Development of a Symmetric EPI Framework for Clinical Translation of Rapid Dynamic Hyperpolarized ¹³C Imaging

Introduction

Advances in hyperpolarization (HP) have enabled real-time molecular imaging that was heretofore impossible with magnetic resonance (1). Using the dynamic nuclear polarization (DNP) technique, spins are exogenously polarized to four orders of magnitude greater than Boltzmann equilibrium. Metabolically active substrates, such as [1-¹³C]pyruvate (2), are transported into cells and undergo enzymatic conversion into downstream metabolites on the timescale of an HP experiment (~1-2 minutes). This technique is being translated into the clinic, primarily for cancer and cardiac metabolism studies. A Phase 1 clinical trial demonstrated both safety and the feasibility of hyperpolarized [1-¹³C]pyruvateMR with up-regulated conversion to [1-¹³C]lactate conversion observed in biopsy-proven prostate cancer patients (3).

This first clinical trial used a ¹³C spectroscopic data acquisition method, which even with an echo-planar spectroscopic readout, required an acquisition time of 8-12 seconds for 3D volumetric coverage of the prostate with a spatial resolution of 0.34cm³.This spectroscopic data demonstrated increased HP [1-¹³C]lactate to [1-¹³C]pyruvate ratios in the prostate cancer patients, but the acquisition method had the drawbacks of relatively long temporal resolution, preventing accurate dynamic measurements of metabolism, and the complex EPSI reconstruction was performed offline, not on the clinical scanner.A different approach for rapid volumetric HP ¹³C MRI acquisition is selective excitation of a single resonance followed by a single shot flyback EPI readout (4). This technique was developed and applied to study the dynamics of [1-¹³C]lactate in preclinical cancer models following injection of [1-¹³C]pyruvate with a 3.5s temporal resolution over a 100s time window (4). Although this study showed significant differences in [1-¹³C]lactate dynamics between early and late stage transgenic prostate cancers, without the detection of [1-¹³C]pyruvate signals the measurement of conversion rates could not be made. Utilizing the chemical shift differences between the lactate and pyruvate resonances, acquisitions of both compounds can be acquired simultaneously (5), but suffer from FOV restrictions and sub-optimal SNR.

A major challenge facing these HP acquisition techniques is the limited imaging window, an outcome of the inexorable decay to the Boltzmann (or thermal) equilibrium polarization. Rapid and RF efficient imaging sequences are therefore of paramount importance to maximize the signal-to-noise ratio (SNR) and resolution in ¹³C experiments. Using a singleband spectral-spatial RF pulse (4,6), individual metabolites can be excited and rapidly imaged with a single excitation. Due to the slow transverse decay of ¹³C-labeled metabolites (7), a low bandwidth/long duration readout can be used to maximize SNR efficiency. An echo-planar imaging (EPI) readout trajectory in particular is well-suited for hyperpolarized imaging because of its rapid, efficient sampling and relative robustness to off-resonance artifacts. Flyback EPI trajectories have been used (5,8) for rapid metabolic imaging due to their insensitivity to timing delays and ease of implementation, but this approach suffers from a reduced SNR efficiency because of the lack of encoding during the flyback gradient. The lower ¹³C gyromagnetic ratio also results in reduced scan efficiency and longer echo times (TE) that exacerbate EPI-related image artifacts. Translating this technique to clinical ¹³C imaging requires further improvements in scan efficiency in order to improve SNR and enable rapid and robust metabolic imaging in vivo.

