Robust cardiac T1ρ$$ {\mathrm{T}}_{1_{\boldsymbol{\rho}}} $$ mapping at 3T using adiabatic spin‐lock
 preparations

The aim of this study is to develop and optimize an adiabatic T1ρ$$ {\mathrm{T}}_{1\uprho} $$ ( T1ρ,adiab$$ {\mathrm{T}}_{1\uprho, \mathrm{adiab}} $$ ) mapping method for robust quantification of spin‐lock (SL) relaxation in the myocardium at 3T.


INTRODUCTION
Cardiac MRI is the clinical gold standard for the assessment of scar and fibrosis in ischemic and nonischemic heart diseases. 1-4 Late gadolinium enhancement (LGE) imaging can be used to differentiate between scar and healthy myocardium based on the retention of gadolinium-based contrast agents. 5 However, gadolinium-based contrast agent injection is contraindicated in patients with severe renal impairment due to the risk of necrotic systemic fibrosis. 6 In addition, gadolinium retention in the brain after injection of gadolinium-based contrast agents has been reported. 7 Thus, contrast-free alternatives are highly desired. Quantitative myocardial tissue characterization has emerged with a wide spectrum of applications in various cardiomyopathies. 8 Native T 1 mapping has been explored for the assessment of myocardial infarction (MI) without the need for contrast agents. [9][10][11] However, mixed results have been reported on its sensitivity to focal scar and the approach remains the subject of ongoing research. [12][13][14] T 1ρ mapping has been proposed as a promising non-contrast alternative for scar assessment, due to its increased sensitivity to slow molecular motion in the kilohertz range. 15,16 First, Muthupillai et al. reported stronger postcontrast enhancement in acute MI cases for T 1ρ -weighted imaging compared with conventional T 1 -weighted LGE imaging. 17,18 More recently, quantitative T 1ρ maps have demonstrated improved differentiation between infarcted and remote myocardium in swine models, compared with native T 1 and T 2 maps, yielding comparable contrast-to-noise ratio (CNR) to LGE images. 13,19,20 Similar results have been reported in mice [21][22][23] and monkeys. 24 In vivo T 1ρ mapping has been successfully applied in patients with ischemic and nonischemic cardiomyopathies at 1.5T. [25][26][27][28][29][30] Implementing T 1ρ mapping at 3T could further improve the diagnostic value of this approach, due to an increase in signal-to-noise ratio and CNR, and the applicability in a growing number of 3T cardiac examinations. However, at 3T, only a few studies have been reported, [31][32][33] highlighting limitations related to system imperfections and the specific absorption rate (SAR) at high field strengths.
Conventional T 1ρ maps are obtained using spin-lock (SL) preparation pulses with various durations, which are most commonly based on continuous-wave radiofrequency (RF) irradiation. These preparations are inherently susceptible to B 0 and B + 1 field inhomogeneities. 34,35 To compensate for these inhomogeneities, continuous-wave SL pulses, in combination with refocusing pulses and phase cycling of SL modules, have been proposed. 34,36,37 An alternative strategy to achieve resilience against system imperfections is the use of adiabatic pulses. 38 The robustness of adiabatic pulses against field inhomogeneities has been studied in other 3T cardiac MRI methods, such as inversion-recovery T 1 mapping 39 or refocusing in T 2 preparations. 40 Recently, similar adiabatic pulses have also been employed for refocusing in conventional SL preparations for cardiac T 1ρ mapping 29 at 1.5T. Alternatively, SL preparations consisting of trains of adiabatic full passage pulses have been proposed to generate T 1ρ contrast in other anatomies. 41,42 During the adiabatic full passage frequency sweep, the magnetization is locked along the effective field. This induces T 1ρ,adiab as the dominant relaxation mechanism during the pulse application. 43,44 T 1ρ,adiab will be used throughout the manuscript to indicate the rotating frame of reference relaxation constant measured by adiabatic preparations.
In this work, we sought to investigate the use of fully adiabatic SL (aSL) preparations for T 1ρ,adiab mapping of the myocardium at 3T. Bloch simulations were performed to optimize aSL pulse shapes for resilience against system imperfections. Phantom and in vivo imaging of the calf muscle were then carried out to compare aSL preparations against fully compensated conventional SL preparations. In vivo performance was shown with cardiac mapping in healthy subjects. Finally, clinical feasibility was evaluated in a small proof-of-principle cohort of patients.

