Sidebands in CEST MR—How to recognize and avoid them

Clinical scanners require pulsed CEST sequences to maintain amplifier and specific absorption rate limits. During off‐resonant RF irradiation and interpulse delay, the magnetization can accumulate specific relative phases within the pulse train. In this work, we show that these phases are important to consider, as they can lead to unexpected artifacts when no interpulse gradient spoiling is performed during the saturation train.

Z-spectrum also contains direct off-resonant saturation effects that are the focus of the present work.
In preclinical scanners, RF saturation can be achieved using continuous wave (CW) irradiation (i.e., a long rectangular pulse of several seconds).The advantages of CW is the efficient continuous saturation, easy implementation at the scanner, and analytical description of the resulting data via Bloch-McConnell theory. 2,3owever, for clinical MR scanners, hardware limitations such as a maximal pulse duration, due to amplifier restrictions and limits in specific absorption rate (SAR), constrain the application of CW irradiation for CEST imaging.Pulsed saturation (i.e., saturation using a long train of rf pulses) is required to overcome amplifier and SAR limits.Unlike CW, pulsed saturation introduces the delay time (t d ) between the RF pulses of the duration t p and allows an optimization of the CEST effect by using high duty cycles (DCs) to generate a CW-like saturation.Thus, several rectangular pulses connected in pulse train series could be used for saturation.However, the spectral response of shorter rectangular pulses can produce sever Rabi oscillations, which lead to unwanted off-resonant saturation effects and which is why shaped RF pulses are often used for smooth labeling. 4t is commonly accepted that CW equivalent pulse trains can be achieved when adjusting different parameters such as the saturation time (t sat ), the B 1 root-mean-square (B 1rms ) RF amplitude, and a high DC. 5 The pulse shape as well as the recovery time T rec played out, as relaxation delay before the saturation also affects the final CEST effect.Thus, with the parameter set (B 1rms , t sat , DC, t p , t d , T rec , pulse shape) CEST experiments are supposed to be well-defined as published in the amide proton transfer-weighted (APTw) CEST-MRI consensus protocol. 6e show herein that this definition is incomplete, because the RF phase of each saturation pulse is typically not defined, and different choices for the RF phase can lead to very different Z-spectra in simulation and in real measurements, despite all other parameters being identical.In this regard, we point out a so-called sideband pattern that appears in the Z-spectrum that can interfere with postprocessing steps to generate pseudo-CEST effects.
The observed patterns and Z-spectra sidebands were reported before 4 but were until now not investigated in detail.As we observe them now in a consensus protocol, a careful reexamination of the phenomenon is considered herein.Last but not least, this work shows that an exact and easy reproducible definition of CEST pulse trains is important for both measurement and simulation, which is realized here using the Pulseq-CEST standard (https:/ /pulseq-cest.github.io/).

CEST sequence definition
To achieve CW-like CEST spectra, several sequence parameters have to be adjusted for pulsed saturation.A typical CEST pulse train is depicted in Figure 1 and is defined by a frequency-selective RF pulse of a certain shape and amplitude B 1 , the pulse duration (t p ) and interpulse delay time (t d ), and the number of pulses (n).The saturation duration (t sat ) is then given by t sat = n⋅(t p + t d ), whereas the saturation duty cycle is defined by DC = t p /(t p + t d ).The DC is important to be high to achieve a similar CEST weighting as for CW.For the APTw consensus, the recommended DC is at least 90%. 6he exact amplitude of the pulse shape can be defined in different ways, such as peak amplitude or average amplitude, but spectra most similar to CW are achieved when the RMS B 1 (B 1rms ), defined by Eq. ( 1), matches the CW B 1 .
The APTw CEST-MRI consensus protocol 6 is defined by the following parameters: a pulse train of 36 sinc pulses of t p = 50 ms, t d = 5 ms, B 1,rms = 2 μT, t sat = 2 s, and DC = 91%.The protocol with a recovery time before saturation of T rec = 3.5 s is shown in Figure 1A.Another option in pulsed saturation is the addition of gradient spoiler events between the pulses to spoil transverse magnetization, as shown here in x-, y-, and z-direction (Figure 1E,F).Such gradient spoiling requires a DC of less than 100%.
In addition to these parameters, the individual phase Φ i of each RF pulse is an often-overseen degree of freedom.This is interesting, as off-resonant pulses as used in CEST will always lead to an offset-dependent linear phase increment as depicted in Figure 1B, where the initial phase for each RF pulse is set to the phase accumulated during the previous pulses.We call the phase cycling of Eq. ( 2) continued phase cycling in the following: In contrast, Figure 1C shows another scheme in which the phase is reset to 0 before each RF pulse, as follows: A further phase increment pattern is used in this work, where Φ i follows the typical FLASH RF spoiling pattern with a 117 • starting phase angle as described by Zur et al. 7 For exact and complete definition of the CEST preparation, the schemes are described by the Pulseq-CEST

