Imaging of blood flow using hyperpolarized [13C]Urea in preclinical cancer models




To demonstrate dynamic imaging of a diffusible perfusion tracer, hyperpolarized [13C]urea, for regional measurement of blood flow in preclinical cancer models.

Materials and Methods:

A pulse sequence using balanced steady state free precession (bSSFP) was developed, with progressively increasing flip angles for efficient sampling of the hyperpolarized magnetization. This allowed temporal and volumetric imaging of the [13C]urea signal. Regional signal dynamics were quantified for kidneys and liver, and estimates of relative blood flows were derived from the data. Detailed perfusion simulations were performed to validate the methodology.


Significant differences were observed in the signal patterns between normal and cancerous murine hepatic tissues. In particular, a 19% reduction in mean blood flow was observed in tumors, with 26% elevation in the tumor rim. The blood flow maps were also compared with metabolic imaging results with hyperpolarized [1-13C]pyruvate.


Regional assessment of perfusion is possible by imaging of hyperpolarized [13C]urea, which is significant for the imaging of cancer. J. Magn. Reson. Imaging 2011;33:692–697. © 2011 Wiley-Liss, Inc.

THE DYNAMIC NUCLEAR polarization (DNP) dissolution method has enabled >10,000-fold (relative to 3 Tesla [T] thermal polarization) liquid state signal enhancement of many 13C-labeled compounds in MR, introducing a new set of potential intravenous contrast agents for MRI (1–4). While most of the recent in vivo research of hyperpolarized 13C-labeled compounds has focused on probing metabolism using metabolically active substrates (5–8), initial studies have indicated that metabolically inactive agents such as [13C]urea may also hold great potential for angiography or perfusion imaging (9–11). The use of hyperpolarized tracers like [13C]urea for MR imaging of blood flow has several advantages over the use of conventional gadolinium (Gd) -containing (paramagnetic) compounds, including direct proportionality of signal to tracer concentration (subject to relaxation), inherently high contrast-to-noise ratio (CNR) due to absence of background signal, and the potential for combination with hyperpolarized [13C]pyruvate for combined perfusion and metabolic imaging (12). Urea quickly diffuses into interstitial space in most tissues, and is rapidly taken up by cells in several tissues through facilitated transport (13, 14). Urea also meets the key criteria of safety and efficacy, as an endogenous compound with normally high concentrations in vivo (typically 1–10 mM, and much higher in renal medullary interstitium), low toxicity, and neutral pH, as well as the key MR properties of sufficiently long T1 relaxation times and being highly polarized by the DNP method.

Tumors typically exhibit altered, heterogeneous patterns of blood flow due to abnormal neovascularization, and measurements of tumor blood flow are clinically significant (15). For example, in human trials, changes in tumor blood flow during chemotherapy, as measured by dynamic H215O positron emission tomography (PET), can predict pathologic response in locally advanced breast cancer (16).

In this study, we demonstrate dynamic volumetric imaging of hyperpolarized [13C]urea in vivo over a time period of 30 s, in 5 normal rats and 11 mice (5 normal, 6 tumor-bearing). A new rapid pulse sequence with highly efficient usage of the hyperpolarized magnetization was developed, using fully balanced steady state free precession (bSSFP) with progressively increasing flip angle over time. This sequence was designed to ameliorate the primary challenge associated with the imaging of hyperpolarized agents: the rapid loss of polarization due to T1 and, following radiofrequency (RF) excitation, T2 decay.

This technique allows direct imaging of the heterogeneous blood flow within tumors, differentiating well-perfused from poorly-perfused regions. These data are complementary to the metabolic information available from MR spectroscopic imaging (MRSI) of hyperpolarized [1-13C]pyruvate (5, 7, 8), particularly as [1-13C]lactate detection can be increased in tumors independently of poor perfusion (17). For comparison, 13C MRSI data were also acquired following injections of hyperpolarized [1-13C]pyruvate.