Adiabatic SL preparation design
In this work, aSL preparations were based on a train of adiabatic full passage pulses with an identical duration ( Figure 1B). An even number of pulses was used to ensure that, at the end of the preparation (t = SL ), the magnetization M( SL ) was stored along the +z direction. Hyperbolic secant (HS) pulse shapes were employed, as commonly used in other imaging applications. [39][40][41]45,46 These are characterized by the following amplitude and frequency modulation functions: Here B 1 (t) represents the pulse amplitude, B max 1 the peak amplitude, and a constant that characterizes the width of the pulse bell. The single HS pulse duration is indicated by HS . Δ 1 (t) = dΦ 1 (t)∕ dt is the frequency modulation with respect to the Larmor frequency 0 , where Φ 1 (t) represents the pulse phase as a function of time, and 2f max is the amplitude of the frequency sweep. The polarity of the frequency sweep was alternated between consecutive HS pulses to compensate for residual pulse imperfections.
Preparations with variable SL durations were achieved by concatenating identical pulse modules multiple times. The total duration of a single aSL module ( SL ) was fixed to 60 ms. This value was chosen as a trade-off between adequate sampling of the expected range of in vivo T 1ρ,adiab times and restrictions imposed by the SAR limits (whole-body SAR <2.0 W/kg) and the RF amplifier chain. To obtain constant preparation times when changing the pulse duration ( HS ), modules containing 2, 4, or 8 HS pulses (2HS-aSL, 4HS-aSL, and 8HS-aSL) with relative pulse duration SL ∕2, SL ∕4, and SL ∕8, were implemented. For SL modules with 4 and 8 HS pulses, phase cycling was adopted between pairs of HS pulses to achieve a full Malcolm-Levitt (MLEV) scheme compensation. 47

Bloch simulations
Bloch simulations were used to optimize , f max , and HS in the aSL preparations. All simulations were performed in MATLAB (MathWorks).
The preparation efficiency was determined as M z ( SL )∕M(0) and used as a metric to optimize the design of the aSL module. The aSL preparation modules were simulated using the maximum RF pulse power, within the limits imposed by the peak B + 1 (B max 1 = 13.5 μT) and SAR (whole-body SAR <2.0 W/kg). The preparation efficiency was averaged over a design window covering the expected range of in vivo off-resonances (Δ off 1 ∈ {−150, −149, ... + 150} Hz) and B + Two sets of optimizations were performed to identify the optimal pulse duration and amplitude/frequency modulation functions, respectively. First, the 2HS-aSL, 4HS-aSL, and 8HS-aSL modules were compared in terms of preparation efficiency. Then, the module that produced the highest preparation efficiency was selected to derive the optimal values of and f max . Bloch simulations covering the range of expected in vivo variability of B 0 and B + 1 were performed to obtain optimized pulses for three design regions: (1) original balanced design region (Bal-aSL) ( For each SL module, Δ off 1 and 1 , two snap-shot balanced steady-state free-precession (bSSFP) images were acquired: one preceded by the aSL preparation ( SL = 60ms) and one with no preparation. The two scans were interleaved by a 5-s pause to allow longitudinal magnetization recovery. Low imaging resolution was used (10 × 10 × 10 mm 3 ), with pulse repetition time = 1.9 ms, echo time = 0.72 ms, flip angle = 90 • and a SENSE factor of 2. The preparation efficiency M z ( SL )∕M(0) was then calculated as the ratio of the two magnitude images. Signal polarity was restored using the corresponding phase images prior to further processing. In phantoms, the entire phantom area was evaluated, while in the calf, manually drawn circular regions of interest (ROIs) were used.

T 1 mapping
The proposed T 1ρ,adiab mapping approaches were compared to each other and to a conventional, continuous-wave T 1ρ mapping implementation in phantom and through in vivo experiments in the calf muscle and in the myocardium for healthy subjects and patients. Phantoms and healthy subjects were scanned on a 3T Ingenia system (Philips). Patient data were acquired on a 3T Achieva system (Philips). In vivo imaging was ethically approved by the competent review authorities (METC NL73381.078.20, UK National Research Ethics Service 15/NS/0030). Written informed consent has been obtained prior to all imaging sessions according to institutional guidelines.
The aSL preparations were compared to a fully balanced non-aSL pulse 37 (RefSL in Figure 1A). Three phase-cycled SL blocks with equal amplitude and durations of SL ∕4, SL ∕2, and SL ∕4, respectively, were played. The SL amplitude was chosen based on the RF amplifier constraints as B + 1 ∕ = 300 Hz. T 1ρ and T 1ρ,adiab mapping was performed using a cardiac triggered breath-hold sequence ( Figure 2). Five baseline single-shot bSSFP images were acquired: the first with no SL preparation, then three with increasing SL durations, and finally a saturation-prepared image used to approximate infinite SL length. 51 A composite "Water suppression Enhanced through T1-effects" pulse was used to achieve robust saturation in the presence of field inhomogeneities. 52 Total preparation durations were SL = 0, 60, 120, 180 ms for aSL modules. Shorter preparations were employed for RefSL ( SL = 0, 12, 24, 36 ms) to account for higher SAR levels, heavier RF amplifier load, and significantly shorter nonadiabatic T 1ρ times. Scans were acquired in the end-diastolic phase. All images, except the saturation-prepared image, were preceded by a pause to allow for longitudinal magnetization recovery. Other imaging parameters were: in-plane resolution = 2 × 2 mm 2 , field of view = 220 × 220 mm 2 , slice thickness = 8 mm, echo time/pulse repetition time = 1.2/2.4 ms, flip angle = 70 • , SENSE = 2.
T 1ρ and T 1ρ,adiab maps were reconstructed in MATLAB using the following three-parameter model, 51 to account for the effect of the imaging pulses:

