to ω₁ = 267.5 rad/s) is much larger than R₂, which in turn is an order of magnitude larger than R₁.

Phantom Study to simulate infusion:

Two phantoms were made using plastic tubes. Phantom-1 contained 20% w/v bovine serum albumin (BSA) dissolved in phosphate buffer saline (PBS) with pH adjusted to 7.3. Phantom-2 was identical to phantom-1 except that glucose was mixed in at a concentration of 10 mM. Then 25 μL/mL glutaraldehyde was added to both phantoms to induce cross-linking of the BSA. The phantoms were put in an ice bath for at least 12 hours to ensure adequate, uniform cross-linking. Prior to the experiments, the phantoms were immersed in a jar of water and the assembly was put in a water bath set to 37.2 °C. After an hour of equilibrating, the entire jar containing the phantoms was imaged as shown in Supporting Information Figure S1a. The imaging parameters were the same as in the in vivo experiments described below but only a single slice was acquired to speed up the acquisition. Full Z-spectra with four different B₁ levels were recorded within a total scan time of 10 minutes in which the temperature change was assumed to be minimal.

In Vivo Human Study

Five healthy volunteers and three postoperative patients previously diagnosed with malignant glioma (two oligodendroglioma grade III and one glioblastomas) were studied. The study was approved by the Institutional Review Board at the Johns Hopkins University School of Medicine and written informed consent was obtained prior to the study from all participants. The inclusion and exclusion criteria were the same as previously reported (9). Participants were asked to fast (8 hours, but clear liquids allowed) before the study and their baseline glucose levels were checked to be within the normal range (3.9 - 7 mM). The intravenous infusion lines were then set up and the participants positioned in the scanner. During the dynamic glucose scan, a brief hyperglycemic state was established by intravenous infusion of hospital grade D50 glucose, (D50, 25 g dextrose in 50 mL water sterile solution, Hospira, Lake Forest, IL, USA) followed by 20 mL saline solution in one arm. The glucose infusion was performed using a power injector at infusion rate of 0.8 mL/s, 0.4 mL/s or 0.3 mL/s, corresponding to total infusion times of 63 s, 125 s or 167 s, respectively. The venous glucose level was sampled at baseline and 20 or 30 min post glucose infusion. Venous glucose concentration was measured immediately upon collection by a blood analyzer (Radiometer).

Data Acquisition:

Subjects were scanned on a 3 T Philips MRI scanner (Elition R5.4, Philips Healthcare, NL). A 32-channel head coil was used for signal reception and a body-coil with parallel transmission sub-system was used for RF transmission. Acquisition software was modified to operate the two RF amplifiers of the system in an alternating fashion during the RF saturation as described by Keupp et al. (35). Therefore, each amplifier can be operated at full power while maintaining 50 % duty cycle. The quasi-continuous wave saturation consisted of a train of sinc-gauss pulses, each 50 ms in duration. Images were acquired using a multi-shot 3D turbo spin echo (TSE) sequence with TR/TE/FA of 3.5 s/6.1 ms/90°; 15 slices of thickness 4.4 mm each with in-plane resolution of 3.3×3.3 mm² were acquired within a FOV of 180×220 mm². The TSE factor was 80 and the k-space data were collected in a centric, radial order. The SENSE acceleration factor was 1.8 (Anterior-Posterior direction). The scan time for acquiring each image volume was 17.5 s.

Z-spectra were acquired before and after glucose infusion scan using a saturation B₁ field of 1.6 μT for 1 s. For the 5^th volunteer the saturation parameters, B₁ of 1.0 μT for 1 s at an offset of 1.5 ppm, were used, with a temporal resolution of 14 s. For the 3^rd patient, the saturation parameters were the same as for the other healthy volunteers and tumor patients, except that the dynamic scan time was 14 s. For Z-spectra, the saturation frequency was varied from −5 to 5 ppm with an increment of 0.4 ppm. For DGE-MRI, dynamic images were acquired at a single frequency of 2 ppm. Images with no saturation (S₀ images) were also collected (N = 4) prior to glucose infusion. The length of the dynamic scans was 21 min. Glucose was administrated during the dynamic scan starting at 3 min. Anatomical images such as MPRAGE (1 mm³ isotropic resolution) and FLAIR (2.2 mm³) images were acquired prior to glucose infusion.