Samples were prepared by dissolving 99% 13C-urea (Sigma-Aldrich, St. Louis, MO) in glycerol (6.4 M), with the trityl radical OXO63 (23 mM) (Oxford Instruments, UK). For each experiment, a 105-mg sample was weighed and loaded into a HyperSense DNP polarizer (Oxford Instruments, UK), where it was cooled to 1.4 K and irradiated by microwaves (94.095 GHz and 25 mW) in a magnetic field of 3.35T for approximately 1 h, and subsequently rapidly dissolved in 4.5 mL of heated phosphate buffered saline (1× PBS). The solution was rapidly transported to the scanner bore and the injection was started within 30 s of dissolution. The estimated mean polarization at the time of imaging was 3% and the ex vivo T1 was measured to be 47 s at 3T, based on separate experiments where the hyperpolarized material was identically prepared but drawn into a syringe that was placed into the coil for nonlocalized, low flip angle spectroscopic acquisitions. The pH of the injected solution was ∼7.5. Rats and mice received tail vein injections of boluses of 2.4 mL and 350 μL, respectively, of urea solution (115 mM) over 12 s, followed by a 500-μL and 150-μL saline flush.

Sixteen animal subjects (five male Sprague Dawley rats, five normal mice, six liver tumor-bearing mice from a transgenic model (18, 19)) were scanned on a GE 3T full body human scanner, with separate dual-tuned 1H/13C transmit-receive quadrature RF coils for imaging of mice and rats (rat coil 1H channel linear only). The RF coil diameter was 5 cm, and the length was 8 cm. All animal studies were conducted in accordance with Institutional Animal Care and Utilization Committee approved protocols. The custom two-dimensional (2D) multi-slice bSSFP pulse sequence (Fig. 1) was used for efficient imaging of the hyperpolarized media. The pulse sequence was prescribed with the following parameters: matrix = 32 × 32, field of view (FOV) = 8 cm, frequency direction = right–left, fractional phase FOV = 0.75, 8 slices of 8 mm thickness, repetition time (TR) = 11.5 ms, echo time (TE) = 5.75 ms, RF pulse = 3.2 ms sinc. The spatial resolution was 2.5 mm × 2.5 mm × 8 mm. The imaging time was 2.2 s per stack. Initial and final θ/2 pulses were used at each slice for good signal stability and for recycling of transverse magnetization, respectively (11). Scanning commenced at the start of the injection, and the stack of eight slices was repeated every 6 s, for a total of six repetitions over 30 s. The transmit gain was calibrated before the hyperpolarized experiment by using the nonlocalized signal from a syringe containing enriched [13C]urea (8 M, doped with Magnevist 1% by volume) placed near the center of the RF coil.

Figure 1.

The bSSFP pulse sequence for dynamic imaging of hyperpolarized 13C media.

To maintain a nearly constant proportionality between signal and tracer concentration across multiple time points, successive image stacks were acquired according to a progressively increasing flip angle schedule that accounts for loss of polarization due to T1 and T2 relaxation. This proportionality is crucial to the ability to accurately measure tracer concentration–time curves. Applying a postprocessing correction for T1 decay is another approach to this problem but is suboptimal due to inefficient use of the magnetization. For the experimental conditions described here (polarization, relaxation times, spatiotemporal resolution of the sequence, etc.), any chosen constant flip angle caused the measured signal in some key regions, most notably the liver, to dip below a correctable SNR threshold before the late data could be obtained. The ramped flip angle approach extends the imaging window by allocating signal much more equally over the imaging time points (Fig. 2). The flip angles chosen were based on a simulation of the hyperpolarized SSFP signal, taken from Svensson et al (11), modified for dynamic acquisitions. In this model, the signal in a SSFP RF pulse train decays according to a weighted sum of E1 = exp(−TR/T1) and E2 = exp(−TR/T2). The signal from an initial longitudinal hyperpolarized magnetization Mz,0 following n RF pulses is:

equation image

In our model, Mz,0 was also updated for each successive time point, and the effect of the delay between time points was included. An initial flip angle of θ = 4° was empirically chosen because this flip angle produced sufficient signal at the first time point in testing. The flip angles for subsequent repetitions were chosen to match the mid-train signal levels. A roughly linear flip angle ramp was selected: 4° (i.e., 2°−4°−4°−4°−…−4°−2°) − 7°−13°−25°−50°−120°. The model was based on an approximate T1 relaxation time for urea of 15 s, somewhat shorter than the published value of ∼20 ± 2 s in vivo for 2.4T (urea T1 is expected to decrease with increasing field strength) (9), and an approximate T2 of 300 ms.