Phantom and in vivo calf experiments
The T1MES phantom was used for phantom experiments to mimic blood and myocardium relaxation times at 3T. 53 Approximate T 1 and T 2 times of the phantom vials were estimated, using a MOLLI sequence for T 1 and a Gradient Spin Echo sequence for T 2 . To study repeatability, ten repetitions of T 1ρ and T 1ρ,adiab mapping scans were acquired for each preparation (B 0 -aSL, Bal-aSL, B 1 -aSL and RefSL). Manually drawn circular ROIs were used to extract T 1ρ and T 1ρ,adiab values for further processing. Repeatability was assessed using the coefficient of variability (CV): where N v is the number of samples, corresponding to the number of vials in this case, and wCV i is the coefficient of variability within the sample computed for every vial as Here, R = 10 represents the number of repetitions, i,r is the average T 1ρ or T 1ρ,adiab value for each vial i and F I G U R E 2 (A) T 1ρ mapping sequence diagram with (B) corresponding baseline images from a representative healthy subject. Five images are acquired, one without preparation, three with different T 1ρ,adiab preparations ( SL = 60, 120, 180ms), and one with saturation preparation, to allow for accurate mapping of the induced T 1ρ relaxation.
repetition r and i is the average T 1ρ or T 1ρ,adiab value for each vial across all repetitions.
In a second experiment, T 1ρ,adiab time was assessed as a function of the HS shape parameter by acquiring phantom and calf T 1ρ,adiab maps for ∈ 1, 2, … 10. For each , a constant sweep amplitude f max value was acquired. The dependence between the parameter and the measured T 1ρ,adiab values was tested using linear regression. R 2 coefficient, slope, and intercept values were reported for a single exemplary vial and a manually drawn circular calf ROI.

Healthy subjects experiments
The proposed aSL preparations were tested in six healthy subjects (four males, two females, 21.5 ± 1.9 years old). For each subject, B 0 -aSL, Bal-aSL, and B 1 -aSL T 1ρ,adiab maps were acquired in three short-axis (SAX) slices (basal, mid, and apical) and a four-chamber view. To assess reproducibility, the 12 maps were re-acquired following the repositioning of the subject. 54 In this cohort of healthy subjects, the magnetization recovery pause was 2.5 s to limit the total scan time to 13 s. In a second cohort of seven healthy subjects (five males, two females, 24.7 ± 2.5 years old), the best-performing aSL preparation was compared to RefSL. Similarly to the first cohort, three SAX slices and a four-chamber view were acquired for each subject and preparation. Here, a magnetization recovery pause of 3.5 s was employed to avoid relaxation time over-estimation (see Figure S1). To assess robustness to B 0 and B + 1 inhomogeneities, a second repetition of each map was acquired by moving the shimming volume only on the right ventricle, while keeping the position of the patient fixed.
The myocardium was automatically segmented using the nnU-Net framework 55 with uncertainty estimation. 56 Segmentation maps with predictive confidence below 75% were discarded and the segmentation was performed manually. The average values of T 1ρ or T 1ρ,adiab and their corresponding SD values (SD) in the segmented myocardium were extracted according to the American Heart Association (AHA) 16 segment model. A group-wise ANOVA test followed by paired t-tests was used to assess statistical differences between the T 1ρ and T 1ρ,adiab times with different preparations.
T 1ρ and T 1ρ,adiab quantification precision was assessed for each myocardial segment and SL module through the within-subject coefficient of variability (wCV): computed for every repetition r and subject i, where and are the T 1ρ or T 1ρ,adiab mean and SD, respectively. Then, the mean and SD of T 1ρ or T 1ρ,adiab values across repetitions were computed as: and, therefore, the reproducibility as: where R = 2 indicates the number of repetitions. Finally, the inter-subject variability was computed as a summary of the deviation of each subject's average T 1ρ or T 1ρ,adiab value from the overall mean: and N s indicates the number of subjects. Statistical differences between the different SL preparations in terms of precision and reproducibility were investigated using a group-wise Kruskal-Wallis test and subsequently right-tailed pair-wise Mann-Whitney U-tests.

Patients experiments
Clinical feasibility was tested in a small proof-of-principle cohort of six patients (two males, four females, 50.2 ± 11.0 years old) referred to clinical CMR. All patients were imaged using standard clinical protocols, including MOLLI-based native T 1 mapping, LGE imaging and CINE scans.
LGE imaging was performed 10-15 minutes after injection of 0.15 mmol/kg of Gadobutrol (Gadovist, Bayer Schering). Extracellular volume (ECV) maps were estimated from native and postcontrast T 1 values. Synthetic hematocrit values were computed for each patient as Hct = 0.88 − (T 1,blood ∕3240). 57 The proposed T 1ρ,adiab mapping sequence and conventional T 1ρ mapping of a single mid-ventricular SAX slice were included in the scan protocol prior to contrast administration. Imaging parameters were chosen to closely match those used in the healthy subjects. PCA-based group-wise registration was used to mitigate residual cardiac and respiratory motion for baseline T 1ρ,adiab and T 1ρ images. 58 T 1 and T 1 (,adiab) baseline images were spatially co-registered to the corresponding LGE images applying a PCA-based group-wise method to the baseline images. 58 Finally, the resulting deformation matrices were transferred to the previously reconstructed maps. Manually drawn ROIs were defined on LGE images and then superimposed on the co-registered quantitative maps to extract scar and remote T 1 , ECV, T 1ρ , and T 1ρ,adiab times.

