Fast and motion-robust saturation transfer MRI with inherent B₀ correction using rosette trajectories and compressed sensing

Abstract

Purpose:

To implement rosette readout trajectories with compressed sensing (CS) reconstruction for fast and motion-robust chemical exchange saturation transfer (CEST) and magnetization transfer contrast (MTC) imaging with inherent correction of B₀ inhomogeneity.

Methods:

A pulse sequence was developed for fast saturation transfer imaging using a stack of rosette trajectories with a higher sampling density near the k-space center. Each rosette lobe was segmented into two halves to generate dual-echo images. B₀ inhomogeneities were estimated using the phase difference between the images and corrected subsequently. The rosette-based imaging was evaluated in comparison to a fully sampled Cartesian trajectory and demonstrated on CEST phantoms (creatine solutions and egg white) and healthy volunteers at 3T.

Results:

Compared to the conventional Cartesian acquisition, CS-reconstructed rosette images provided image quality with overall higher CNR and significantly faster readout time. Accurate B₀ map estimation was achieved from the rosette acquisition with a negligible bias of 0.01 Hz between the rosette and dual-echo Cartesian GRE B₀ maps, using the latter as ground truth. The water saturation spectra (Z-spectra) and amide proton transfer weighted (APTw) signals obtained from the rosette-based sequence were well preserved compared to the fully sampled data, both in the phantom and human studies.

Conclusions:

Fast, motion-robust, and inherent B₀-corrected CEST and MTC imaging using rosette trajectories could improve subject comfort and compliance, CNR, and provide inherent B₀ homogeneity information. This work is expected to significantly accelerate the translation of CEST-MRI into a robust, clinically viable approach.

1. INTRODUCTION

Chemical exchange saturation transfer (CEST) imaging has become an important molecular imaging technique that can identify low-concentration, water-exchangeable organic compounds in living tissue, including proteins, peptides, and various metabolites (1–3). The clinical applications of CEST-MRI are promising; for instance, amide proton transfer (APT) imaging has been successfully used to assess brain tumors, strokes, and neurodegenerative diseases (4–15). However, a limitation of the routine clinical use of CEST-MRI is the requirement to add another scan to the already long clinical imaging protocol. For conventional CEST imaging, the long scan time sometimes required due to the use of saturation acquisitions at multiple frequencies is vulnerable to subject motion, which can result in spatial blurring, distortion, loss of contrast, and poor signal-to-noise ratio (SNR) in the images acquired with Cartesian readouts. Furthermore, CEST imaging is significantly influenced by spatial as well as temporal B₀ variation, the latter due to motion potentially affecting the local saturation frequency in each voxel. The acquisition of an independent B₀ map for inhomogeneity correction requires additional imaging time.

Non-Cartesian acquisition techniques with oversampling near the center of the k-space are inherently less sensitive to bulk motion artifacts and significantly accelerate the imaging time compared to conventional Cartesian acquisition techniques. The advantage of radial sampling and periodically rotated overlapping parallel lines with enhanced reconstruction (PROPELLER) readout for CEST imaging has been demonstrated in previous studies (16, 17). However, these imaging techniques still require a relatively long scan time for whole-brain coverage and additional time for B₀ map scans. Compressed sensing (CS) is another effective technique to accelerate imaging time by leveraging the sparsity property of MR signals to restore missing data from an under-sampled k-space (18–20). Combining CS with highly under-sampled non-Cartesian acquisition could further accelerate image acquisition. Previous studies have shown that the rosette readout trajectory can provide more incoherent under-sampling compared to other non-Cartesian trajectories, which is ideal for CS reconstruction (21–23). Another advantage of rosette trajectories is the self-derived correction of B₀ inhomogeneity (24, 25). The rosette pattern crosses the center of the k-space multiple times in each shot, which can be segmented to generate multiple echo images. This can be used for B₀ estimation, eliminating the need for additional scans for B₀ mapping and their co-registration with CEST images. In addition, a flexible trajectory design and smoothly varying gradient amplitudes make the rosette trajectory very compelling. Herein, we developed a fast and motion-robust CEST imaging technique using rosette readout trajectories with inherent B₀ correction and CS reconstruction. The rosette-CEST imaging technique was evaluated against fully sampled Cartesian trajectories and validated on CEST phantoms and healthy volunteers.