F I G U R E 1
(A) Schematic of a pulsed CEST sequence with 36 sinc pulses, of which two consecutive off-resonant RF pulses and their interpulse delay t d are highlighted in the inlay.The corresponding phase is linearly increasing during each RF-pulse, leading to a finite accumulated phase at the end of the pulse.(B) Typical implementations keep this phase and start the next RF pulse with exactly this accumulated phase.(C) However, one could also just start again with the next RF pulse in the x' direction, meaning Phase 0. (D-F) Placement of feasible spoiling gradients.Generating files of the case ([A] and [B]) is given in the Pulseq-CEST standard here: https://github .com/kherz/pulseq-cest-library/tree/master/seqlibrary/APTw_3T_001_2uT_36SincGauss_DC90_2s_braintumor. standard, with direct links in each caption.All files of the paper are collected and shared via Github (https://github .com/kherz/pulseq-cest-library/tree/master) to enhance reproducibility.

MR sequence
Experiments were performed via Bloch-McConnell simulations using the software tool Pulseq-CEST 8 and by in vitro and in vivo measurements on a 3T MR scanner (Prisma; Siemens Healthineers, Erlangen, Germany) with a 64-channel transmit/receive head coil.CEST preparation was realized by the Pulseq interpreter, combined with a conventional readout; this hybrid sequence was described by Herz et al. 8 Thus, CEST preparation and offsets are completely defined by the provided Pulseq-CEST seq files.The readout was a 3D snapshot gradient-echo sequence protocol 9,10 that was used with TE = 2 s, TR = 4 s, FOV = 180 × 220 mm, and 12 slices at resolution of 2 × 2 × 5 mm, as suggested by Herz et al. 11 With this hybrid setup, we can ensure that exactly the same CEST preparations are played out in simulations, in vitro, and in vivo.

Simulation
Bloch-McConnell simulations were carried out with Pulseq-CEST in MATLAB (version 2022b; MathWorks, Natick, MA, USA) to investigate the effect of several factors, such as the sampling rate, field inhomogeneities, and different phase-cycling schemes on the emergence of the sidebands.Therefore, the CEST preparation, the main version of which can be found in the Pulseq-CEST library (see previous link), was adjusted for each experiment with regard to the RF phase Φ i .Furthermore, we investigated the effect of different pulse shapes as well as the influence through gradient spoiling between the RF pulses.The spoiler gradient strength was set to 80% of the maximum adjustable gradient strength (32 mT/m).For a reproducible simulation of the provided CEST sequence, Bloch-McConnell pool parameters files, the so-called bmsim.yamlfiles, were created.Inhere, the pool model and scanner parameters, such as the field strength and B 0 inhomogeneity, were defined.For this purpose, a standard two-pool model was designed that is most similar to the phantom described in Section 2.4 and defined by a water pool (T 1,W = 1.5 s and T 2,W = 1 s, fb = 1) and a minor CEST pool (T 1,S = 1 s, T 2,S = 0.1 s, k ex = 350 Hz, Δω = 3 ppm) with a concentration of 20 mM, leading to a proton fraction (fb) of fb = 3.6036e −4 .Furthermore, bmsim.yaml-files3][14] All yaml files and further simulation files, which have been used in this paper, are provided in Github (https://github.com/mrmeeseeksjr/Sideband_Example/tree/Sideband_MRM /matlabskripts).