Bloch simulations
The simulated preparation efficiency achieved with the 2HS-aSL, 4HS-aSL, and 8HS-aSL preparations is shown in Figure 3A. For all three modules, the highest preparation efficiency was obtained for low to intermediate frequency sweep amplitudes and showed an inversely proportional relationship with the parameter . However, very low values of required a reduction of the pulse peak power to satisfy SAR limitations. In all three cases, the optimal region is well-defined and separated from the non-adiabatic region at high sweep velocities (top-right corner). Overall, 2HS-aSL shows higher overall preparation efficiency than 4HS-aSL and 8HS-aSL. The 2HS-aSL module also presents a larger optimal region, indicating higher stability to the choice of parameters Simulation results for 2HS-aSL preparation with three different design regions are shown in Figure 3B. For B 0 -aSL and B 1 -aSL, similar patterns to the previously analyzed Bal-aSL case ( Figure 3A) can be observed, with an inversely proportional relationship with the parameter . The optimal region becomes narrower when using a more B + 1 compensated preparation, with overall decreasing optimal values and f max . Optimal values of { , f max } were identified as {6.9, 450 Hz } for B 0 -aSL and {4.4, 200 Hz } for B 1 -aSL, yielding an average efficiency M z ∕M 0 of 0.99 and 0.94, respectively. A summary of the parameters used for the optimized aSL preparations can be found in Table 1. Figure 3C illustrates how the preparation efficiency M z ( SL )∕M(0) varies over a range of off-resonant frequencies and B + 1 inhomogeneities for the optimized B 0 -aSL, Bal-aSL and B 1 -aSL modules according to Bloch simulations. The corresponding design region used for the parameter optimization of each preparation is marked by the dashed rectangle. For all three aSL modules, the regions characterized by low preparation efficiency (in blue) are outside the design region.

Phantom and in vivo calf experiments
The experimental preparation efficiency measured in the phantom experiments with varying Δ off

T A B L E 1
Adiabatic spin-lock preparations design parameters. conditions is depicted in Figure 4A. Good agreement between the simulated and experimental results can be observed. Broad areas of lower preparation efficiency are present for intermediate to low 1 values with B 0 -aSL, low to very-low 1 values with Bal-aSL, and very low 1 as well as high absolute Δ off 1 values with B 1 -aSL. The results in terms of preparation efficiency obtained in vivo in the calf muscle of a healthy subject are shown in Figure 4B. These results are in good agreement with both simulations and phantom data. In vivo preparation efficiency is compromised for 1 < 0.6 with the B 0 -aSL module, while no substantial degradation was observed over the entire off-resonance range studied. On the opposite side, B 1 -aSL yields robust preparation efficiency for 1 values down to 0.2, but lower efficiency for |Δ off 1 | >150 Hz. The overall efficiency score measured in the phantom F I G U R E 4 (A) Experimental preparation efficiency measured in phantoms for a range of Δ off 1 and 1 with three adiabatic spin-lock (aSL) preparations. Experimental results were in agreement with simulations in Figure 3C, minus a scaling factor given by relaxation, which was ignored in simulations. (B) Adiabatic preparations efficiency was measured in vivo on a healthy subject's calf muscle for the same range of Δ off 1 and 1 . Overall, the results were in good agreement with the phantom experiments (A) and the numerical simulations ( Figure 3C). Representative calf T 1ρ,adiab maps for different values of Δ off 1 and 1 illustrate the variation in image artifacts. and calf experiments is lower than in simulations, as no relaxation contributions have been simulated. Complete T 1ρ and T 1ρ,adiab mapping results for the T1MES phantom can be found in Figure S2. Improved repeatability was observed (p < 0.05) in T 1ρ,adiab maps (wCV i = 0.29 ± 0.15% for B 0 -aSL, p < 0.01; wCV i = 0.23 ± 0.13% for Bal-aSL, p < 0.01; wCV i = 0.21 ± 0.11% for B 1 -aSL, p < 0.001) with respect to conventional T 1ρ maps (wCV i = 1.30 ± 1.34% for RefSL).

Pulse shape
In Figure 5, examples of phantom and calf T 1ρ,adiab maps acquired with different values are displayed. T 1ρ,adiab values increase with an approximately linear trend for higher in both cases (R 2 = 0.99, slope = 9.56, intercept = 32.15 for phantoms, R 2 = 0.91, slope = 12.46, intercept = 26.53 for the calf). A higher deviation from linearity was observed in the calf values for ∈ {3, 4, 5}.