2. METHODS

2.1. Rosette trajectory

The rosette trajectory oscillates in the radial direction around the center of the k-space with an angular frequency of ${\omega }_1=2\pi f_1$, while also simultaneously rotating in the k_x-k_y plane with an angular frequency of ${\omega }_2=2\pi f_2$. The k-space trajectory is given by (26, 27)

$$k\left(t\right)=k_m\text{sin}\left({\omega }_{1}t\right)e^{i{\omega }_{2}t}$$ (1)

where $k_m=\frac{N_{x,y}}{2\bullet FOV}$ is the maximum spatial frequency, FOV is the field of view, and N_x,y is the matrix size. The values of f₁ and f₂ can be chosen based on the intended shape of the k-space trajectory, and the maximum gradient strength and slew rate limitations of the scanner. The corresponding gradient amplitude is calculated by $G\left(t\right)=\frac{2\pi }{\gamma }\frac{dk\left(t\right)}{dt}$, where $\gamma$ = ¹H gyromagnetic ratio. Figure 1A illustrates the 2D multi-slice gradient recalled echo (GRE) rosette-CEST sequence, consisting of RF saturation preparation and fat suppression, followed by a rosette sampling. The corresponding rosette readout trajectory is shown in Fig. 1B.

Figure 1:

Sequence diagram and the k-space trajectory for CEST MRI with rosette readout. (A) The sequence consists of twenty rectangular pulses (100 ms each), each followed by a crusher gradient (5 ms duration and 15 mT/m strength), and a fat suppression pulse. Slice selection was followed by rosette gradients to acquire the first shots in each slice. This was repeated to acquire all the shots to fill up the k-space. The corresponding rosette k-space trajectories are shown in (B), where the bold black line represents the k-space trajectory for a single shot.

2.2. Compressed sensing reconstruction

CS methods employ non-uniform under-sampling strategies to generate incoherent aliasing artifacts and nonlinear reconstruction techniques to enforce sparsity in an appropriate transform domain, capitalizing on spatial and temporal correlations to expedite MRI. Essentially, it is an optimization problem that can be formulated as follows (18, 23, 28)

$$arg\underset{x}{\text{min}}({\parallel Y-Fx\parallel }_2^2+\lambda {\parallel \psi x\parallel }_1)$$ (2)

where Y is the under-sampled k-space data; F is the under-sampled Fourier transform; x is the reconstructed data; $\psi$ is a sparsifying transform; and $\lambda$ is a regularization parameter that weights the relative importance of the two terms. The symbols ${\parallel \dots \parallel }_1$ and ${\parallel \dots \parallel }_2^2$ represent summations of absolute values and their squares, respectively. Eq. 2 can be solved using a nonlinear conjugate gradient descent algorithm (18).

2.3. MRI experiments

The rosette-CEST sequence was tested on an egg white phantom and 10% (wt/wt) cross-linked bovine serum albumin (BSA; Sigma Aldrich, St. Louis, MO, USA) in a phosphate-buffered solution (PBS) with different RF saturation strengths (B₁). The creatine (Cr) concentrations in the BSA phantom ranged from 50 to 100 mM and the pH ranged from 6 to 7. The in vivo study was approved by the Johns Hopkins University Institutional Review Board (IRB). All experiments were performed in accordance with the IRB guidelines and regulations. All the subjects provided informed consent prior to participation in the study. Four healthy subjects (one female, three males, age = 26 ± 9 years, weight = 63 ± 10 kg) participated in the study.