Phantom
A phantom was created consisting of 50-mL falcon tubes with L-arginine for providing the CEST effect.The concentration was slightly increased up to 20 mM, which had to be additionally adjusted in the yaml file.To ensure an exchange rate of about 350 Hz, 15 the pH value was titrated to about 4.0 using HCl in a 0.9% NaCl solution.The T 1 relaxation time was adjusted to about 1500 ms by the addition of 3.9 μL Gadovist to the phantom tube.

Subjects
One healthy subject (male, 27 years old) was scanned after written informed consent, approved by the local ethics committee.

RESULTS
Figure 2 shows a simulated two-pool model for WM, GM, CSF, and a liquid phantom for the APTw-CEST sequence (Figure 1A) with the accumulated phase-cycling scheme (Figure 1B).Here, the WM and GM reveal a negligible substructure around the resonance frequency of water at Δω = ±0.4ppm.On the other hand, the Z-spectrum of CSF (Figure 1C) shows two prominent (Δω = ±1.75ppm) sidebands as well as several smaller ones.This indicates that sideband artifacts emerge even in the consensus APTw protocol and that this pattern becomes more pronounced and even produces new sidebands with increasing T 1 and T 2 relaxation times.This is problematic for tissue compartments with liquid environments such as in cysts or necrotic tumor tissue.In further investigations we focus on the liquid two-pool system (Figure 2D) corresponding to the aqueous arginine phantom described previously.
The observed sideband pattern depends on the chosen sampling rate as well as on the respective B 0 shift Simulated white-matter (WM; A) and gray-matter (GM; B) data show negligible sidebands around the water frequency.For higher T 1 /T 2 relaxation times such as in CSF (C) or for the T 1 adjusted phantom tube (D), the pattern amplifies and new sidebands emerge.Figures can be recreated https://github.com/mrmeeseeksjr/Sideband_Example/tree/Sideband_MRM/matlabskripts/Fig2. of the acquired spectrum, as demonstrated for the phantom tube in Figure 3. Thus, (i) sidebands can be overseen when the sampling rate is low, and (ii) sidebands can lead to clear asymmetric effects as shown in Figure 3E,F that can be misinterpreted as pseudo-CEST effects by MTR asym analysis.Even at very high sampling, the sideband pattern does not stay symmetric around the B 0 shifts, but a complex behavior of the sidebands is observed (Figure 3H).
In the next experiment we investigate the influence of different phase-cycling schemes on the observed sidebands.Figure 4D-F shows simulated Z-spectra for accumulated phase (A), phase reset (B), and a quadratic phase increment (C) typically used in FLASH imaging for RF spoiling of transverse magnetization. 7None of the different phase-cycling schemes appear to eliminate the sidebands.Instead, they reappear in different structural patterns.That the RF phase has an influence on the sidebands also explains the complex B 0 shift influence, as a B 0 shift will lead to a phase shift during the delay between the pulses, with opposite influence for positive or negative offsets.
To answer the question, if these effects are only visible in simulations or if they can be detected in vivo, the same sequences were played out on a real 3T scanner in an aqueous L-arginine phantom.The data show that the sideband pattern is not just a simulation artifact but it can also be detected on the phantom in a real measurement (Figure 4G-I), with close similarity to the Pulseq-CEST simulation.