Figure 2.

Simulated hyperpolarized bSSFP signal as a function of pulse number (dashed, constant flip angle 12°; solid, progressive flip angle as described in text, progressive over acquisitions but uniform within each acquisition). Sharp signal drops in dotted line correspond to T1 decay during periods of delay between acquisitions. Sharp drop at end of solid line corresponds to larger T2 decay component at higher flip angles.

Signal dynamics were quantified for the kidneys and the liver. Peak signal levels for each voxel were measured. The rat data were not normalized. However, to permit comparison among normal and tumor mice despite variable polarization levels at the time of the experiment, signals from the mouse studies were normalized by the last value of solid state polarization attained by each sample during polarization, which was automatically measured and recorded by the polarizer.

Blood flow maps were derived from the dynamic urea data. The tissue tracer concentration Ct (which is directly proportional to the hyperpolarized signal) can be described as the product of the blood flow and a convolution of the arterial concentration Ca (or arterial input function, AIF) with the residue function R

equation image

(20). The model-independent linear algebraic approach was taken, generating the matrix equation

equation image

where N = 6 for six time points according to the experimental details, which was inverted to find F· R(t) by linear least squares with an added regularization term, which was required in order for the analysis to yield meaningful parameter values, and was empirically determined (from the simulation study described below). Vascular signal (abdominal segment of the descending aorta) from the same slice as the target organ for each time point served as the AIF. Following convention, the maximum of the flow-residue product was taken as the blood flow, in arbitrary units.

Hyperpolarized magnetization is destroyed by each excitation, and, therefore, it is important not to select too fine a temporal resolution and thereby expend the magnetization too rapidly, a consideration that does not apply to Gd-based studies. In contrast to a previous hyperpolarized study of cerebral perfusion (10), much longer mean transit times (MTTs) of this diffusible tracer outside of the brain (21) allow sufficient sampling of the concentration–time curve at the chosen temporal resolution (6 s) for computation of tissue blood flow. A detailed software simulation was undertaken to verify this point.

In this simulation, a 2D virtual tumor perfusion phantom (Fig. 3) was created. Intended to model the heterogeneous perfusion characteristics of a liver tumor nodule with increased peripheral vascular density, the phantom had a circular shape with radially increasing tissue blood flows (ranging over a factor of 5), and an exponential tracer residue function with radially decreasing MTTs (ranging from 20 to 60 s, based on literature values for urea in rat liver) (21). The arterial input function was generated from a gamma-variate function,

equation image

where t0 is the tracer arrival time. Because the SNR is generally high in the vasculature, it was possible to verify in separate experiments at high temporal resolution (2 s, single slice imaging of an abdominal region of the descending aorta) that this equation (parameterized by α = 3.3, β = 4.0) closely models the vascular signal following the injection procedure as described (see Fig. 3). Phantom concentration–time curves were generated at a temporal resolution of 0.1 s, and then undersampled (along with the AIF) according to the actual experimental details described above. Noise was added to the data to produce comparable SNR levels to the experimental images. More precisely, because magnitude processing was used in our study, the simulated concentration data were adjusted to take on a Rician distribution. The noise level was set such that the maximum concentration curve SNR matched the maximum SNR from the experimental images (∼50:1). Finally, importantly, the tracer arrival time t0 was varied randomly over a 3-s window to simulate experimental variation. The flow-residue products were then deconvolved from the subsampled data. The maximum of this product was taken as the blood flow, and the residue function was fit to an exponential to determine the MTT. This is the same analysis as was performed on the actual data. The simulation was repeated for a total of 500 repetitions, and the computed perfusion parameters were analyzed in comparison to their true values.