Approximately 35 min post glucose infusion, Gadoteridol (ProHance, Bracco Diagnostics Inc) was given at a dosage of 0.1 mmol/kg via a power injector. Post contrast MPRAGE image (1 mm³ isotropic resolution) was acquired after the Gd contrast injection.

Data Analysis:

Motion Correction.

When performing dynamic difference experiments at a single frequency, motion can induce small signal changes, as was recently demonstrated by Zaiss et al (36). To correct for possible motion, the dynamic images were re-aligned using SPM12 (37). The first images acquired without saturation were removed from the series and the 2^nd image with saturation was chosen as a reference image. All other saturation images were aligned to the reference image using rigid body transformation. The first S₀ image without saturation was discarded and the rest of the S₀ images were aligned to the last S₀ image acquired immediately before the application of the saturation pulse. Alignment was done by applying a 4 mm (FWHM) Gaussian smoothing kernel to the images and the interpolation/re-slicing was done with 4^th degree B-spline. The realignment process was repeated twice to refine the motion correction results.

Co-registration and Segmentation.

The MPRAGE and FLAIR images were co-registered to the DGE images using SPM12. For the healthy volunteers, automated image segmentation for gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) was performed over the entire imaging volume. For patients with glioma, Gd-enhanced and FLAIR ROIs were drawn manually by selecting the voxels of the entire hyperintense area within the tumor volume (i.e. over multiple slices) of these particular images and the DGE intensity in the corresponding voxels of the co-registered images analyzed for the curves. The white matter area was drawn in the area contralateral to the tumor and analyzed similarly. Since this is not a clinical study, but a first presentation of measured effects at 3T, an inter-rater consistency study was not performed. Instead, tumor segmentation was done with the simple guideline of selecting all hyperintense voxels within the tumor for the FLAIR or Gd enhanced regions.

Calculating AUC images.

For the DGE dynamic series, the first S₀ image was discarded to assure proper steady state and the rest of the S₀ images (N = 3) were averaged. All other images were normalized to this average, pre-infusion S₀ image. The first image with saturation was also discarded. Then, a baseline image with intensity S_base was generated by averaging the rest of the pre-infusion saturated images (N = 5). The DGE difference images were calculated by taking the difference between S_base and each dynamic image S(t_n). Normalized area-under-curve (AUC) images were calculated using:

$$AUC=\frac{1}{n}{\sum }_1^n\frac{S_{base}}{S_0}-\frac{S(t_n)}{S_0}=\frac{1}{n}{\sum }_1^n\frac{\Delta S(t_n)}{S_0}$$ (2)

For the 4 healthy volunteers, the mean AUC values of the three tissue compartments (GM, WM, CSF ) and two different time periods were compared by paired student’s t-test (p < 0.05).

Results

Simulations:

Figure 1 shows simulations of DGE difference spectra at 3 T, 7 T, and 11.7 T with different saturation powers for t_sat = 1 s. The DGE changes reflect a concentration increase of 10 mM D-glucose in blood and of 2.5 mM in tissue. It can be seen that the D-glucose signal change decreases strongly when the static field changes from 11.7 T to 3 T and the resulting signal becomes broadened on the ppm frequency scale. Interestingly, while higher B₁ power results in more glucose signal difference at 11.7 T in the arteries, this is no longer the case at lower fields, where the maximum difference remains about the same but the peak position moves to higher frequency offsets from water due to the increased competition of direct water saturation closer to water. In the veins, the DGE signal difference reduces from 11.7 T to 7 T, but not from 7 T to 3 T, with minimal apparent B₁ dependence at each field for signal intensity and again an apparent shift to higher frequency offset with lowering field. In the gray matter, the DGE signal difference reduces with B₁ at each field strength, but at 3 T the difference between 1.0 μT and 2 μT is minimal when looking at a frequency of about 2 ppm (see inset with enlarged effects). In white matter, due to competition of MTC and direct saturation, the DGE difference signal is less than 0.5 % at 11.7 T and one third of that at 3 T.

Figure 1.

Simulations of the DGE difference Z-spectrum as a function of static field strength (11.7 T, 7.0 T and 3.0 T) in different tissue compartments. Four different RF powers from 0.4 to 2.0 μT are simulated for a saturation time of 1 s. The simulation assumes that blood glucose concentration increase from 5 mM to 15 mM after infusion and the glucose in other tissue compartments is ¼ of the blood concentration. The plots for the gray and white matter are enlarged in the black box below. Parameters are listed in Table 1.