All experiments were performed using a 3T whole-body MRI scanner (Magnetom Prisma, Siemens, Erlangen, Germany) using a 20-channel head and neck coil. The maximum gradient and slew rate of the scanner were 80 mT/m and 200 T/m/s, respectively. The RF saturation module consisted of twenty 100-ms long rectangular pulses with an inter-pulse delay of 6 ms. Five-ms long crusher gradients with an amplitude of 15 mT/m were executed during the inter-pulse delay. The total preparation time for the RF saturation module was 2.1 s. For the phantom studies, sixty-five saturated images (S_sat) were acquired in the offset range of −30 to 30 ppm (with an interval of 2 ppm from ±30 ppm to ±6 ppm and an interval of 0.25 ppm from ±4.75 ppm to 0 ppm) with B₁ values of 0.5 and 1.5 μT. An unsaturated image was acquired for signal normalization. For in vivo studies, forty-one rosette-CEST images were acquired in the offset range of −8 to 8 ppm (± 8 ppm, ±6 ppm, then with an interval of 0.25 ppm from ±4.5 ppm to 0 ppm) with B₁ values of 1 and 2 μT. For the in vivo study, we decreased the number of frequency offsets to reduce the scan time for subject comfort and used higher B₁ values (up to 2 μT), as the consensus paper suggested (7). Other rosette acquisition parameters were as follows: FOV = 256 X 256 mm²; N_x,y = 128; slice thickness = 4 mm; number of slices = 10; flip angle = 10°; ω₁ = 3500 rad/s; ω₂ = 4500 rad/s; number of shots = 100; sampling time = 2 μs; TE = 1.04 ms; and TR (time between successive excitations) = 2.9 ms. The total readout time (10 slices) for rosette acquisition was ~2.9 s and the total imaging time per frequency offset including the RF saturation module was ~5.1 s. For comparison, fully sampled Cartesian (fast low angle shot, FLASH) images were acquired with the same geometry parameters as the rosette sampling acquisition. Other FLASH acquisition parameters were as follows: TR = 5.75 ms; TE = 2.9 ms; and flip angle = 15°. As the FLASH TR was longer than the rosette TR, a slightly higher flip angle was used for the FLASH acquisition than the rosette for a fair comparison between the images, as both the signal and contrast in the images depend on the flip angle for a given TR. The total readout time for the FLASH acquisition was ~7.3 s. The robustness to motion for both Cartesian and rosette trajectories was also evaluated. During the experiment, a subject was instructed to perform moderate motion by moving hands and legs during Cartesian and rosette image acquisition. The motion-related MRI experiments were repeated three times. To evaluate the B₀ map estimated from rosette trajectories, an additional B₀ map was acquired using the conventional dual-echo Cartesian GRE sequence with TE₁ / TE₂ / TR = 4.92 / 9.84 / 29 ms, which was used as the gold standard.

2.4. Data analysis

Rosette images were reconstructed from different numbers of shots using gridding and CS, as described before (23). A total variation (TV) sparsifying transform ($\psi$) with $\lambda$ = 0.05×S₀ was used (Eq. 2) for CS reconstruction from rosette acquisition, where S₀ was the peak magnitude of the image. The contrast-to-noise ratio (CNR) between the gray matter (GM) and the white matter (WM) of the human brain image was calculated using (29)

$$CNR=\frac{\mathit{\text{mean}}\left(S_{GM}\right)-\mathit{\text{mean}}\left(S_{WM}\right)}{SD\left(S_{\mathit{\text{noise}}}\right)}$$ (3)

where S_tissue/noise is the signal intensity from the tissue of interest or the noise (air/background), and SD is the standard deviation. The ROIs were drawn manually on the unsaturated image (S₀).