Figure 3.

Perfusion simulation data curves and virtual tumor phantom. Data curves: true measured AIF function (dip is due to saline flush), simulated gamma-variate AIF and resulting liver tissue concentration–time curve (Ct) after convolution with exponential residue function and addition of noise. Sampled time points indicated by dotted lines. Images: True phantom blood flows and MTTs with linear radial parameter variation over a range of likely values. [Color figure can be viewed in the online issue, which is available at]

Metabolic activity of the normal and tumor-bearing mice was studied by spectroscopic imaging of hyperpolarized [1-13C]pyruvate. The mice were injected with 350 μL 80 mM [1-13C]pyruvate in a TRIS buffer (pH ∼7.3). At 30 s following intravenous administration of [1-13C]pyruvate, high resolution 3D spectroscopic images (16 × 16 × 16, 2.5 mm × 2.5 mm × 5.4 mm, spectral bandwidth = 581 Hz, covering lactate to pyruvate at 3T) were acquired using previously described methods: echo planar spectroscopic imaging (EPSI) with adiabatic double spin echo preparation (22), and four-fold resolution enhancement using compressed sensing (18). The scan duration was 16 s.

Full sets of axial, sagittal, and coronal multi-slice T2-weighted fast spin echo (FSE) 1H images were also acquired for each animal (192 × 192, FOV = 10 cm, 2-mm slices, NEX = 6). All hyperpolarized imaging data were registered to these anatomic scans, using in-house software. Spectral peaks were integrated to form metabolite maps. Some maps were resized for display, using either zero-filling or cubic interpolation. Color image overlays were generated using NIH ImageJ (23).


Six imaging time points consisting of eight slices each were acquired over 30 s (Fig. 4). In all of the animal subjects, rapid distribution of the intravenous [13C]urea to the heart and the kidneys was observed by the second time point (6 s), as well as diffuse pulmonary signal. The [13C]urea signal in the heart and kidneys peaked at 12 s (i.e., coincident with the end of the 12-s bolus injection). The right and left chambers of the rat heart were distinguishable. Diffuse hepatic signal was observed in all animals within 18 s. By quantitative region of interest (ROI) analysis, the mean ratio of renal to hepatic blood flows (per unit respective tissue volume) among the rats was 3.45 ± 0.80, and the ratio of peak signals was 2.96 ± 0.65.

Figure 4.

Dynamic [13C]urea data. Left: Axial images of hyperpolarized [13C]urea over 30 s in a normal rat (6 s between frames, arranged left-to-right), overlaid on T2-weighted FSE images. Top half: Six time points for slice containing kidneys. Bright spot is vena cava/descending aorta. Bottom half: Six time points for heart and lungs. Right: Dynamic signal curves from representative kidney and liver voxels, and corresponding AIF. [Color figure can be viewed in the online issue, which is available at]

Differences in liver blood flows were observed in mice bearing liver tumors as compared with normal controls. While normal mouse liver exhibited a more uniform, medium level of blood flow, tumor tissue blood flow was slightly lower overall but with regions of increased signal at tumor margins. By ROI analysis, the mean blood flow per unit normal liver tissue volume among the normal mice was 1470 ± 720, while the mean blood flow per unit tumor volume was 1190 ± 258, and the mean blood flow in a manually defined 2.5-mm tumor rim was elevated by up to 26% over the mean full tumor signal. These findings are consistent with published descriptions of inhomogeneous blood flow in solid tumors (15). Tumor was defined by manually tracing ROIs on the FSE 1H images.

The perfusion simulations demonstrated that reasonably accurate measurements of blood flow are possible using the described experimental protocol and deconvolution procedure, for the range of relevant tissue blood flows and MTTs, and experimental SNR. The mean absolute error in blood flow over the 500 repetitions did not exceed 17.7% (Fig. 5). While all results were reasonably close, the error was largest at the center of the virtual tumor (i.e., for the slowest blood flow rate and lowest SNR) with overestimated flow, and slightly underestimated blood flow for the tumor periphery. On the other hand, MTTs were systematically underestimated for all regions by 45–89%, with errors due both to temporal undersampling (every 6 s) and limited data acquisition window (30 s). In summary, although the residue function did not produce an accurate measurement of MTT, its maximum could be taken as a good measure of true blood flow.