Simulation of the standard two-pool model for the investigation of sideband artifacts with increasing sampling rate without (A-D) and with (E-H) the presence of B 0 field inhomogeneities.The comparison demonstrates that sidebands appear just by using another sampling rate through skipping the frequency offset at ±1.75 ppm according to the different sampling of the Z-spectrum.In addition to this effect, the smaller CEST pool on the Z-spectrum becomes more visible.By further increasing the sampling rate to 0.05 ppm (C) or 0.01 ppm (D), even more sidebands appear on the Z-spectrum or get further amplified.When adding a small B 0 drift of 0.07 ppm to the simulation (E,F), which can already occur due to field inhomogeneities or the scanner drift during the measurement, the sideband pattern loses its symmetrical appearance and becomes asymmetric distributed over the Z-spectrum.Depending on the chosen sampling rate, this asymmetric distribution can hamper a sufficient B 0 correction such as in Figure 2E.Figures can be recreated via https://github.com/mrmeeseeksjr/Sideband_Example/tree/Sideband_MRM/matlabskripts/Fig3.
Let us now study the influence of the pulse shape itself and the influence of interpulse gradient spoiling.In Figure 5, three different RF-pulse shapes (sinc, Gaussian, block) were simulated using Pulseq-CEST with the same B 1,rms , without (A-C) and with interpulse gradient spoiling (D-F), as well as with interpulse gradient spoiling and an additional B 0 drift (G-I).We can see that sideband patterns depend strongly on the pulse shape revealing a unique pattern; and when interpulse gradient spoiling is applied, the sidebands pattern alters strongly and can lead to a considerable reduction of sidebands as it can be seen for sinc and Gaussian pulses.However, this is not true for every pulse shape, as the block pulse train shows more visible sidebands once gradient spoiling is activated (Figure 5F).Furthermore, gradient-spoiled B 0 -shifted Z-spectra are again symmetric around their center for all pulse shapes as compared to acquired Z-spectra without gradient spoiling and B 0 drift (Figure 3E-H).This translation invariance in the B 0 is an important property required for any B 0 correction.
As the use of spoiler gradients during the interpulse delay time can lead to an improvement in the avoidance of sidebands, we further investigated this matter in a real measurement and tested whether the use of a different phase-cycling scheme has an additional impact.Figure 6 shows the phantom measurement by using sinc RF pulses for saturation.Herein, the measurement with and without gradient spoiling reveals similar Z-spectra and sideband patterns compared with the simulated data (Figure 4A,D).Furthermore, we can learn from Figure 6 that the standard accumulated (B) and optimized 117 • (C) phase-cycling schemes lead to almost similar shapes of the Z-spectrum when gradient spoiling is switched on.This means that the phase cycling can be neglected once gradient spoiling is performed between RF pulses.
Interestingly, the spoiler gradients do not remove all sideband patterns.This hints already to a different origin of different sideband structures, which analyze now for an increasing number of pulses.
In Figure 7, we demonstrate Z-spectra with and without gradient spoiling for increasing number of RF pulses.With gradient spoiling (Figure 7A-F), the Z-spectrum structure for many RF pulses is very similar to the structure for a single pulse; the gradient removes all transverse magnetization and leads to the fact that the pulse acts like a single pulse.Thus, we call these kinds of residual sidebands single-pulse sidebands, which are mostly affected by the single-pulse response and thus the pulse shape.) Schematic of a pulsed CEST sequence with 36 sinc pulses, of which five consecutive off-resonant RF pulses and their interpulse delay t d are highlighted in the inlay.(A) For the typical implementation of continued phase cycling, the corresponding phase is linearly increasing during each RF pulse, leading to a finite accumulated phase at the end of the pulse.This phase is kept and the next RF pulse starts with exactly this accumulated phase.