To evaluate the spatial point spread function (PSF_x,y) of the rosette trajectory for different types of tissues (i.e., GM, WM, and CSF), a delta function (1 pixel wide in all dimensions) was simulated in the image domain. The initial signal relaxation during the rosette shots was simulated using the extended phase graph (EPG) algorithm for GM, WM, and CSF (30). In MRI experiments, the acquired signal represents an average of all spins within a voxel, imposing the simulation of signal evolution during each TR using multiple isochromats (magnetization vectors). The EPG algorithm, widely used in MR fingerprinting (MRF) studies (31), efficiently simulates signal evolution by considering numerous isochromats with different dephasing angles. It accounts for the effects of RF pulses, T₁ and T₂/T₂* relaxations, and dephasing due to balanced/unbalanced gradient moments. For this study, signal relaxation during the rosette shots was simulated using the EPG algorithm with a flip angle of 10⁰; TE/TR of 1.04/2.9 ms; and T₁/T₂* of 1300/70 ms, 800/40 ms, and 4000/1200 ms for GM, WM, and CSF, respectively. Finally, PSF_x,y was calculated using rosette k-space sampling (100 shots) and subsequent reconstruction (32). The impact of k-space filtering due to the sampling pattern and relaxation effects was assessed by convoluting the resulting PSF_x,y for GM, WM, and CSF with a delta function. The spectral PSF (PSF_Δω) for the rosette trajectory was also evaluated using (27)

$$\left|{PSF}_{\Delta \omega }\right|=\frac{1}{N_{\delta }}\left|\frac{\text{sin}\left(\Delta \omega .N_{\delta }.\frac{t_{\delta }}{2}\right)}{\text{sin}\left(\Delta \omega .\frac{\Delta t_{\delta }}{2}\right)}\right|$$ (4)

where Δω is the frequency offset, N_δ is the number of spectral points, and t_δ is the spectral dwell time (= the duration of each rosette shot for each spectral point).

As the rosette trajectory crosses the k-space center twice for each shot, two low-resolution images at two echo times (TE₁ / TE₂ = 1.04 / 1.5 ms) were generated from a single acquisition (Fig. 3A), and their phases (ϕ₁ and ϕ₂, respectively) were used to estimate the B₀ map. Bland-Altman analysis was performed to evaluate the agreement between the B₀ maps acquired with dual-echo Cartesian and the rosette readouts. After B₀ correction of the Z-spectra in each voxel, MTR_asym for each frequency offset was calculated using (33)

$${MTR}_{\mathit{\text{asym}}}\left(\Delta \omega \right)=Z\left(-\Delta \omega \right)-Z\left(\Delta \omega \right)$$ (5)

Two-tailed t-tests were performed for statistical analysis. In all cases, p<0.05 was considered the level of significance. All results are reported as mean ± SD. The image and statistical analyses were performed offline using MATLAB (MathWorks, Natick, MA).

Figure 3:

Estimation of a B₀ map from the rosette encoding from a representative subject. (A) As the rosette sampling pattern crosses the center of the k-space twice for each shot, a single lobe can be segmented into two halves to generate low-resolution dual-echo images (black and red lines for echo 1 and echo 2, respectively). Here, only five shots are shown for clarity. (B) The phase information (ϕ₁ and ϕ₂) of the dual-echo images was used to estimate a B₀ map, which was compared quantitatively with the B₀ map from a conventional dual-echo Cartesian GRE acquisition. (C) Bland-Altman analysis among the estimated B₀ values from all the voxels of the brain region shows very good agreement between the two methods.