Symmetric EPI is an appealing alternative to flyback EPI. By encoding k-space on both the positive and negative gradient lobes, the echo-spacing can be substantially reduced, resulting in improved SNR and less sensitivity to off-resonance. The SNR efficiency is also increased with symmetric EPI due to a higher sampling duty-cycle, and can be further increased by using ramp-sampling. The downside of this approach is that symmetric EPI is prone to Nyquist ghosting artifacts because of delays and inconsistencies between even and odd lines of k-space, which primarily arise from eddy currents and concomitant magnetic fields (9). These errors manifest as an offset between even and odd lines of k-space, resulting in misalignment between the even and odd echoes and resulting in image ghosts that are shifted by half the FOV if left uncorrected (10). These timing delays and inconsistencies can be estimated with a reference scan (11,12), and, by applying a line-by-line phase correction in x-k_y space, Nyquist ghost artifacts can be minimized. However, a reference scan will be difficult to directly acquire from the hyperpolarized substrate due to the transient magnetization and the narrow window in which to image metabolism.

The goal of this project was to develop and test a new specialized HP ¹³C EPI acquisition approach with high SNR efficiency and temporal resolution that could be readily translated into future human studies. To that end, we developed and implemented a symmetric, ramp-sampled, partial Fourier EPI sequence for efficient HP ¹³C imaging. We also developed a framework for acquiring a reference scan from ¹H data for application to hyperpolarized ¹³C data on a clinical 3T MR scanner. Our approach was designed to be robust and SNR-efficient, have reduced EPI off-resonance artifacts, and to eliminate EPI ghosting artifacts, making it well-suited for clinical metabolic imaging studies with hyperpolarized ¹³C agents.

Methods

All experiments were performed on a clinical 3T scanner (MR 750, GE Healthcare, WI) with a maximum gradient strength of 5 G/cm and a maximum slew-rate of 20 G/cm/ms. A ramp-sampled, symmetric EPI readout with singleband spectral-spatial (SPSP) excitation and partial Fourier reconstruction was developed for use on the clinical GE platform. A reference scan, acquired prior to imaging by disabling the phase-encoding gradients, was used to correct for phase errors that lead to Nyquist ghosts. As the sequence was built off of a product GE sequence, it leveraged the ability to use the reference scan for online reconstruction of ¹³C metabolite maps, and retained all of the requisite safety and performance features.

A custom built dual-tune ¹H/¹³C volume coil (13) was used for both the thermal phantom and hyperpolarized ¹³C experiments. For in vivo experiments, an enriched 8M ¹³C urea phantom was used for center frequency and B₁ calibration. The SPSP RF pulse used in all of the EPI acquisitions was designed to excite [1-¹³C]pyruvate and metabolites at 3T, with a passband FWHM of 120Hz and a stopband of 600Hz. The RF pulse was designed in Matlab (R2014b) using the SPSP RF pulse design toolbox (14). Further information and downloads can be found at http://rsl.stanford.edu/research/software.html or https://github.com/agentmess/hyperpolarized-mri-toolbox.

The SNR efficiency is a function of the duty cycle and k-space trajectory, and will increase from a flyback to a symmetric EPI readout. This gain in SNR efficiency, which is equivalent to the product of the duty cycle and trajectory efficiency, can be defined as (15,16):

