Motion‐compensated fat‐water imaging for 3D cardiac MRI at ultra‐high fields

Respiratory motion‐compensated (MC) 3D cardiac fat‐water imaging at 7T.


| INTRODUCTION
Fat-water (FW) cardiac MR (CMR) imaging is an essential technique for assessing fatty infiltration of the myocardium. 1 Especially the differentiation of epicardial fat from intramyocardial fat remains difficult for the right ventricle with its thin structure, and overall small, affected areas in the myocardium compared to the limited resolution of current systems, and currently fail for the assessment of disease like arrhythmogenic right ventricular cardiomyopathy. 2,3 Therefore, a histologic evaluation of the biopsied right ventricular endomyocardial tissue is still part of the recommendation. 4 To determine the FW content in the human body, chemical shift encoding-based techniques have become an established clinical imaging method. [5][6][7] These techniques use different TEs to encode the FW information into the magnitude and phase of the MR signal. To resolve ambiguity problems, particularly in the presence of main magnetic field deviations (ΔB 0 ), triple gradient-recalled echo (GRE) methods 8,9 were proposed. The latter enables direct estimation of the ΔB 0 inhomogeneities, which can be used to resolve the ambiguity problem. [10][11][12] Thereby, the separation of the images into fat and water content can be performed sequentially 13,14 or in a joint reconstruction. 15,16 To date, 3D multi-echo GRE FW imaging techniques have been applied to the entire human body 10 as well as to various organs 8,17,18 at field strengths up to 3T.
However, by increasing the field strength, toward ultrahigh fields (UHF >= 7T), especially 3D FW imaging may benefit due to an increased SNR and spectral separation. Moreover, another benefit may be the reduction of the relative influence of j-coupling, which could improve the quantification of saturated and unsaturated fatty acids, as recently shown for lower fields. 19,20 In previous 7T body studies, the water or fat signal at UHF has been obtained by using 3D fat or water selective rectangular pulses for pelvic lymph node detection, 21 or by suppressing the fat signal with spectral pulses in 3D coronary angiography. 22 While, successful water and fat imaging has been demonstrated in breast imaging at 7T using a dual-echo GRE acquisition, 23 to the best of the author's knowledge, no investigations have been performed for UHF cardiac MRI using a multi-echo, cardiac binned (CB), and respiration resolved (RR) acquisition.
This gap may be linked to the fact that 3D imaging at UHF in the human body and, in particular, the heart is highly challenging due to multiple factors. First, the increased transmit field (B + 1 ) heterogeneity in the body at UHFs, leads to an inhomogeneous flip angle (FA) distribution, including regions of complete signal cancelation within the volume of interest. 24,25 This issue has been addressed by volunteer-tailored parallel transmission techniques (pTx), such as B + 1 phase shimming, 26-28 spokes, 29,30 kT-point 24 and, recently, using calibrationfree universal pulses. 31 Another technique that does not require volunteer-specific B + 1 information but provides homogeneous image without signal dropouts is the timeinterleaved acquisition of modes (TIAMO) technique. 21,32 Second, cardiovascular MRI (CMR), and particularly 3D CMR, is generally more challenging compared with brain MR applications because of the cardiac and respiratory motion that can decrease image quality. 33,34 Different strategies to handle cardiac motion were proposed 35,36 , however, most strategies rely on the acquisition of an electrocardiogram (ECG), and it has been shown that the ECG signal is strongly affected by the magnetohydrodynamic effect at 7T yielding incorrect ECG trigger points. 37 This has been especially a problem in long 3D hemodynamic scans performed under free-breathing at 7T that had to be aborted after some time and restarted with a different ECG detection algorithm. 25 The 3D FW imaging scans also require scan times of approximately 8-12 minutes 17,38 ; thus, similar issues could arise in FW scans performed at UHF. To mitigate respiratory-induced artifacts, breath-holds can be used in 2D acquisitions, but scan durations of 3D acquisitions covering the body are typically longer than a breath-hold. Thus, 3D acquisitions are usually obtained under free-breathing in combination with imagebased navigators, which track the motion of the liver. However, while this method is effective at lower field strengths, it may fail at UHF, since the navigator performance strongly depends on the B + 1 shim setting. 25 One efficient way to overcome this problem consists in using free-breathing acquisitions in combination with selfnavigation approaches 39,40 that bin or gate the acquired k-space data into different motion phases reflecting the respiratory position. 41 Additionally, self-navigation can Conclusion: This work demonstrates the feasibility of 3D fat-water imaging at UHF for the entire human heart despite spatial and temporal B + 1 and B 0 variations, as well as respiratory and cardiac motion.