3. RESULTS

The maximum gradient amplitude and slew rate for the designed rosette trajectory were 30.6 mT/m and 190.8 T/m/s, respectively. The results of the PSF simulations are shown in Fig. 2. Figure 2A shows the signal relaxations for GM, WM, and CSF during 100 rosette shots (normalized by the signal after the 1^st rosette shot for each respective tissue type). The spatial and spectral PSF are shown in Figs. 2B and C, respectively. The spatial PSF (PSF_x,y) shown in Fig. 2B exhibits an irregular distribution for all tissue types, as demonstrated by the diagonal and central projections. This irregularity reflects the incoherent sampling pattern of rosette trajectory, which is advantageous for CS reconstruction (18, 21). The convolutions of PSF_x,y for GM, WM, and CSF with a delta function show the effect of k-space filtering resulting from the sampling pattern and signal relaxation during rosette shots. Although the ringing effect is not clearly visible in the regular scale, the convolution results in the logarithmic scale clearly reveal its presence across all tissue types. The spectral PSF (PSF_Δω) shows that the local maxima and minima are a function of the duration of each rosette shot (Fig. 2C). The cut-off frequency for the passband filter is determined by the duration of each shot (26, 27, 32).

Figure 2:

Spatial and spectral point spread functions (PSF) for the rosette trajectories. The spatial PSF (PSF_x,y) was calculated for the gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) by simulating a delta function in the image domain with signal relaxation during rosette shots using the extended phase graph (EPG) algorithm, followed by rosette k-space sampling and subsequent reconstruction. (A) Signal evolutions during 100 rosette shots (normalized by the signal after the 1^st rosette shot) for GM, WM, and CSF. (B) Zoomed-in view of PSF_x,y for GM, WM, and CSF in logarithmic scale, showing projections in the central (x) and diagonal (x,y) directions as well as the spatial distribution in the x-y plane. PSF_x,y exhibits an irregular distribution, evident from both central and diagonal projects, demonstrating the irregular sampling pattern advantageous for compressed sensing (CS) reconstruction. Convolution results between a delta function and PSF_x,y for GM, WM, and CSF on a logarithmic scale reveal the presence of ringing effects. (C) The spectral PSF (PSF_Δω, using Eq. 4) of the rosette trajectory demonstrates local maxima and minima as a function of the duration of each rosette shot.

Figure 3B shows a comparison between the B₀ maps estimated from rosette and the ground truth dual-echo Cartesian GRE readouts for one slice from a representative subject. Bland-Altman analysis (Fig. 3C) shows that an accurate B₀ estimation can be achieved from the rosette-CEST acquisition. No significant bias for the estimation of the B₀ map was observed and the mean difference between the two measurements was 0.01 Hz, with limits of agreement (±1.96 SD) of +1.31 and −1.63 Hz.

A comparison between in vivo brain images from a healthy volunteer under Cartesian full k-space and rosette acquisitions is shown in Fig. 4. The overall CNR between the GM and WM was higher in the rosette images compared to the Cartesian images (2.2 ± 0.3 vs 1.8 ± 0.5 for rosette with 100 shots and Cartesian, respectively, p < 0.01). The total readout time for the rosette acquisition with 100 shots was significantly faster than the Cartesian readout time (~2.9 vs ~7.3 s for the rosette and the full k-space reference, respectively). Rosette images with higher acceleration factors (by using a smaller number of rosette shots) also demonstrated higher CNR than Cartesian images (CNR = 2.1 ± 0.3, 1.9 ± 0.3, and 1.9 ± 0.4 for 75, 50, and 25 shots, respectively, p < 0.05), which can further accelerate the readout time by a factor of 4 (Fig. 4). The CS reconstruction time for rosette images (10 slices) was 35 s. The effects of moderate motion on the Cartesian and rosette images are compared in Fig. 4C. While severe motion artifacts were apparent in the Cartesian images, rosette images showed minimal artifacts, indicating higher robustness to bulk motion. The normalized root mean squared errors (nRMSE) between reference (motion-free) and motion-corrupted images were 10.3 ± 6.2% and 2.7 ± 1.8% for Cartesian and rosette images, respectively.