(B) However, one could also just start again with the next RF pulse in x-direction, meaning Phase 0. (C) Another option is using quadratic phase-cycling schemes with a growing phase increment by 117 • .Corresponding adjusted simulations with Pulseq-CEST (D-F) and real measurements in the L-arginine phantom (G-I) directly show similar sideband patterns that are different for all three phase settings.Parts (A)-(F) can be recreated via https://github.com/mrmeeseeksjr/Sideband_Example/tree/Sideband_MRM/matlabskripts/Fig4.Without gradient spoiling (Figure 7G-L), the pattern is more complex and the structures change strongly even after few additional pulses.As residual transverse magnetization is left after each previous pulse and delay, multiple pulses generate certain sidebands by acting together.We therefore call these kinds of sidebands multi-pulse sidebands.A more detailed analysis of the development and dynamic of these multi-pulse sidebands is shown in Supporting Information S2 and S3.The dependency on the transverse magnetization after the previous pulse and delay also explains why these multi-pulse sidebands are affected by phase cycling and B 0 shifts, as these will directly affect the phase of the transverse magnetization.
For the APTw-CEST preparation, the pulses are relatively long (50 ms) and have a high flip angle of 1364 • .To investigate the influence of shorter pulses and the rotation regime of flip angle, we switch now to another established CEST protocol, namely the low B 1 power protocol by Deshmane et al., 10,16 which uses 20-ms RF pulses at a flip angle of about 180 • .In Figure 8, the Z-spectra of this pulse train at different flip angles (90 • , 180 • , 270 • , and 360 • ) with and without gradient spoiling reveal that both multi-pulse and single-pulse sidebands are also present here, but behave differently.For shorter pulses, single-pulse sidebands affect a broader range, as expected from the pulse bandwidth.However, multi-pulse sidebands affect a more narrow range, as they are more dependent on the tilting dynamics, which decreases with decreasing flip angle and B 1 .Altering the on-resonant flip angle appears to affect the sideband structure, but the actual on-resonant flip angle is not the governing factor, as sidebands occur due to off-resonant behavior in which rotation around the effective field and its tilting angle are more important, which is governed by the B 1 level.
From the previous experiments, we can conclude that the CEST preparation with gradient spoiling is more reliable, as it suppresses additional multi-pulse sidebands and leads to B 0 -shiftable spectra.Moreover, the pulse shape also influences the pattern of multi-pulse sidebands as well as the chosen sampling rate of the Z-spectrum.This knowledge is now applied again in vivo, where Figure 9 presents a Z-spectra of a voxel in ventricular CSF acquired with Simulation of the of pulse shapes on the pattern in the Z-spectrum using the delay time without intermediate gradient spoiling (A-C), with intermediate gradient spoiling (D-F), as well as with an additional simulated B 0 drift (G-I).Figures can be recreated via https://github.com/mrmeeseeksjr/Sideband_Example/tree/Sideband_MRM/matlabskripts/Fig5. the APTw protocol from the consensus paper 6 using sinc pulses for saturation with the a standard phase-cycling scheme.The data were acquired once with and once without interpulse gradient spoiling.The data confirm our previous in silico and in vitro results that the use of interpulse gradient spoiling leads to a visible reduction of sidebands (Figure 9A,B).Moreover, a pseudo-CEST effect at about 1.3 ppm emerges in the MTR asym spectra with low sampling rate (Figure 9C), which originates from the interplay of multi-pulse sidebands (see corresponding points in Figure 9A), the B 0 shift, and the low sampling.Spoiler gradients suppress these pseudo-CEST effects also for the low sampling rate (Figure 9D).Observing the Z-spectra for solid tissue, the sidebands are not clearly evident in WM (Figure 9E,F) but are slightly visible in GM (Figure 9G,H).However, the effects of gradient spoiling are less pronounced than in liquid tissue with higher T 2 times.