3.3
Healthy subjects experiments Figure 6A shows mid-ventricular SAX and four-chamber T 1ρ,adiab maps for one representative subject, displaying overall strong myocardium-to-blood contrast. No major off-resonance or B + 1 artifacts are visually apparent on the T 1ρ,adiab maps. In agreement with phantom and calf results, myocardial T 1ρ,adiab values obtained with the B 0 -aSL preparation ( = 6.9) are higher than those obtained with the Bal-aSL preparation ( = 5.5), which in turn are higher than those obtained with B 1 -aSL preparations ( = 4.4). Myocardial T 1ρ,adiab values averaged over slices, segments, and subjects were 194.22 ± 24.54 ms, 155.59 ± 18.09 ms, and 87.48 ± 11.55 ms for B 0 -aSL, Bal-aSL, and B 1 -aSL, respectively. The bull's-eye plots in Figure 6B show that the inter-subject average T 1ρ,adiab values for all three aSL preparations are homogeneous across all segments. Bal-aSL and B 1 -aSL bullseye plots depict lower T 1ρ,adiab values in the apical slice (apical vs. basal slice: −2.64%, p < 0.001 for Bal-aSL, −6.62%, p < 0.001 for B 1 -aSL) but not for B 0 -aSL (−0.97%, p = 0.12). Figure 6C depicts good reproducibility across the 16 AHA segments for all aSL preparations. Trends of improved precision and reproducibility were observed for B 0 -aSL compared with B 1 -aSL, but the differences were not significant (p > 0.08). However, B 0 -aSL yielded significantly lower inter-subject variability than B 1 -aSL (p < 0.05).
B 0 -aSL T 1ρ,adiab and RefSL T 1ρ maps obtained in two repetitions under different shimming conditions for a representative subject are shown in Figure 7. RefSL preparations yield lower T 1ρ values than B 0 -aSL (average T 1ρ over subjects, slices and segments = 38.21 ± 14.37 ms for RefSL, compared with 183.28 ± 25.53 ms for B 0 -aSL, Figure 8A). RefSL-based T 1ρ maps display pronounced artifacts over large portions of the myocardium and poor reproducibility across the shimming conditions. B 0 -aSL preparations, on the other hand, present comparable image quality for both cases, free of visually apparent artifacts. The adiabatic B 0 -aSL preparation resulted in significantly better precision compared with RefSL (B 0 -aSL: wCV i,r = 14.51 ± 3.71%, RefSL: wCV i,r = 37.61 ± 19.42%; p < 0.01, Figure 8C).
At least 10 times higher reproducibility was obtained with the B 0 -aSL preparation compared with the RefSL module (average wCV i = 4.64 ± 2.18% for B 0 -aSL against average wCV i = 47.39 ± 12.06% for RefSL, p < 0.0001), as shown in Figure 8C.
A complete overview of the in vivo myocardial T 1ρ,adiab and T 1ρ values, as well as precision, reproducibility, and inter-subject variability for each healthy subject across the two cohorts, can be found in Tables S1-S4.

Patients experiments
Four of the six patients presented as LGE-positive in the CMR. For two of those four patients, the mid-SAX slice intersected with the area of focal scar identified on the LGE images. Figure 9 shows the clinical sequences as well as aSL-based T 1ρ,adiab maps and RefSL-based T 1ρ maps for the two subjects with LGE-identified scars in the mid-ventricular SAX slice. T 1ρ,adiab maps show visually discernable alteration in the myocardium, that spatially coincides with the areas of hyper-enhancement in the LGE images. Any potential alteration in the RefSL-based T 1ρ maps is obfuscated by the presence of substantial artifacts. B 1 -aSL yielded the best maps quality among adiabatic preparations, with no visible B 0 or B + 1 -related artifacts. B 0 -aSL and Bal-aSL maps were characterized by overall lower quality and presented visible artifacts across the myocardium, as shown in Figure S3.

F I G U R E 6
(A) Mid SAX and four-chamber T 1ρ,adiab maps obtained with B 0 -aSL, Bal-aSL, and B 1 -aSL preparations in a representative healthy subject of the first cohort. T 1ρ,adiab maps achieved good visual map quality, with a homogeneous myocardium and clear delineation against the blood pool across all acquired slices. (B) Bull's-eye plots showing the T 1ρ,adiab values, averaged over all subjects and repetitions, for 16 AHA myocardial segments. T 1ρ,adiab values are homogeneous across the 16 segments for all preparations. Average T 1ρ,adiab increases with increasing beta . (C) Bullseye plots report the average reproducibility (wCV) coefficients, measured over two acquisitions interleaved by subject repositioning, for adiabatic spin-lock (aSL)-prepared maps in 16 AHA myocardial segments. Global average values are reported at the center of each bullseye plot. A mild improvement in reproducibility is observed for B 0 -aSL and Bal-aSL preparations, compared to B 1 -aSL, but the difference was not statistically significant (p > 0.05).
Patient 1 shows near transmural enhancement in the LGE images. T 1ρ,adiab in this subject shows a +47.48% elevation in the LGE-positive area compared with the remote myocardium for B 1 -aSL, while RefSL-based T 1ρ maps show a −33.26% difference. In comparison, native T 1 and ECV values for the same patient showed, respectively, +17.12% and +80.53% in the LGE-positive area. In patient 2, who showed signs of lipomatous metaplasia in bSSFP CINE images (Figure 9), decreased relaxation times were measured for the LGE positive area, compared with remote healthy myocardium (−6.04% for B 1 -aSL T 1ρ,adiab , −67.19% for RefSL T 1ρ , −3.11% for native T 1 , −41.19% for ECV), as expected in the presence of fatty infiltration. Major

DISCUSSION
In this study, we proposed a new cardiac T 1ρ,adiab mapping technique based on fully aSL preparation for myocardial tissue characterization at 3T. Numerical optimization yielded aSL preparations with tuneable resilience