For flyback EPI, T_gradient is the duration of the trapezoidal readout gradient plus the flyback gradient. For symmetric EPI, T_gradient is the duration of the trapezoidal readout gradient. T_readout is the time the analog-to-digital converter is acquiring data. For non ramp-sampled EPI, T_readout is equal to the plateau of the readout gradient, while for ramp-sampled EPI, T_readout is equal to T_gradient minus the duration of the phase encode blip. A_k is the area of the sampled k-space, and D_k is the local k-space sampling density. For partial Fourier, the theoretical SNR efficiency is reduced by 1.15 to account for the reduction in acquisition time for 75% k-space coverage. To quantify the SNR advantages of symmetric EPI, a ¹³C enriched phantom was used (44mm outer diameter (OD)), consisting of four cylindrical tubes with either 1M ¹³C formic acid (14.8mm OD), 1M [1-¹³C]lactate (11.9mm OD), 1M [1-¹³C]alanine (10.2mm OD) or 1M ¹³C bicarbonate (19.6mm OD) to simulate the expected number of metabolites and chemical shift in a HP ¹³C experiment. The RF pulse was centered on bicarbonate, and data were acquired with either a symmetric or flyback echo-planar trajectory, with a 1s TR, a FOV of 96 × 96mm, a matrix size of 32 × 32, and 60 averages, yielding a total 1 minute scan time. For the symmetric readout (Fig. 1) data were either fully encoded or acquired with a ramp-sampled, symmetric readout with partial Fourier reconstruction (75% k-space coverage in the k_y direction; Table 1). A reference scan was obtained directly on the ¹³C channel by disabling the phase-encoding gradients and was acquired prior to imaging using 60 averages. Flyback EPI was reconstructed with a 2D Fourier Transform, while for ramp-sampled symmetric EPI the reference scan phase correction was applied (17), then data were interpolated to a Cartesian grid and 2D Fourier Transformed to image-space using the Orchestra reconstruction toolbox (GE Healthcare, WI, USA). The SNR was measured by calculating the mean signal within the phantom and then dividing by the standard deviation of the signal from a noise ROI. No changes to the Orchestra EPI routines were required for our ¹³C EPI application, and the routines are fully automatic once the location of the reference scan and data are specified.

Unlike flyback EPI, symmetric EPI requires a reference scan to correct for timing delays and inconsistencies between even and odd lines of k-space. This can potentially be acquired on either the ¹H channel or directly on the ¹³C channel.Due to the lack of endogenous ¹³C, using the ¹³C channel requires either an external ¹³C enriched syringe or expending some hyperpolarized magnetization. To determine if a ¹H reference scan could provide a robust correction for ¹³C images, reference scans were acquired on either the ¹H channel or on the ¹³C channel from a natural abundance ethylene glycol phantom (outer diameter of 54mm, inner diameter of 24 mm) using the same trajectory (designed for a ¹³C acquisition). The singleband SPSP pulse described above was used to avoid excitation of the J-coupled ¹³C resonances. The matrix size and FOV were 32 × 32 and 72 × 72 mm, respectively, with partial Fourier (75% k-space coverage) in the phase-encode direction and ramp-sampling along the readout (0.492ms ramp duration, 0.380ms plateau duration, 1.364ms echo-spacing). A single 50 mm thick slice was excited, with 360 averages, a 0.5 s TR, and a 21.9 ms TE, with a total readout duration of 32.7 ms. Reference scans were applied to the ¹³C data retroactively using the Orchestra Toolbox, and image quality was compared by calculating the fractional difference between the phantom data reconstructed with the two different reference scans.

To explore the utility of the ¹H reference scan for in vivo HP ¹³C dynamic imaging, metabolites maps of pyruvate and lactate were acquired in the study of renal metabolism in a healthy rat. Metabolites were excited with a singleband SPSP RF pulse, with independent flip-angles designed to achieve the best estimates of perfusion and metabolic conversion rate parameters in a physiological model (18). A total of 25 timeframes for each metabolite were acquired over 50s, for a temporal resolution of 2s, with a FOV of 96 × 96mm and a matrix size of 32 × 32. Partial Fourier (75% k-space coverage) was used in the phase-encode direction along with ramp-sampling (0.452ms ramp duration, 0.240ms plateau duration, 1.144ms echo-spacing) to decrease the echo-time, yielding a TE of 20 ms with a 100ms TR per metabolite. Data acquisition started immediately prior to injection covering a 20mm axial slice centered on the kidneys. A reference scan was acquired on the ¹H channel using the ¹³C trajectory prior to injection of the HP substrate. For comparison, a separate reference scan was directly acquired on the ¹³C channel from the remaining hyperpolarized substrate magnetization following the final image. The efficacy of the ¹H reference scan was determined by calculating the structural similarity (SSIM) index (19) of [1-¹³C]pyruvate for all time frames, and by comparing the time course between data reconstructed with either the ¹H or ¹³C reference scan. For the SSIM index, the ¹³C images reconstructed with the ¹³C reference scan were considered the ground truth. SSIM is a measure of image quality and signal fidelity between two images, with an SSIM of 1 implying complete agreement between the two datasets. SSIM was chosen as the metric of comparison because it has been shown to be more sensitive to image shifts and overall changes in signal than mean square error (19,20).