Figure 4:

(A) Comparison of images acquired with a rosette trajectory using different numbers of shots and a conventional fully sampled Cartesian trajectory (fast low-angle shot, FLASH) from the same subject with the same FOV, in-plane resolution, slice-thickness, and the total number of slices. (B) Bar plots showing comparisons of CNR between gray matter (GM) and white matter (WM), and the total readout time for 10 slices between Cartesian and rosette readouts. Rosette acquisition resulted in an overall higher CNR with significantly faster readout time compared to the Cartesian acquisition. (C) Comparison of motion robustness between Cartesian and rosette acquisitions. The Cartesian images exhibited severe artifacts under moderate motion, whereas the rosette images showed minimal artifacts, demonstrating a higher robustness to bulk motion.

Figure 5A shows the effect of inter-pulse crusher gradients for the RF saturation module on the Z-spectra acquired from the egg white phantom. Without inter-pulse crusher gradients (where only one crusher gradient was applied after the RF pulse train), a poor-quality Z-spectrum was obtained. Using inter-pulse crusher gradients resulted in a high-quality Z-spectrum, where the APT and rNOE effects can be identified clearly at ±3.5 ppm. Therefore, the inter-pulse crusher gradients were applied for the rest of the experiments. Figures 5B and C show the Z-spectra and corresponding MTR_asym curves from the cross-linked BSA/Cr phantoms. The guanidinium protons of Cr were detected around 2 ppm using the rosette-CEST sequence. As expected, the CEST signals were proportional to the Cr concentrations for these small saturations. The CEST signal increased with higher B₁ due to improved saturation efficiency.

Figure 5:

Egg white and cross-linked BSA/Cr phantom experiments. (A) Comparison of Z-spectra from an egg white phantom acquired without and with crusher gradients between RF saturation pulses (B₁ = 0.5 μT). In the case of ‘w/o crushers’, only one crusher gradient was applied after the RF saturation pulse train, resulting in a poor-quality Z-spectrum. The use of the inter-pulse crusher gradients provided a high-quality appearance of the APT and rNOE signals in the range of +/− 5 ppm around the water resonance. (B) Measured Z-spectra of 50 mM Cr concentration acquired at B₁ of 0.5 and 1.5 μT. (C) MTR_asym plots for Cr concentrations ranging from 50–100 mM at pH = 6.8.

In the in vivo study, the acquired Z-spectra, particularly in the WM, were asymmetric around the water resonance frequency, with lower signal intensities at the negative frequency offsets (Fig. 6A). The MTR_asym curves showed APT peaks around 3.5 ppm from water, particularly in the GM. High-quality APTw (MTR_asym at 3.5 ppm) images obtained from the rosette-CEST sequence are shown in Fig. 6B. A representative comparison between Cartesian and rosette APTw images (B₁ = 2 μT) from the same subject is also shown in Fig. 6B, which demonstrates that the rosette-CEST sequence can serve as an effective surrogate for conventional Cartesian sequence for CEST imaging. The rosette APTw signals were negative at B₁ = 1 μT due to higher upfield rNOE and MTC signals compared to the downfield APT signals. A small contrast between the GM and the WM was observed in the APTw images. The mean, median, and SD of APTw values (N = 4) were −0.9%, −0.7%, and 0.9% for GM and −2.1%, −2.0%, and 1.1% for WM at 1 μT; and 1.6%, 1.2%, and 1.0% for GM and 0.6%, 0.7%, and 0.8% for WM at 2 μT.

Figure 6:

In vivo healthy volunteer studies. (A) The average measured Z-spectra and corresponding MTR_asym curves from gray matter (GM) and white matter (WM). Note that shaded error bars depict standard deviations. (B) Top: a representative comparison between Cartesian and rosette APTw (MTR_asym at 3.5 ppm) images from the same subject. Bottom: rosette APTw images at different B₁ values. (C) Box plot showing the quantitative rosette APTw values at different B₁ values.

4. DISCUSSION

This study is the first to demonstrate the feasibility of a fast and motion-robust CEST MRI using rosette trajectories with inherent B₀ correction. The CEST image quality from the rosette acquisition concordant with the conventional fully sampled Cartesian acquisition, but significantly reduced readout time at least by ~2.5-fold. The self-derived B₀ map from rosette trajectories was accurate and comparable with the B₀ map estimated from the conventional dual-echo Cartesian GRE acquisition. Therefore, no extra scans were needed for B₀ correction due to the inherent B₀ mapping capability.