Figure 2 shows the simulated glucoCEST signal at 2 ppm as a function of saturation time and power for the four different tissue compartments at 3 T. We simulated 31 B₁ values between 0.5 to 2.0 μT and 31 saturation times from 0.5 to 2.0 s. While in general, the glucoCEST signal increases with increasing saturation power, this is no longer the case at longer saturation times where the DGE signal difference reaches a maximum and decays with further increases in t_sat. When choosing appropriate parameters for in vivo, it has to be kept in mind that we want contrast between healthy brain tissue (GM and WM) and the tumors that are characterized by compromised BBB and an increase in blood volume. We do not know the water relaxation times in the interstitial space, but these are probably somewhat longer than intracellular. Examining the signal changes from different tissue compartments simulated in Figure 2, we decided t_sat = 1 s and B₁ = 1.6 μT to be appropriate parameter choices for a frequency offset of 2 ppm.

Figure 2.

Simulated glucoCEST signal at 2 ppm as a function of saturation time and power in different tissue compartments at 3 T. Parameters are listed in Table 1.

Phantom Studies:

In order to check the simulations, we prepared some phantoms to mimic the in vivo conditions, namely tubes with cross-linked BSA (phantom-1) to reflect tissue and cross-linked BSA plus 10 mM glucose (phantom-2) to represent tissue after infusion. The Supporting Information Figure S1b shows the difference spectra calculated by subtracting the Z-spectra of phantom-2 from those of phantom-1. Qualitatively the shapes of the experimental difference spectra agree well with the simulations in Figure 1, with maxima at both positive and negative frequency. At 2 ppm, the 10 mM glucose difference signals using t_sat = 1 s and B₁ saturation powers of 0.4, 0.8, 1.2, 1.6 μT were 0.8 %, 1.9 %, 2.6 % and 2.6 % respectively.

Human Studies:

Figure 3a shows the motion corrected dynamic difference images in one of 15 slices before, during and after glucose infusion in a healthy volunteer. Figure 3b shows the DGE signal change as a function of infusion time in whole brain, GM, WM, and CSF. In this participant, the tissue signal reduction was transient and disappeared after about 800 s. Figures 3c and 3d show the mean AUC images for ten slices calculated between 0-2 min and 2-7 min since the start of the infusion.

Figure 3.

Illustration of a DGE study in a healthy volunteer. (a) The dynamic difference images; (b) the DGE signal as a function of infusion time in the brain as a whole, gray matter, white matter and CSF; (c) the mean AUC during glucose infusion (0-2 min); and (d) the mean AUC 5 min post glucose infusion (2-7 min) of a healthy volunteer.

The mean AUC_0-2min and AUC_2-7min images for three other healthy volunteers are shown in the Supporting Information Figures S2-4a. The DGE signal changes as a function of infusion time in whole brain, GM, WM and CSF are shown in Figures S2-4b. The mean AUC values in each of these tissue compartments are presented in Figure 4. When looking at these healthy volunteers (n = 4), we found that, in the initial 2 min post infusion, the mean AUC change was not significant in GM and WM, while the CSF was slightly elevated (p = 0.012). In the following 5 min, the signal in GM and WM decreased significantly compared to that in the initial 2 min (p = 0.015 and p = 0.017, respectively). During this period, GM signal was also decreased significantly when compared to the pre-infusion baseline (p = 0.018), while the reduction (Fig. 4) compared to this baseline was not significant for WM. The CSF signal was not significantly different compared to baseline (p = 0.26) nor the initial 2 min (p = 0.065). The dynamic images of one slice and the AUC during infusion and 5 minutes post infusion for the healthy volunteer scanned with different saturation parameters and faster infusion time are shown in the Supporting Information Figure S5. The time dependent dynamic trends show a similar decrease in the signal in all tissue compartments.

Figure 4.

Mean AUC_0-2min and AUC_2-7min in the WM, GM, and CSF tissue compartments. One-sample t-tests were performed for comparing each mean AUC to the average baseline (0 by definition). Paired t-test were performed when comparing AUC_0-2min and AUC_2-7min. * denotes a p-value < 0.05, for the t-tests.