F I G U R E 7
Apical, mid, and basal short-axis (A) B 0 -aSL-prepared T 1ρ,adiab maps and (B) RefSL-prepared T 1ρ maps in a representative healthy subject. Two repetitions of each slice and preparation were acquired with different shim volumes: one covering the entire heart, the other covering only the right ventricle. T 1ρ,adiab maps retain comparable map quality across repetitions with a nearly identical visual appearance of the maps. RefSL maps depict significant artifacts degrading the map quality in the myocardium, particularly in the second repetition. against B 0 and B + 1 inhomogeneities. Phantom and in vivo measurements demonstrated that T 1ρ,adiab mapping achieved more robust results than conventional T 1ρ mapping approaches. T 1ρ,adiab maps showed fewer artifacts, higher precision and reproducibility, and lower inter-subject variability. Initial data showed feasibility in patients and visual alignment of areas with altered T 1ρ,adiab and hyper-enhancement in LGE images.
Conventional T 1ρ values obtained with the RefSL preparation in this study were comparable to those reported in previous studies at 3T. [31][32][33] However, our results show slightly lower precision for the RefSL maps than in previous studies. This difference in variability may be because previous studies only evaluated a small ROI in the anteroseptal segment of the myocardium, while in this work, an automatic segmentation of the entire myocardium was used. Significant inhomogeneities are visible in conventional RefSL maps, both in our results and in other studies. [31][32][33] Han et al. found that at 1.5T B 0 variations over 10% of the SL field amplitude (typically B 1 ∕ = 500 Hz) cause T 1ρ quantification errors and visible image artifacts. 35 At 3T, this limit is easily exceeded. 39 Furthermore, B + 1 inhomogeneities have a much higher impact at high fields in cardiac imaging, 48 thus necessitating more robust T 1ρ mapping techniques.
Both adiabatic and conventional T 1ρ maps showed lower T 1ρ,adiab or T 1ρ values in the apical slice, compared to the mid and basal slices. This effect is less evident for the B 0 -aSL preparations (T 1ρ,adiab values comparison apical vs. mid and basal slices: p = 0.77 for B 0 -aSL, p < 0.01 for B 1 -aSL and Bal-aSL, Figure 6). Hence, the lower T 1ρ,adiab and T 1ρ values in the apical slice may be explained with the higher contribution of B 0 inhomogeneities at the apex.
Using fully aSL preparations has four major advantages. First, they yield more robust T 1ρ,adiab quantification in the presence of field inhomogeneities. Our results have shown that the T 1ρ,adiab maps have a lower level of noise and do not present significant B 0 -or B + 1 -related artifacts, overcoming the limitations observed in the previous studies. 31-33 T 1ρ,adiab preparations also yielded higher F I G U R E 8 (A) Bull's-eye plots showing the T 1ρ,adiab and T 1ρ values, averaged over all cohort two subjects and repetitions, for 16 AHA myocardial segments. T 1ρ,adiab values are consistently higher, but more homogeneous across the 16 segments for all preparations, compared with RefSL-based T 1ρ values. (B) Bullseye plots report the average precision (wCV), reproducibility (wCV), and inter-subject variability (CV) coefficients for B 0 -aSL-based T 1ρ,adiab maps and RefSL T 1ρ maps in 16 AHA myocardial segments. Global average values are reported at the center of each bull's-eye plot. Improved precision, reproducibility, and inter-subject variability are obtained with aSL preparations, compared to RefSL. (C) Bar plots comparing precision, reproducibility, and inter-subject variability for each preparation per slice and averaged across all slices (A = apical, M = mid-ventricular, B = basal, o = overall). Pair-wise statistical significance is marked by * or * * and the corresponding p-values are shown on top of each plot. Significantly higher wCV r,i , wCV, and CV values are measured for conventional RefSL-based T 1ρ mapping compared with T 1ρ,adiab . precision, reproducibility, and lower inter-subject variability. Resilience to artifacts is of particular importance for applications at high field strengths, like 3T, which have the potential advantage of increased signal-to-noise ratio and CNR. Second, the use of amplitude-modulated HS pulses lowers the SAR demands compared to conventional continuous-wave preparations for the same duration. Wang et al. reported a SL pulse amplitude B 1 ∕ of 298 Hz, 33 limited by SAR constraints and comparable with our findings. Low SL pulse amplitudes result in lower measured T 1ρ values and further compromise the CNR and robustness to system imperfections. The aSL pulses used in this study, on the other hand, allowed us to use maximum peak power and longer preparation times, while still satisfying SAR limitations. Third, T 1ρ,adiab preparations eliminate the need for the initial 90 • tip of the magnetization, which introduces further imperfections in the presence of B + 1 inhomogeneities. 42,51 Finally, conventional SL preparations are orientation dependent. 59 The high anisotropy of myocardial fibers yields orientation-dependent T 1ρ times with conventional preparations. 60 Adiabatic T 1ρ preparations, on the other hand, have been shown to be orientation independent. 59 This may further contribute to more homogeneous and reproducible T 1ρ,adiab maps across the myocardium.
Besides the advantages in terms of robustness given by aSL preparations, the mechanism behind T 1ρ,adiab relaxation is intrinsically different from conventional T 1ρ . Each T 1ρ,adiab preparation probes a wider spectrum of SL frequencies through the adiabatic sweep, compared to mono-frequency conventional SL. Effective field strength and orientation vary during aSL preparations, as well as