Various k-space sampling strategies using Cartesian readouts have been used for CEST imaging in the human brain, such as turbo spin echo (TSE), gradient echo (GRE), gradient and spin echo (GRASE), and echo planar imaging (EPI) (34–40). A few studies have explored the potential of non-Cartesian acquisition for CEST imaging, mostly with radial acquisition (16, 17, 41). The improved robustness to bulk motion effects for CEST imaging using such non-Cartesian trajectories has been successfully demonstrated in previous studies (17). However, these techniques are still limited by the significant imaging time required for volumetric imaging and the requirement for additional B₀ map scans. For CEST-MRI to be adopted as a routine clinical scan, the scan time ideally needs to be as short as possible with a minimum number of additional scans. The rosette trajectory could be very useful in achieving this goal of translating CEST imaging into a routine clinical scan with its rapid sampling property combined with efficient under-sampling capability for CS reconstruction. In addition, the rosette acquisition could eliminate the need for an additional scan for B₀ mapping. As the center of the k-space is sampled repeatedly, rosette acquisition is robust to bulk motion artifacts, which can improve the image quality compared to conventional Cartesian acquisition (Fig. 4C).

CEST imaging is vulnerable to variations in B₀, as the magnetic field inhomogeneity induces spatial variation in resonance frequencies, and therefore, B₀ correction is essential for CEST imaging when comparing voxel-based Z-spectra. The simplest approach for B₀ correction in CEST imaging is to determine the water resonance frequency from the Z-spectrum itself, which requires the full sampling of a relatively high spectral resolution Z-spectrum (42). This approach works only when a low RF saturation power is used to remove direct water saturation and semisolid MT effects, since a sharp and narrow Z-spectrum facilitates the measurement of a water frequency shift according to a field inhomogeneity. However, the approach is sometimes limited by saturation efficiency for CEST imaging. In vivo CEST measurements with higher saturation power and more accurate B₀ correction can be achieved by acquiring an external B₀ map. Conventional CEST imaging techniques typically acquire the external B₀ map using multi-echo GRE acquisition or water saturation shift referencing (WASSR) (43–45), which require additional imaging time. This additional scan can be avoided using the rosette-CEST imaging technique, as demonstrated in this study. From the phase difference of dual-echo rosette images generated from the two halves of each shot, an accurate B₀ map can be estimated. As the B₀ map is calculated from the CEST image itself, no registration between the B₀ map and CEST images is required for inhomogeneity correction. Furthermore, dynamic B₀ correction for each saturation frequency can allow the independent B₀ correction of individual Z-spectral points and minimize the temporal B₀ variation that may arise from the system instabilities, cardiac effects, and subject movements (46–49). As the B₀ map is estimated from the CEST image itself in the rosette-CEST imaging technique, dynamic B₀ correction could be possible.

The imaging time for the rosette-CEST technique with a brain coverage of 256 × 256 × 50 mm³ was ~5.1 s per saturation frequency offset (including 2.1 s RF saturation) at the spatial resolution of 2 × 2 × 4 mm³. This is faster than previous studies with radial and Cartesian acquisitions using TSE, GRE, EPI, and GRASE readouts (acquisition time ranging from 6.9 – 20 s per saturation frequency offset) (17, 34, 36, 50, 51). In terms of a readout time, the rosette readout time with 100 shots (~2.9 s) was relatively longer than the previously proposed 3D EPI readout (<1 s) and snapshot++ 3D GRE readout (~2.1 s) times (52, 53). However, the EPI sequence required a separate reference scan for the GRAPPA calibration (3 × 2-fold under-sampling) and an external EPI phase correction scan. Additionally, the EPI images are susceptible to geometric distortions and signal dropouts due to local B₀ variation, leading to increased errors in CEST quantification. Snapshot++ with 3D GRE readout also used a very high acceleration factor (= 10.98) with CS reconstruction but resulted in lower SNR and CNR compared to the conventional snapshot sequence. The rosette readout time can be further reduced by 4-fold with 25 shots, as demonstrated in Fig. 4. This acceleration in imaging time could be traded off for higher spatial resolution and/or more brain coverage for CEST imaging. Using the same acquisition and RF saturation parameters used in this study, it would be possible to acquire 30 slices in ~15.3 s per frequency offset for a whole brain coverage, which can be further accelerated to ~4.4 s per frequency offset by using a minimum rosette shot of 25. Note that retrospective under-sampling was performed using different number of shots for the rosette acquisition, which might be different from the prospective under-sampling result.