Because motion related artifacts tend to be more pronounced at tissue interfaces, motion can be a serious issue when imaging lesions. Patient 1 was previously diagnosed with an IDH mutant anaplastic oligodendroglioma (WHO Grade III), resected 18 months prior to the current imaging (XRT stopped 3 months prior to the scan) showing a stable heterogeneously enhancing infiltrative mass surrounding the right frontal lobe resection cavity. Figure 5b and S6 show the motion profiles as a function of infusion time before and after image realignment. It can be seen that even sub-voxel translational (1-1.5 mm) and rotational (1 degree) motion can lead to serious artifacts in the dynamic image (Figure 5a). The dynamic image series for slice 9 before motion correction is shown in Figure 5a, while this series for the adjacent slices (8 and 10) can be found in Supporting Information (Figures S7a and S8a). Areas in proximity to the ventricles, around the tumor, and at the GM/WM and GM/CSF interfaces were more severely affected by motion. After realignment, an additional motion correction procedure lead to a change of only about 1/100^th of a voxel, indicating no further correction needed and the dynamic images were improved (Figures 5c, Supporting Information Figures S7b, and S8b). However, an additional effect of more motion appears to occur after about 400 s, despite the minimal residual translation and rotation, and we therefore limited the AUC determination to two periods only, 0-2 min and 2-7 min. Figure 6a compares the post-Gd T₁-MPRAGE, T₂-FLAIR, mean DGE AUC_0-2min, and mean DGE AUC_2-7min images of this same patient for 10 slices. We can observe that the DGE enhancement is minimal for AUC_0-2min and more pronounced in the AUC_2-7min, but mainly located in the areas surrounding the surgical cavity, partially overlapping with areas of Gd enhancement. When plotting the DGE signal change as a function of infusion time (Figure 6b), it can be seen that the enhancement peaks around 1 % in the Gd enhanced areas, approximately 0.5 % in the FLAIR hyperintense areas, but that it was negligible in the posterior unaffected white matter and the brain as a whole. The mean AUC_0-2min and AUC_2-7min in the Gd-enhanced, FLAIR hyperintense and white matter regions are summarized in Table 2. We also isolated a few voxels in the blood vessels to assess the arterial input function (AIF) of the glucose infusion (Supporting Information Figure S9).

Figure 5.

DGE MRI of a patient previously diagnosed with an IDH mutant anaplastic oligodendroglioma (WHO Grade III) at the time point of 18 months post-surgery (Patient 1), (a) without motion correction. (b) shows the translation and rotation motion correction profile leading to the dynamic images in (c).

Figure 6.

Multi-contrast images for Patient 1: (a) the post Gd T1 MPRAGE, T2 FLAIR, mean DGE AUC_0-2min image, mean DGE AUC_2-7min image; and (b) the DGE signal as a function of infusion time in the Gd-enhanced, FLAIR hyperintense and posterior normal appearing white matter.

Table 2:

DGE-based mean AUC in brain tumor patients

	Patient 1	Patient 2	Patient 3
FLAIR hyperintense region, AUC0-2 min, %	0.20	0.26	0.80
AUC2-7 min, %	0.33	0.86	0.53
Gd enhanced region, AUC0-2 min, %	0.16	0.28	0.88
AUC2-7 min, %	0.82	0.97	0.52
white matter, AUC0-2 min, %	0.037	−0.43	−0.066
white matter, AUC2-7 min, %	0.095	−0.057	−1.81

Figure 7 shows DGE results for a patient previously diagnosed with IDH mutant glioblastoma, resected 3 months prior to this imaging. The motion profiles as a function of infusion time before (Figure 7b) and after (Supporting Information Figure S10) image realignment and the resulting dynamic difference images for slice 6 before (Figure 7a) and after (Figure 7c) motion correction are shown. The adjacent slices (7 and 8) are in Supporting Information Figures S11 and S12. The dynamic difference images (Figure 7c) show some fast enhancement that later reduces. This is followed by a large change, again due to motion as judged from the motion profile after correction (Supporting Information Fig. S10). We therefore again limited the AUC analysis to the initial uptake periods. Figure 8a shows post-Gd T1-MPRAGE, T2-FLAIR, mean DGE AUC_0-2min, and mean DGE AUC_{2-7 min} images of this patient. There are a few nodules with Gd enhancement and a ring enhancing lesion based on the post contrast MPRAGE images. T2-FLAIR shows extensive regions of hyperintensity. Radiological impression was consistent with likely progression and this patient went for resection soon after the scan, resulting in diagnosis of very limited residual glioma with extensive necrosis. The DGE AUC_0-2min images show slight scattered enhancement in the frontal region and this becomes more prominent in the DGE AUC_2-7min images. Several regions within the tumor showed DGE enhancement. Note that these enhanced regions do not particularly need to coincide with the Gd enhanced region. The right cortical area posterior to the tumor also shows some delayed signal enhancement with glucose. Figure 8b shows the DGE signal as a function of infusion time in the Gd-enhanced, FLAIR hyperintense, and the left posterior unaffected white matter regions. Around 1-1.5 % DGE enhancement is seen in the Gd-enhanced and FLAIR hyperintense regions. The signal change in the unaffected white matter was negligible after the glucose infusion was completed. The mean AUC_0-2min and AUC_2-7min in the Gd-enhanced, FLAIR hyperintense and the white matter regions are summarized in Table 2. We also assessed the AIF for this patient (Figure S13).