F I G U R E 9
(A) 53-year-old female patient suffering from ischemic cardiomyopathy. LGE images demonstrate myocardial infarction in the mid anteroseptal and anterior wall. A circular ghosting artifact is visible on the LGE image and partially overlaps with myocardial scar (red arrow). The B 1 -aSL-based T 1ρ,adiab map shows elevation co-localized with LGE positive regions (T 1ρ,adiab = 146.24 ± 25.34 scar -black arrow-, 99.40 ± 11.58 remote). Native T 1 and ECV values are also focally elevated in the anterior and anteroseptal segments. Due to changes in the imaging slice position, however, the visible abnormalities in native T 1 maps are not precisely co-localized with the LGE positive area. Due to mapping inhomogeneity in the anterior and lateral regions (yellow arrows), no focal alteration is unambiguously identified in the conventional T 1ρ maps. (B) 59-year-old male patient with a history of ischemic cardiomyopathy. LGE images demonstrate transmural myocardial enhancement in the mid-anterolateral segments (red arrow). Chemical shift artifacts in the bSSFP CINE images indicate lipomatous metaplasia. T 1ρ,adiab values decrease in the scar region (T 1ρ,adiab = 67.06 ± 14.69 scar -black arrow-, 96.57 ± 15.03 remote). Native T 1 and ECV values are also lower in correspondence of the scar region with respect to remote myocardium, although major artifacts impair the quality of the maps. In this patient, significant artifacts obfuscate any potential focal alteration in the RefSL-based T 1ρ maps (yellow arrows). the angle between the effective field and the magnetization. These variations lead to relaxation rate changes throughout the preparation module, rather than sampling a uniform T 1ρ . 43,44 On the other hand, the variable transverse relaxation T 2ρ contribution in the rotating frame of reference results in different T 1ρ /T 2ρ ratios for any given time point. Furthermore, we observed higher T 1ρ,adiab times for preparations with higher and, thus, a faster frequency sweep velocity. This indicates that the spectrum of relaxations rates probed during aSL varies depending on the pulse profiles. These factors may lead to a different sensitivity profile in pathological remodeling and its clinical value remains to be evaluated. An in-depth theoretical analysis of the mechanisms behind T 1ρ,adiab relaxation would be beneficial for the comprehension of its relationship with the underlying physiology.
In patients, the poor resilience of RefSL preparations to system imperfections significantly compromised the map quality. Artifacts in the area around the coronary sinus, as well as the lateral wall, appeared in all cases, preventing the unambiguous identification of focal alteration. Compared to healthy subjects, image artifacts were substantially more pronounced in the patient cohort. This likely stemmed from lower B + 1 shim quality in the clinical setting. aSL-based preparations, in particular when tuned for B + 1 -resilience, yielded good map quality, comparable to the healthy subject cohort. This indicates fair resilience to system imperfections in clinical use.
Cardiac T 1ρ,adiab maps showed visible focal alteration that spatially coincided with areas of hyper-enhancement in the LGE images. This is in line with previous studies indicating sensitivity to a range of diseases. Wang et al. found a +24% T 1ρ elevation for hypertrophic cardiomyopathy patients with diffuse fibrosis. 33 At 1.5T, van Oorschot et al. measured +52% T 1ρ elevation in infarcted myocardium of patients suffering from ischemic heart disease and +46% in a second ischemic cohort. 27 Furthermore, Bustin et al. have found a +40% elevation in infarcted myocardium of LGE-positive patients. 29 These trends are in agreement with the T 1ρ,adiab enhancement measured in patient 1. On the other hand, T 1ρ,adiab in the scar area of patient 2 was decreased. This is in good agreement with the CMR finding of lipomatous metaplasia, [61][62][63] and expected in these cases due to the short T 1ρ,adiab component of the intramyocardial fat. Our preliminary results indicate that fully adiabatic T 1ρ mapping can potentially yield more robust quantification than conventional continuous-wave SL in clinical use at high fields. However, clinical sensitivity of T 1 ,adiab mapping may differ from conventional continuous wave T 1ρ mapping due to the mechanistic differences and among different adiabatic preparations due to differences in the effective and fictitious fields. Consequently, larger dedicated cohorts of healthy controls and a targeted patient population are warranted to determine clinical sensitivity and potential cut-off values for the differentiation of healthy and infarcted myocardium.
Pulse design optimization was the key to achieving the desired resilience against B 0 and B + 1 inhomogeneities. The HS pulse shape was chosen specifically for its enhanced resilience to B 0 inhomogeneities, superior to TANH/TAN pulses, as previously reported. 39 First, we observed that shorter aSL pulses (4HS-aSL and 8HS-aSL) performed worse than the longer one (2HS-aSL) despite allowing for complete Malcolm-Levitt compensation. Longer HS pulses are thus preferred for T 1ρ,adiab preparations. Second, we found that the optimal HS pulse shape varies significantly under different B 0 and B + 1 conditions. Bloch simulations were in very good agreement with the experimental data acquired in both the phantoms and the calf muscle. Our in vivo results show that B 0 -aSL preparations achieve better precision and inter-subject variability than Bal-aSL and B 1 -aSL in healthy subjects. However, B 1 -aSL has proven most robust in the clinical setup where B 1 -shim quality was reduced.
Increased wCV r,i , wCV i , and CV values were observed in the basal and mid-inferolateral segment, as well as the apical lateral segment for B 1 -aSL preparations (see Figure 8). These values were reflected in the B 1 -aSL T 1ρ,adiab maps, which, for some subjects, presented residual B 0 artifacts in the same segments (Figure 7). These effects were not observed for B 0 -aSL and Bal-aSL maps. Thus, depending on the application and the technical characteristics of the scanner either of the optimized preparations may be most suitable for robust T 1ρ,adiab quantification in the clinic. Adiabatic pulses that were previously used for other cardiac MRI applications were found to be closest to those used for B 1 -aSL preparations ( = 4.8, f max = 215 Hz 40 ). These pulses may be particularly warranted on systems where B 1 quality is the main concern, such as systems with a single transmit channel or a lack of advanced shim modes. On other systems, B 0 -aSL and Bal-aSL preparations may be preferred for the observed increase in precision and reproducibility.
In our study, patient scans showed pronounced cardiac and respiratory motion, despite cardiac triggering and breath-holding. Residual motion due to heart rate variability and poor breath-holding capacity in patients rendered retrospective image registration necessary to achieve satisfactory map quality in the final T 1ρ,adiab and T 1ρ maps. Recently, specific attention has been dedicated to the development of accelerated, free-breathing, whole-heart T 1ρ mapping sequences to facilitate its clinical implementation. 28,64,65 Furthermore, several motion correction approaches have been proposed to improve the quality of reconstructed T 1ρ maps and mitigate the contribution of motion. 26,29 These efforts are key to enabling the widespread use of quantitative parametric mapping sequences in clinical practice. Our aSL preparations are fully compatible with these sequence designs and reconstruction approaches and could, in the future, be integrated into accelerated and motion-corrected T 1ρ mapping sequences. This may be particularly helpful to facilitate testing of the proposed T 1ρ,adiab mapping in large, relevant patient cohorts in order to demonstrate its clinical value.