The rosette-CEST sequence demonstrated good sensitivity to different saturation parameters and metabolite concentrations, as shown from phantom and in vivo studies. The sequence was able to generate high-quality Z-spectra and APTw images. Although it is difficult to compare the APTw images among different studies due to the different off-resonance saturation parameters, acquisition parameters, and readout techniques used, the APTw images reported in this study were in line with the previous results using similar saturation parameters (7, 34, 54). Note that it was found necessary to apply crusher gradients between the RF saturation pulses for the high-quality appearance of the CEST and rNOE signals (Fig. 5), similar to previous findings (34, 54). This is also consistent with the sideband effect in Z-spectra when no crusher gradients are applied during the inter-pulse delay in the RF saturation pulse train (55). Nevertheless, a very high duty cycle (> 95 %) was achieved for the RF saturation module.

The rosette-CEST sequence developed in this study used conventional slice selection followed by in-plane rosette encoding to stack rosette acquisition. Volumetric rosette encoding (i.e., applying rosette gradients simultaneously in the x, y, and z directions) could potentially enhance efficiency for volumetric under-sampling and CS reconstruction, thereby further accelerating imaging time (32). As our current implementation involves a 2D multi-slice rosette acquisition, adopting a 3D rosette readout could yield an SNR gain, as previously done in 3D GRE and EPI readouts (52, 53). Our future work will focus on implementing CEST imaging with 3D volumetric rosette encoding. Although the regularization parameters used in this study achieved high-quality reconstructions similar to previous studies (23, 28), residual ringing artifacts were still observed in the reconstructed images. These artifacts are probably attributable to the different intensities of adjacent k-space points influenced by both the shape of the k-space trajectory and signal relaxation during the rosette shots. Several post-processing techniques have been proposed to mitigate ringing artifacts (56–58), which will be explored in our future studies. As shown in Fig. 4, the CS reconstruction can further accelerate the readout time by an additional maximum factor of 4. Further optimization of the CS reconstruction parameters may improve the reconstruction of such highly under-sampled data. We used a fat suppression pulse in the rosette-CEST sequence, as recommended by the tumor APTw imaging consensus paper (7). Although a previous study showed that using a frequency-selective fat suppression pulse can improve the APTw image, especially in the brain region near the skull (34), it might interfere with RF saturation at −3.5 ppm targeted for rNOE effects. Alternatively, a water-selective excitation using binomial pulses can be used (53). Interestingly, the fat signal could be also suppressed in the rosette trajectory itself. As shown in Fig. 2C, the cut-off frequency for the passband filter in the rosette spectral PSF can be adjusted by changing the duration of each rosette shot, which can also be used for fat suppression (24). We will evaluate the performance of these different techniques in our future study.

5. CONCLUSIONS

We demonstrated the feasibility of saturation transfer imaging (CEST and MTC) using rosette trajectories, which offers faster acquisition, increased robustness to bulk motion, and inherent B₀ correction compared to conventional imaging techniques. This technique can significantly accelerate the imaging time for CEST imaging and can be very beneficial for clinical applications.