Figure 7.

DGE MRI of a patient previously diagnosed with an IDH mutant glioblastoma at the time point of 3 months post-surgery (Patient 2). The dynamic difference images before (a) and after (c) motion correction. b) The translation and rotation profile before the motion correction. The color scale for this patient differs from the one from Patient 1.

Figure 8.

Multi-contrast images for Patient 2: (a) the post Gd T1 MPRAGE, T2 FLAIR, mean DGE AUC_0-2min image, mean DGE AUC_2-7min image; and (b) the DGE signal as a function of infusion time in the Gd-enhanced, FLAIR hyperintense and posterior normal appearing white matter. The color scale for this patient differs from the one from Patient 1.

The analysis for a third patient, diagnosed with oligodendroglioma resected 2 months prior to the scan, is shown in Supporting Information Figure S14. The T2-FLAIR enhancement in S14a shows large hyperintense regions due to edema in all 10 slices and the post-Gd MPRAGE images have smaller but clear enhancement areas. In this case, we did not observe DG enhancement in the corresponding FLAIR and Gd-enhanced areas up to the 5 min post infusion, but it gradually build up after that. This enhanced signal persisted throughout the length of the dynamic scan, while the signal change in the unaffected white matter decreased post glucose infusion. Supporting Information Figure S15 shows the effects due to motion and DGE images before (Figure S15a) and after motion correction (Figure S15c) for slice 5. Supporting Information Figure S16 shows the motion profile after correction. The adjacent slices (4 and 6) are in Supporting Information Figures S17 and S18.

Discussion

Simulations:

In order to assess the feasibility of glucoCEST-based DGE experiments at 3 T, we performed simulations, phantom experiments, and in vivo experiments in healthy controls and patients with malignant glioma. These simulation results (Figures 1, 2) showed that the expected signal differences at 3 T are very small (on the order of 1 % or less), which was confirmed by the in vivo results. It is furthermore clear that, due to the fast exchange rate, the individual -OH resonances are not visible at physiological pH at any field and the composite -OH resonance is so broad that apparent maxima are seen on both sides of the water frequency in these DGE difference spectra. There is an asymmetry of these maxima between positive and negative frequency, which is expected based on the resonance frequencies of the -OH groups being at positive offset (Table 1). These maxima in the spectral difference shift further away from the water resonance with increasing power of the saturation pulse, due to the interference of direct water saturation in all tissues and additional interference from MTC in gray and white matter. While, at 11.7 T, this shifting increases the signal difference at positive offsets and reduces it at negative offsets (thus increasing the asymmetry), this is no longer the case at lower field due to the much broader line in ppm units and the faster exchange regime. In addition, the offset from water is much less in frequency units and thus more affected by direct saturation and MTC. These combined effects reduce the signal that can be detected at a certain frequency but on the other hand make the detection more flexible with respect to the choice of offset because of flattening of the difference spectrum. Based on the simulations, when performing glucoCEST-based DGE at 3T, we concluded that it is desirable to detect glucoCEST signal at an offset of around 2 ppm from the water resonance. Simulating such signal changes at 2 ppm as a function of saturation time and power, it became clear that the signal reaches its maximum at different saturation time/power combinations for different tissue compartments (Figure 2). Based on the simulations and the need to optimize for tumor contrast with respect to normal gray and white matter based on (i) leakage of the agent into the extracellular extravascular space (EES) and (ii) increased blood volume due to angiogenesis, we chose saturation power of 1.6 μT and saturation time of 1.0 s for human studies on the clinical scanner.