CONCLUSIONS
In this work, T 1ρ,adiab mapping was proposed as an alternative to conventional T 1ρ mapping to enable its application in the human myocardium at 3T. Our results show that aSL preparations enable more robust mapping in the presence of B 0 and B + 1 inhomogeneities while satisfying SAR limitations. Adiabatic preparation modules yielded quantification with high precision and reproducibility in healthy subjects. In patients, aSL-based T 1 ,adiab maps depicted focal alterations in agreement with the reference LGE scans. Thus, T 1ρ mapping can be a promising candidate for reproducible myocardial tissue characterization and bears potential as a contrast-free imaging biomarker for scar and fibrosis.

SUPPORTING INFORMATION
Additional supporting information may be found in the online version of the article at the publisher's website. Figure S1. Phantom T 1ρ,adiab maps acquired with different rest periods for longitudinal magnetization recovery delays. T 1ρ,adiab values (± SD) reported in the plot are measured from the normal myocardium-mimicking vial (left column, middle row). For longitudinal magnetization recovery delays ≥ 3000 ms, the measured T 1ρ,adiab values deviate less than 5% from the asymptotic value. Figure S2. (A) Example of T 1ρ,adiab and T 1ρ maps of the tissue-mimicking T1MES phantom. Good map quality was achieved with aSL preparations, whereas visible artifacts are apparent in most vials in the maps obtained with RefSL preparation. Approximate T 1 and T 2 maps are displayed for reference. (B) T 1ρ,adiab and T 1ρ values with SD bars for each vial, averaged over 10 repetitions. T 1ρ,adiab values are consistently higher than T 1ρ values measured with RefSL preparations. T 1ρ,adiab dispersion is observed across B 0 , Bal and B 1 optimized pulses, due to a progressively lower value. (C) Repeatability measured as the coefficient of variability (wCV i ) for each vial. Averaging across all the vials, aSL preparations yielded significantly improved repeatability (wCV i = 0.29 ± 0.15 for B 0 -aSL, p < 0.01; wCV i = 0.23 ± 0.13 for Bal-aSL, p < 0.01; wCV i = 0.21 ± 0.11 for B 1 -aSL, p < 0.001 versus wCV i = 1.30 ± 1.34 for RefSL). Figure S3. T 1ρ,adiab maps obtained with B 0 -aSL, Bal-aSL and B 1 -aSL preparations. Image quality is compromised due to artifacts visible in the maps for B 0 -aSL in (A) and for Bal-aSL in (B). Furthermore Bal-aSL prepared baseline images were subject to substantial residual motion in both patients, lowering the image quality. Table S1. In vivo myocardial T 1ρ,adiab values (ms), averaged over all repetitions and segments for six healthy volunteers of cohort 1. Table S2. In vivo myocardial T 1ρ,adiab precision, reproducibility and inter-subject variability (ISV), averaged over segments and repetitions for 6 healthy volunteers of cohort 1.