DISCUSSION
Z-spectra sidebands were previously reported in simulations and measurements by Schmitt et al. 4 These artifacts were described by an inherent symmetrical pattern on the Z-spectra, where the acquired Z-magnetization drops and rises abruptly over a sampled range between the saturated frequency offsets.The authors found that the appearance of sidebands was strongest for short saturation times,

F I G U R E 6
Phantom measurement without gradient spoiling and the continued phase-cycling scheme (A), with gradient spoiling using the continued phase cycling scheme (B), and with gradient spoiling (C) using quadratic phase cycling with growing 117 • phase increment.The spectra with gradient spoiling show no major differences, indicating that the choice of phase cycling can be neglected when gradient spoiling is switched on during the interpulse delay time.Figures can be recreated via https://github.com/mrmeeseeksjr/Sideband_Example/tree/Sideband_MRM/matlabskripts/Fig6.

F I G U R E 7
Simulation of the sideband artifact once with (left) and once without (right) interpulse gradient spoiling between the RF pulses.Although (A)-(F) show the development for increasing the number of RF pulses (1-5, 10), two main single-pulse sidebands near the resonance frequency of water are visible, which increase with each RF pulse.When interpulse gradient spoiling is turned off, each RF pulse suffers from the previous transversal magnetization, which leads to multi-pulse sidebands.Figures can be recreated via https://github.com/mrmeeseeksjr/Sideband_Example/tree/Sideband_MRM/matlabskripts/Fig7.which came along with a broadening of the FWHM, while the opposite behavior was observed for longer saturation times.In that work, the sideband effect was traced back to the power spectral density function of the saturation pulse train and attempted to be explained via the Fourier transform, which behaves inversely proportional to the saturation duration. 4e show herein that (i) phase cycling, (ii) B 0 , and (iii) gradient spoiling affect the sideband pattern, which indicates that a simple Fourier transform cannot describe all the observations.Even the single-pulse response required the typically large flip angles used in CEST, a full Bloch simulation of the nonlinear response, which cannot be described by the linear Fourier transform.
Especially the described multi-pulse sidebands can only be revealed in simulation and with match to measurements when the full pulse train is dynamically simulated, and when the RF phases are considered carefully to match between measurements and simulations.Thus, for a set of defining parameters of a CEST experiment, we suggest adding the interpulse gradient spoiling and phase-cycling scheme Φ i to nine parameters (B 1rms , t sat , DC, t p , t d , T rec , Simulation of different flip angles (FAs; 90 • , 180 • , 270 • , 360 • ) by modifying B 1 without (A-D) and with (E-H) B 0 drift as well as with B 0 drift and additional gradient spoiling (I-L).For simulation, a DeepCEST sequence was used for practical reasons, as it already provides a flip angle close to 180 • .The data reveal that multi-pulse sidebands cannot be avoided by only changing the flip angle, but lower B 1 lowers their influence range.Again, gradient spoiling helps to avoid multi-pulse sidebands, whereas single-pulse sidebands increase with B 1 .Figures can be recreated via https://github.com/mrmeeseeksjr/Sideband_Example/tree/Sideband_MRM/matlabskripts/Fig8.pulse shape, Φ i , G x/y/z ) that describe the saturation process completely using a reproducible standard such as Pulseq-CEST.
The fact that some sideband patterns can appear like noise or are hidden in a low sampling rate is an important guidance for validating CEST effects in vivo and in vitro.We therefore suggest testing different sampling rates before reducing them for reasons of time efficiency.The close match between Pulseq-CEST simulations and measurements (Figure 4) allows us to further study in silico the magnetization dynamics leading to sidebands, as described in detail in the Supporting Information.
This analysis helped us to understand that there are two types of sidebands: single-pulse sidebands that only depend on the pulse shape itself (Figure 5) and multi-pulse sidebands that build up over multiple pulses when a certain complex resonance condition is achieved.
These multi-pulse sidebands are not translations invariant in the frequency domain, which can hamper a correct B 0 correction.To avoid multi-pulse sidebands, we suggest adding interpulse spoiler gradients to avoid an enlargement of the magnetization cone, as suggested previously. 4As the multi-pulse sideband effects depend on non-decayed transverse magnetization of the previous pulse, these effects will also vanish for shorter T 2 times as in healthy GM or WM.
Still, pseudo-CEST effects can appear through sidebands in more pronounced liquid compartments.This also implies pathologically altered tissue with high T 2 times, such as cystic or necrotic tissue.Thus, wrong settings for pulsed CEST sequences (e.g., short pulse length, flip angle, phase cycling, sampling rate) can affect the evaluation of all CEST effects in the spectral range up to ±2 ppm (e.g., amine, guanidino, hydroxyl, glucose, myo-inositol). 17,18Therefore, a previous simulation of the pulsed-CEST sequence is always recommended to avoid artifacts.Furthermore, interpulse spoiler gradients are indicated, for which we Acquired in vivo Z-spectra in CSF with an increment of 0.05 ppm without (A) and with (B) interpulse gradient spoiling.(C,D) Reconstructed sampling rate of 0.25 ppm.Although the corresponding MTR asym in (A) is more disturbed through the sidebands, it also leads to a pseudo-CEST effect with lower sapling rate (C) around 1.5 ppm.In contrast, this becomes negligible when gradient spoiling is used (B,D).The sideband artifacts are less visible in white matter (WM; E), but noticeable in gray matter (GM; G).Therefore, gradient spoiling does not appear to have a strong impact as in more fluid tissue types.