Results

The results from the phantom study are shown in Table 1 and Supporting Figure 1. The substantial SNR increase between flyback and symmetric EPI is due to the higher duty cycle for symmetric EPI, which encodes k-space on both the positive and negative readout lobes and reduces the TE. Further gains were observed when using ramp-sampling and a partial Fourier acquisition, as this further reduces the echo-time and increases the readout duty cycle, mitigating signal loss due to T₂* decay.

The ¹H reference scan provides a potential alternative method that is more robust and has higher SNR than direct acquisition on the ¹³C channel. The reference scan in both cases was acquired using the ¹³C trajectory, resulting in the same echo-spacing and timing delays for both approaches. The ¹³C phantom images reconstructed without a reference scan show substantial ghosting (Fig. 2A). Qualitatively, the lack of Nyquist ghosting in the phantom images reconstructed with either the ¹H or ¹³C reference scan (Fig. 2B,C) demonstrates the effectiveness of this technique. Quantitatively, the mean signal difference between both reconstruction approaches was <2% for all phantom voxels, further indicating that a ¹H reference scan was sufficient to correct for Nyquist ghosting in ¹³C images.

This approach was further extended to the in vivo study of [1-¹³C]pyruvate renal metabolism in a healthy rat. Images reconstructed with either a ¹³C reference scan obtained from the residual hyperpolarization (Fig. 3A) or a ¹H reference scan obtained prior to ¹³C imaging (Fig. 3B) are free of Nyquist ghosts. The SSIM index (Fig. 3D) was measured to be greater than 0.91 for all timeframes using the ¹H reference scan, whereas the absence of a reference scan leads to Nyquist ghosting (Fig. 3C) and substantially lower structural similarity for all timeframes. The dynamic time course (Fig. 3E) was not significantly different (t-test, p = 0.98), further supporting the utility of a ¹H reference scan.

The in vivo dynamic data from a separate experiment can be seen in Fig. 4. Both pyruvate and lactate metabolite maps exhibit high SNR and are free of Nyquist ghost artifacts. The variable flip-angle schedule attempts to optimize the rate parameter estimates in vivo, yielding lactate images that provide nearly constant SNR throughout the time course. The absence of ghosting throughout the acquisition indicates that gradient heating and infidelity was not an issue, and a single reference scan was sufficient for phase correction for all time points. The higher spatial resolution of EPI (3mm in-plane) also enabled the visualization of differential metabolism between the renal cortex and medulla.

Discussion

In this study, we implemented and tested a ramp-sampled, symmetric, partial Fourier EPI sequence with spectral-spatial RF excitation for hyperpolarized ¹³C imaging. The benefits of a symmetric readout were validated, and the feasibility of obtaining a suitable reference scan and artifact-free metabolite maps in vivo were demonstrated. In particular, ramp-sampling and partial Fourier further reduced the echo-spacing and TE for symmetric EPI, making it more robust to chemical shift artifacts in the slow phase encoding direction. In principle ramp-sampling and partial Fourier can result in an SNR penalty because of the reduction in total acquisition time and trajectory efficiency, but in practice this reduction is typically offset by the increase in SNR from the shorter TE and reduced echo-spacing. This acceleration will be crucial for the translation to clinical imaging, as the larger patient FOVs will necessitate a concomitant increase in matrix size to maintain adequate spatial resolution.