Phantom Studies:

To mimic tissue, we designed phantoms with cross-linked BSA, which consists of a large solidified network of proteins that generates a strong MTC effect similar to the situation in vivo. The phantom with only cross-linked BSA served as an analog of background brain signal and the cross-linked BSA+Glc (10 mM) phantom imitated the effect of glucose infusion. The difference spectra then represent signal observed in the DGE experiments. The results in Figure S1b show that glucose contrast of approximately 2.6 %/10 mM could be observed using 1.2 or 1.6 μT B₁ saturation power at 2 ppm, which would correspond to 0.65 % in tissue where the concentration is about four time lower. This effect size is between that predicted by the simulations for veins and arteries, but somewhat larger than in gray and white matter. There could be several reasons for this: (1) The glucose hydroxyl exchange rate in a crowded environment such as BSA network is slower than measured in PBS solutions; (2) The MTC exchange rate in the BSA phantom is less than that assumed in the simulations, reducing spillover effects and thus increasing the glucose signal. Overall, the Z-spectral differences look qualitatively similar to the simulations.

Human Studies:

Compared to the single-slice DGE experiments previously reported at 7 T, the current 3D DGE protocol covers a large brain volume that facilitates better tumor/edema visualization and motion correction. After motion correction, the pre-infusion glucose difference images in the healthy volunteer (Figure 3) showed minimal intensity difference, indicating good signal stability and hence temporal SNR. Generally, the signal appeared to decrease slightly in GM and WM compartments upon glucose infusion (Figure 4). This observation was consistent and reproducible for the volunteers and the tumor patients and statistically significant when comparing the first 2 minutes post-infusion with the subsequent 5 minutes. We did not find such effects in the previous studies at 7 T and such differences between two field strengths probably do not have a physiological explanation but more likely are related to an MRI phenomenon that is not dominant at 7 T where the glucoCEST changes are much larger. Possible causes for the slight decrease may be a susceptibility-based frequency shift or a tissue signal change due to osmolarity differences between blood and tissue. However, we are unable to make a conclusive assessment for the reason behind this observation.

Subject motion was also a problem in the dynamic difference images of the brain tumor patients, leading to severe artifacts in the non-corrected dynamic glucose difference images, especially at tissue interfaces (Figures 5a, 7a). These artifacts may cause signal changes that are beyond the magnitude of the actual effect size (36). Therefore there is a risk that they will be mistaken for true glucose effects. Motion correction adjusted the DGE effect to the expected effect size from 0.5-1.5 %. In this study we noticed that motion started shortly after the glucose infusion. One possible intuitive interpretation is that this is due to the sensation of the infusion. For instance, several participants reported feeling a pulsating or warm feeling in the head and groin areas upon glucose infusion for the 1 min infusion studies. The sensation generally disappeared quickly (tens of seconds) but may cause (in)voluntarily motion. We recently switched to 2 min infusion to reduce this effect (used in our examples) but some participants still felt warm/cool sensations at the start of the infusion. Another issue may be that performing procedures (such as blood sampling through the IV line) may lead to the participant trying to look or involuntarily move (reflex) when the nurse moves nearby the magnet. Therefore, we limited the blood draws to three counts for tumor patients: i) before the start of the study, ii) at the end of DGE and post-infusion CEST (~20 min post glucose infusion) and iii) end of the study before releasing the participant. Recently we have demonstrated at the 7 T that AIFs can be measured using DGE and correlated with venous blood glucose levels measured by blood sampling. (13) At 3T we found it more difficult to find these AIFs. There may be several causes for this, for instance the reduced SNR at 3 T. Motion correction could also be a factor, since the algorithm detects the change in image contrast and may interpret this as motion and minimize it. In the current study, since we performed limited blood sampling, no correlation between AIF and the venous blood glucose level was calculated. Nevertheless, the AIFs shown in Supporting Information Figures S8 and S12 show a peak signal of 2.5-5 %, which is approximately 3-4 times of the peak signal observed in the enhancing areas of the tumor.