K E Y W O R D S
7 Tesla, B 0 , body imaging, Dixon, fat-water imaging, parallel transmission, respiration be combined with respiratory motion correction [42][43][44][45] to increase the scan efficiency and use up to 100% of the acquired data.
Another challenge of FW imaging that is even observed at 1.5T or 3T is given by ΔB 0 variations across the target region. At UHFs, however, ΔB 0 , as well as its spatial and temporal variations are larger compared with lower field strengths. 46,47 In addition, potential fluctuations in ΔB 0 as a function of the respiratory phase have been demonstrated at UHFs for the human heart, 48 showing local ΔB 0 changes of up to 100 Hz between exhale and deep inhale. These respiration-induced ΔB 0 changes may have an impact on free-breathing 3D FW imaging at UHFs, especially when different respiratory phases are combined. This is because a respiratory-dependent B 0 would result in different ΔB 0 information from different respiratory phases 16 being combined in one image due to non-respiratory motion resolved reconstruction or motion compensation, which could lead to FW swaps in the resulting separated images.
In the present work, we investigate the feasibility of obtaining free-breathing 3D triple-echo GRE data for respiratory motion-compensated (MC) and CB FW imaging at 7T. After volunteer-tailored homogenous B + 1 shimming across the whole heart the triple-echo GRE data are binned into different respiratory phases, separate datasets for each respiratory phase are reconstructed and 3D respiratory motion vector fields (MF) are estimated using a non-rigid registration approach. 49 Since the MC step assumes static ΔB 0 for each moving voxel, respiratory phase-resolved ΔB 0 maps are calculated to investigate the resulting ΔB 0 changes throughout the respiratory cycle, and motion-corrected ΔB 0 fields are calculated. Finally, a joint reconstruction based on the triple-echo GRE dataset using the electrocardiogram (ECG) information, the calculated MF, a motioncorrected ΔB 0 estimate, and a six-peak signal model for fat is applied to reconstruct respiratory MC and CB 3D FW images of the heart at UHF.

| METHODS
The 3D triple-echo GRE data, based on a radial phaseencoding (RPE) trajectory 39 (cf. Figure 1A), were acquired (cf . Table 1), reconstructed, and separated into fat and water images. For reconstruction, an iterative, modelbased reconstruction framework 15 was used that utilizes parallel imaging, 50 compressed sensing, 51 motion information, 52 and a multi-peak spectral model for fat to account for chemical shift effects. The various steps required to obtain 3D FW separated images and RR ΔB 0 maps of the heart at UHF are presented below.

| Experimental setup
All MRI scans were performed on a 7T system (Magnetom 7T, Siemens Healthineers, Germany) equipped with an eight-channel pTx system. Imaging was performed with a commercial thorax coil array (MRI Tools, Berlin, Germany) consisting of 32 transceiver elements (eight dipoles and 24 loops) operated in an 8Tx\32Rx channel mode. The detailed setup of the coil, as well as the applied power limits, were identical to previous publications. 24,41 The study was performed after obtaining ethical approval by the local institutional review board and after F I G U R E 1 Overview of the trajectory and workflow used to obtain main magnetic field inhomogeneity (ΔB 0 ) maps and FW images. A, Schematic depiction of bipolar RPE trajectory with arrows indicating readout direction, with golden angle increment of 111.24° between successive spokes. B, Block diagram of the reconstruction workflow to obtain ΔB 0 maps of different motion phases, that is, NRR, RR, and respiratory MC. C, Block diagram of reconstruction workflow to obtain FW images in different RR, NCB, and CB motion phases, detailed in the motion handling subsection providing written informed consent from 11 healthy volunteers (age: 30 ± 5 years; max\min: 35\21 years; body mass index: 24 ± 5 kg/m 2 ; max\min: 34\19 kg/m 2 ; male\ female: 7\4). A three-lead wireless ECG was used for retrospective cardiac phase binning, following Ref. [37]. Leads were placed by the operators with volunteers lying in a supine, head-first position on the table prior to coil placement. After placing the coil centrally over the heart and positioning the table into the isocenter, ECG triggering performance was evaluated at the host console. It was visually checked whether (i) other peaks were falsely detected in addition to the R-peak, (ii) the detection of the R-wave correlated with the signal from the infrared plethysmograph attached to the left index finger, and (iii) there was any potential jitter in the R-peak detected by the scanner. If the trigger performance was insufficient, the trigger mode was first changed, and if this was not sufficient, the electrodes were replaced and the procedure was repeated. The default (tune-up) B 0 shim was applied and the center frequency was set to the water peak using the semiautomatic adjustment routine of the system.

calibration
All volunteers underwent the same protocol. In addition, a high-resolution 3D triple-echo GRE-RPE scan was performed for the last volunteer, further described below. Each protocol started with the default B + 1 shim setting 0 of the coil, set by the manufacturer and optimized based on electromagnetic simulations for the ascending and descending aorta. This default phase setting, however, may result in signal cancellations in some volunteers. Therefore, relative 3D B + 1 mapping was performed using a 3.40 minutes long free-breathing 3D GRE-RPE acquisition as outlined in Ref. [41]. Scan parameters are listed in Table 1. Since only a shallow-breathing pattern was requested, strong respiration-induced variations of B + 1 were not expected; therefore, motion-averaged relative 3D B + 1 datasets were extracted. 41 The raw data were transferred to a separate workstation (RAM: 128 GB; CPUs: 12 with 2.1 GHz) and B + 1 maps were reconstructed using MATLAB R2019b (MathWorks, Natick, MA, USA) with 4 mm isotropic resolution within approximately 1 minute. Either B + 1 phase or B + 1 magnitude and phase shimming was subsequently performed on three transversal slices across the heart, as indicated in Figure 2, to obtain a homogenous B + 1 distribution within the heart, similar to Ref. [41]. Actual FA imaging (AFI) was performed 53 during RF shim calculation to quantify the average FA within the target region. This analysis, however, was only performed retrospectively due to the tight timing of the protocol.