Figure 5.

Simulation results from the virtual tumor phantom, demonstrating feasibility of measurement of tissue blood flow by dynamic imaging of [13C]urea according to the experimental details from this study. Top row: True blood flow (left) and instance of blood flow measurement from a single trial (right). Bottom row: Absolute percent error mean (left) and SD (right) over 500 trials. [Color figure can be viewed in the online issue, which is available at]

The perfusion changes seen in liver tumors were associated with metabolic changes measured by MRSI of hyperpolarized [1-13C]pyruvate. The mean lactate-to-pyruvate ratio for liver tumor tissue was 3.17 ± 0.82 versus 1.21 ± 0.56 for normal liver tissue. Regions with low [13C]urea signal and high [1-13C]lactate are likely to represent poorly perfused regions of tumor with high metabolic activity. These regions are shown in pure red in Figure 6, while well perfused regions of tumor are shown in yellow (green plus red).

Figure 6.

Axial images of hyperpolarized [13C]urea (green, 24-s time point) and [1–13C]lactate (red), overlaid on T2-weighted FSE 1H images (grayscale) of mouse liver, combining results of multiple experiments. Yellow indicates overlap of [13C]urea and [1–13C]lactate. Top row: normal mouse. Rows 2&3: liver tumor mice. Same slice shown at left and right. (Mouse in row 2 is prone, while 1&3 are supine.)


In this project, we demonstrated dynamic imaging of hyperpolarized [13C]urea over a period of 30 s following intravenous injection allowing regional estimation of blood flow in animals. This was enabled by the development of a new bSSFP pulse sequence, with progressively increasing flip angles over time.

Estimated renal and hepatic blood flow values were calculated, using techniques validated in a simulation study. In practice, there is substantial variation in the values obtained between exams, as expected due to variation in 13C polarization, delays before injection, etc. Other potential sources of variation are the high permeability of urea across red blood cells (14), which could cause a variable effect on relaxation times in vivo according to hematocrit, as well as variable leakage time of urea into tissue, and T1 variation in different tissues. Nevertheless, the mean ratio of renal to hepatic blood flows and also peak signals (3.45 ± 0.80 by blood flow, 2.96 ± 0.65 by peak signals) among the rats was consistent with published “gold standard” measurements of blood flow obtained using radioactive microspheres (24) of 3.5 (with all measurements made per unit volume of respective organ tissues). Further validation is still needed for other organs and for quantitative measurement of a full set of hemodynamic parameters, which would likely require finer temporal resolution and/or a longer sampling time window.

In tumor models, the described methods provided additional information about tumor tissue perfusion to the metabolic information that can be obtained using metabolically active hyperpolarized substrates. In particular, a perfusion-metabolism mismatch, or high metabolism despite poor perfusion (measured by a combination of H215O- and FDG-PET), is thought to be an important prognostic variable, and has recently been shown to be a stronger predictor of poor survival than either metric alone in human pancreatic cancer (17). The methods developed through this study will allow exploration of this factor in preclinical cancer models using hyperpolarized carbon-13 MR. Furthermore, recently described methods of simultaneous co-polarization of multiple compounds by DNP (12), such as [1-13C]pyruvate and [13C]urea, may allow the assessment of tissue metabolic and perfusion status based on a single injection experiment.

In conclusion, regional assessment of perfusion is possible using hyperpolarized [13C]urea, which is highly significant in the imaging of cancer.


We gratefully acknowledge the assistance of Kristen Scott with animal experiments, as well as Peter Shin, Galen Reed, Ilwoo Park, Brian Hill, and Duan Xu. We also acknowledge Adam Kerr, Mark Van Criekinge, and Albert Chen for helpful discussions.