To interpret the glucose enhancement pattern in post-surgery patients with complicated pathology is difficult. In patient 1, the glucose enhancement was mainly located in the Gd enhanced area around the cystic region but in patient 2, the glucose enhancement was not restricted to the Gd enhanced region, but rather extended to the temporal region of the right hemisphere, partially overlapping with the FLAIR hyperintense. For patient 3, the FLAIR hyperintense signal overlapped and surrounded the nodular enhancement regions of the Gd-scans. The glucose enhancement appeared around both these regions. Patient 1 refused surgery so no further pathological results could be obtained post MRI study. For patient 2 and 3, the radiological anatomical images (without considering the DGE images) suggested progression of disease and the patient underwent a second surgical resection of the right frontal lesion and left frontal nodular enhancing areas around the previously resected cavity, respectively. For patient 2, pathology revealed limited glioma with extensive treatment-related necrosis. The Gd enhancement observed on the anatomical images was thus mainly due to treatment effect. The glucose enhancement did not coincide with Gd enhancement, however it was difficult to pinpoint which region within the tumor enhances with glucose due to insufficient image quality. Interestingly, the time dependent curves seem to have some extra information. For instance, in patient 1 there is minimal signal change in the first few minutes except around the cavity (Figure 5c). In patient 2 (Figure 7c) there is a fast increase in the Gd-enhanced region, so information similar to contrast-enhanced MRI is present. At later times leakage into other tumor areas appears, the significance of which we do not yet know. For patient 3, the motion corrected images are very stable throughout the acquisition (Figure S15c) and enhancement appears to start only around after 5 minute post-infusion, unlike the other two patients. Due to the complicated pathology and the limited number of cases studied in this initial pilot study, it is difficult to determine the exact pattern of enhancement with glucoCEST. Others also found that regions of DGE and Gd enhancement differed in area (10,12). In our previous 7 T study (9), we noticed in one patient case that early enhancement corresponded to the Gd-enhanced ring, while later enhancement spread through a larger tumor area not enhanced in the Gd-T1w scan. Such preliminary results suggest that DGE may have additional information beyond Gd imaging. However, more studies with primary, pre-surgical brain tumor cases are needed to shine light on the uptake and retention of glucose in tumors.

Technical considerations:

Any change in signal during the post infusion period could be a result of either the glucose infusion, motion that could not be corrected or erroneous motion correction, or even local field shifts (36). The current DGE method, while offering the advantage of removing the background signal and thus providing time-resolved signal change is very sensitive to motion artifacts, mainly because of the small effect size. To successfully perform DGE imaging at 3T, movement thus needs to be minimized. In order to achieve this, participants need to be immobilized as much as possible, the infusion duration should be increased and minimal procedures should be performed during the scan. Participants should also be informed about the procedure, including if the nurse will draw blood and the physiological effects of the glucose infusion to further minimize motion.

It has been reported that D-glucose can cause vessel dilatation and enlargement of the ventricles due to the rapid uptake of D-glucose in CSF (39). These physiological volumetric changes may affect the DGE subtraction images and appear as motion. This type of tissue boundary change could be erroneously compensated for by motion correction and more work is needed to assure that the correction procedure does not remove real effects due to infusion. In addition, recent work in animals at 3 T (40) has shown that glucose infusion cause a small field shift, perhaps a susceptibility effect analogous to that occurring during Gd infusion. Zaiss et al. recently showed that motion can cause such field shifts (36), but since animals are immobilized this is unlikely. Reversely, an actual field shift may be erroneously compensated for by motion correction.

In order to improve DGE acquisitions, several approaches are possible, which will need to be tested in future work. Most logical would be work to improve the effect size, an example of which was shown recently in animals at high field using a pulsed CEST approach (41). Similar to other approaches, such as diffusion imaging, the methodology would benefit from fast 3D acquisition and the use of navigator echoes. A limitation of the current study is that possible dynamic changes in the B₀ field (e.g. frequency shifts due to susceptibility changes) cannot be separated out from the effects measured. A possible improvement would be the use of double volumetric navigators before the saturation period, which may allow for both motion and B₀ correction, as recently suggested (42). While residual motion contributions can still occur in dynamic scans, such approaches no doubt would lead to improvements.

Conclusion

The results show that it is possible to perform multi-slice DGE experiments at the clinical field strength of 3 T with sufficient temporal resolution. However, due to the reduced CEST effect size compared to 7 T, minor motion can cause severe tissue interface artifacts in dynamic difference images at 3 T. Motion must thus be mitigated to accomplish a successful DGE experiment. In addition, approaches for increasing the CEST effect for acquiring fast 3D acquisition with navigators need to be implemented. While glucose enhancement in the Gd-enhanced tumor region could be observed in our very limited number of brain tumor cases, this is just an illustration and the clinical value of DGE remains to be further explored with much larger groups of patients.