Estimating the timing delays between even and odd lines of k-space is crucial to minimizing ghosting artifacts in a symmetric EPI acquisition. We explored the reference scan correction using the ¹H channel, an external ¹³C urea syringe, or the remaining magnetization from the hyperpolarized substrate. The external ¹³C syringe works well for single, thick slice phantom studies but is problematic for thinner slices due to poor SNR or for multi-slice experiments when the imaging volume is larger than the phantom. While the reference scan using the residual hyperpolarization can also produce a suitable reference scan, it worked best in conjunction with a single-slice axial technique. This is an outcome of the experimental setup, since in-flowing spins provide fresh magnetization for a reference scan following a VFA scheme that ends at a 90-degree flip. For coronal slices, or for 3D imaging, this approach will likely prove difficult because of the lack of residual magnetization.

Instead, acquiring the reference scan from ¹H data is an appealing alternative. Complementary information from the ¹H channel has been previously shown to aid in ¹³C reconstruction. For instance, a field map can be readily obtained on the ¹H channel and applied to ¹³C data by accounting for the difference in gyromagnetic ratio (21), and the k-space trajectory can be measured using Duyn’s method (22) on the ¹H channel and applied to ¹³C data in a similar manner. Similarly, we have shown that a suitable reference scan can be obtained in an analogous manner. This approach provided a more robust, higher SNR method to calculate the phase coefficients and correct for timing delays. The similarity in the computed coefficients, as well as the structural similarity index between images reconstructed with the ¹H or ¹³C reference scan, implies that a ¹H reference scan can be acquired prior to and applied on hyperpolarized ¹³C imaging to generate images that are free of Nyquist ghost artifacts. This technique is likely to be more robust than external ¹³C syringes due to improved SNR and will be especially useful in clinical imaging where space considerations and patient comfort must be considered.

Other methods of accounting for timing delays in symmetric EPI have limitations. Doubly sampled EPI (23) can yield artifact free images by acquiring each line of k-space twice and separately reconstructing even and odd echoes. However this will substantially increase the echo-spacing and echo-time, resulting in signal loss and further filtering due to T₂* decay. Timing delays can also be estimated by acquiring two images, with the second shifted by one echo such that the even lines of k-space in the first image line up with the odd lines of k-space in the second image (24). This method effectively doubles the scan time and RF deposition, rendering it largely ineffective for hyperpolarized imaging. Navigator echoes are an alternative approach (25) that have also been shown to be effective in minimizing Nyquist ghosts. While it provides an updated reference scan for each timeframe, it does results in an increase in the TE. Calibrationless reconstruction techniques, such as minimizing the entropy in image-space (26), are an appealing alternative to a reference scan to mitigate Nyquist ghosting and will be explored in subsequent research.

Both spiral and Cartesian EPI are well-suited for single-shot imaging of hyperpolarized substrates, as both utilize high duty cycle and efficient trajectories (27-29). While a spiral acquisition has the benefit of a shorter echo-time and high SNR efficiency, it is more sensitive to off-resonance artifacts and gradient infidelities, but is robust to flow effects and pulsatile motion (30,31). Hyperpolarized pyruvate studies using spiral trajectories have relied on auto-focusing reconstruction methods to correct for off-resonance artifacts (8). EPI is more robust in the presence of off-resonance, resulting in either a benign shift in the PE direction for chemical shift or object distortion for an inhomogeneous B₀ field. Partial Fourier and parallel imaging techniques are also readily amenable to EPI trajectories and can be used to accelerate imaging (32), further reduce RF deposition, and reduce the TE for larger FOVs.

Conclusion

This ramp-sampled, symmetric EPI approach with spectral-spatial excitation of a single metabolite provided a fast, robust and clinically efficacious way to acquire hyperpolarized ¹³C dynamic molecular imaging data. The substantial?SNR increase enables improved temporal and spatial localization and?increased scan coverage. The gains of this efficient sampling, combined with?partial Fourier methods, will be crucial for large matrix sizes required for?human-sized FOVs. This pulse sequence is well suited for clinical dynamic imaging of hyperpolarized substrates as it leverages well-established EPI methods to reconstruct ¹³C metabolite maps on a commercial scanner in real time, and can be easily incorporated into existing diagnostic workflows.

Supplementary Material