F I G U R E 10
Simulated Z-spectrum of the amide proton transfer-weighted (APTw) sequence with 36 sinc pulses, t p = 50 ms, t d = 0 ms, and standard phase cycling without (A) and with (B) B 0 shift.The data suggest that the use of 100% duty cycle (DC) without the possibility of interpulse gradient spoiling still results in multi-pulse sidebands and pseudo-CEST effects in the presence of B 0 inhomogeneities.could confirm in vivo that they resolve multi-pulse sidebands.
However, the execution of gradient spoilers requires time, governed by the scanner hardware.Thus, we cannot proceed to the suggested 100% DC preparations, 19 as these would not allow the execution of gradients.It is conceivable that 100% DC preparations might not have problems regarding multi-pulse sidebands, as the system is always driven.However, the simulation in Figure 10, where the interpulse delay t d was adjusted to be 0, reveals that this is not the case.Because the sidebands have to be considered when aiming for 100% DC, it might be safer to use interpulse spoiler gradients between pulses and accept a slightly lower DC.
Although multi-pulse sidebands can be avoided by gradient spoiling, as shown in Figures 7 and 9, the single-pulse response can only be optimized by the individual pulse shape; here, recent approaches using spin-lock pulses 11,20 or optimal control theory 21 might lead to an improved single-pulse response.
Finally, for the mentioned APTw brain tumor consensus protocol, 6 we do not think that our findings lead to the necessity for taking action, as APTw CEST is typically evaluated at Δω = 3.5 ppm, and we did not see any significant sideband effects in this range.
Still, our results should be considered when evaluating offsets close to the resonance frequency of water for this protocol, such as guanidine (∼2 ppm) or hydroxyl groups (∼1.3 ppm) 1 as well as relayed nuclear Overhauser effect at −1.6 ppm. 22Furthermore, these artifacts should be taken into account in the B 0 correction using the water saturation shift referencing (WASSR) method, as the data points that need to be adjusted could be affected by the sidebands.

CONCLUSION
We provide the first detailed insights into sidebands occurring in pulsed CEST experiments and show that, like in imaging sequences, gradient and RF spoiling play an important role.Therefore, we recommend defining nine CEST parameters: B 1rms , t sat , DC, t p , t d , T rec , pulse shape, Φ i , and G x/y/z .Furthermore, we found that one can differentiate between single-pulse sidebands and multi-pulse sidebands, the latter leading to complex and non-B 0 -correctable patterns.To avoid them, gradient-spoiled pulse trains and high enough sampling rates are generally indicated to avoid misinterpretations of multi-pulse sidebands as CEST effects, especially in liquid environments or for CEST resonances close to water.These observations could be measured and replicated by simulations with Pulseq-CEST for different examples, which contribute to a better understanding of CEST pulse trains.