Supplementary Material

Supp figS1-18

Figure S1. Phantom experiment to simulate tissue DGE effects in vivo. (a) phantom assembly: a large water jar containing cross-linked BSA (phantom-1: CL-BSA) and cross-linked BSA with glucose (Phantom-2: CL-BSA+GLC). The other two phantoms in the assembly labeled PBS solution and CL-BSA+Cr are not related to the current experiment. (b) Difference in Z-spectra: phantom-1- phantom-2.

Figure S2. (a) Mean AUC_{0-2 min} and AUC _2-7min and (b) the DGE signal as a function of infusion time in whole brain, gray matter, white matter and CSF of healthy control number 2.

Figure S3. (a) Mean AUC_{0-2 min} and AUC _2-7min and (b) the DGE signal as a function of infusion time in whole brain, gray matter, white matter and CSF of healthy control number 3.

Figure S4. (a) Mean AUC _{0-2 min} and AUC _2-7min and (b) the DGE signal as a function of infusion time in whole brain, gray matter, white matter and CSF of healthy control number 4.

Figure S5. Illustration of a DGE study in healthy volunteer 5. (a) The dynamic difference images; (b) the DGE signal as a function of infusion time in the brain as a whole, gray matter, white matter and CSF; (c) Mean AUC _{0-1 min} and (d)AUC _1-6min. This data set has different saturation parameters and infusion time than the other 4 volunteers.

Figure S6. The translational and rotational motion profile after the motion correction on the images from Figure 5 of Patient 1, showing minimal additional correction.

Figure S7. (a) DGE images of brain tumor Patient 1 without (a) and with (b) motion correction. The figure shows slice 8, which is one slice below the slice shown in main text Figure 5. Time resolution is 17.5 s per image volume.

Figure S8. (a) DGE images of brain tumor Patient 1 without (a) and with (b) motion correction. The figure shows slice 10, which is one slice above the slice shown in main text Figure 5. Time resolution is 17.5s per image volume.

Figure S9. (a) Voxel selection for arterial input function from slice 2 of patient 1. The selected voxels are indicated with yellow arrows in both DGE AUC and post Gd MPRAGE images. (b) DGE signal change in the vessel voxel as a function of infusion time.

Figure S10. The translational and rotational motion profile after the motion correction on the data from Patient 2 in Figure 7, showing minimal additional correction.

Figure S11. (a) DGE images of brain tumor Patient 2 without (a) and with (b) motion correction. The figure shows slice 5, which is one slice below the slice shown in main text Figure 7. Time resolution is 17.5 s per image volume.

Figure S12. (a) DGE images of brain tumor Patient 2 without (a) and with (b) motion correction. The figure shows slice 7, which is one slice above the slice shown in main text Figure 7. Time resolution is 17.5s per image volume.

Figure S13. (a) Voxel selection for arterial input function of patient 2. The selected voxels are indicated with yellow arrows in both DGE AUC and post Gd MPRAGE images of slice 2. (b) DGE signal change in the vessel voxel as a function of infusion time.

Figure S14. Multi-contrast images for Patient 3, diagnosed with oligodendroglioma (a) the post Gd T1 MPRAGE, T2 FLAIR, mean DGE AUC_0-2min image, mean DGE AUC_2-7min image; for 10 slices out of the 15 are shown here and (b) the DGE signal as a function of infusion time in the FLAIR hyperintense, Gd-enhanced, normal appearing white matter.

Figure S15. DGE MRI for Patient 3. The dynamic difference images before (a) and after (c) motion correction. The color scale for this patient differs from the one from Patient 1. (b) shows the motion correction profile. Time resolution is 14s per image volume.

Figure S16. The translational and rotational motion profile after the motion correction on the images from Figure S15 of Patient 3.

Figure S17. (a) DGE images of brain tumor Patient 3 without (a) and with (b) motion correction. The figure shows slice 4, which is one slice below the slice shown in Figure S15.

Figure S18. (a) DGE images of brain tumor Patient 3 without (a) and with (b) motion correction. The figure shows slice 6, which is one slice above the slice shown in Figure S15.