| 3D triple-echo GRE-RPE imaging protocol
After applying the calculated B + 1 shim, 3D triple-echo GRE-RPE data with 1.  Table 1. For the last volunteer (V11) the same protocol was acquired, except that one additional 3D

| Bipolar gradient correction
Prior to the reconstruction of the FW images and the calculation of the estimated ΔB 0 maps, phase errors resulting from the bipolar readout gradient polarity were corrected. 54 Therefore, the bipolar-uncorrected 3D tripleecho GRE-RPE k-space data (K u ) were reconstructed into non-RR (NRR) and non-CB (NCB) images using an iterative SENSE algorithm. 50 Constant and linear phase terms of the second echo with respect to the first and third echo were estimated in the image domain along the readout direction for each volunteer. Then the second echo K u (TE 2 ) is corrected for the constant and linear phase terms to obtain bipolar corrected triple-echo GRE-RPE data (K).

| Motion handling
Respiratory MC was applied to transform data from different respiratory phases into a single respiratory reference phase, which was set as the end-exhale phase. MC increases the sampling density in the resulting k-space dataset compared with the respiratory binned k-space datasets and, thus, reduces undersampling as well as motion artifacts in the resulting reconstructed MC images I MC . A particular aspect to consider in MC is the potential fluctuation of ΔB 0 during respiration, which could lead to erroneous MC, NCB ΔB 0 maps (ΔB 0,MC ) derived from I MC and, thus, may result in erroneous FW images. To investigate this aspect, we first separated the data into different respiratory phases to analyze the impact of respiration as well as MC on the resulting ΔB 0 maps (cf. Figure 1B).
To examine the impact of respiratory motion and MC, the triple-echo GRE-RPE dataset K was binned into different respiratory phases using self-navigation using a sliding window with 30% overlap to the next respiratory bin. Then, the datasets were reconstructed into four RR and NCB triple-echo images I NCB RR = {I exhale ,..., I inhale } using iterative SENSE. For self-navigation, a motion surrogate was estimated from the central k-space line of each RPE spoke from the first echo K(TE1). 40 Subsequently, a non-rigid MF was determined for each respiratory phase from the RR images of the first echo I NCB RR (TE 1 ) using NIFTYREG. 49 Therein, the MFs describe the pixel's displacements from a respiratory phase to the reference phase. These MFs were used for MC image reconstruction 42 of K to obtain MC and NCB images I NCB MC that were used for the calculation of ΔB 0,MC .
In addition, NRR and NCB, NRR and CB, and MC and CB FW images FW NCB NRR ,FW CB NRR , and FW CB MC were derived from the same datasets K (cf., Figure 1C), with FW CB MC being reconstructed with ΔB 0,MC , and NRR, NCB ΔB 0 maps (ΔB 0,NRR ). Therefore, the bipolar-corrected dataset K was retrospectively split into five different cardiac phases using the recorded ECG signal with a sliding window and a 30% overlap to the next cardiac bin to obtain FW CB NRR images with compressed sensing 15 reconstruction techniques. The acquisition window of each phase depends on the heart rate and corresponds to ~200 ms per cardiac phase. For the reconstruction of MC and CB images, the bipolar-corrected dataset K was first divided into the five cardiac phases and each cardiac phase dataset was divided into the different respiratory motion phases to apply the MF during compressed sensing reconstruction of FW CB MC .

| Estimation of ΔB 0
To investigate changes of ΔB 0 as a function of the respiratory cycle images I NCB NRR , I NCB RR and I NCB MC (cf. Figure 1B) were used to calculate the corresponding ΔB 0 maps with a multi-seeded safest-first region-growing algorithm outlined in Ref. [10]. It uses an analytic function to find ΔB 0 values for each voxel, identifies true seed voxels over the entire volume, and grows a ΔB 0 map.

| Reconstruction of fat and water images
Three different types of FW images were reconstructed: . ΔB 0,NRR were used to reconstruct (1), (2), and (3), whereas ΔB 0,MC were also used to reconstruct (3), as shown in Figure 1C. To obtain the fat and water images, a signal-model-based compressed sensing reconstruction was performed afterward. 15,38 The acquired multi-echo data are assumed to be composed of the water signal W, precessing at a single frequency, and the fat signal F, which is modeled by a six-peak spectrum as used by the "FattyRiot" toolbox, 55  A compressed sensing reconstruction was performed on bipolar-corrected raw data K by solving the following minimization problem: with spatial total variation term TV for each motion phase and the total variation term TVT along the motion phases. λ FW and λ t are the spatial and temporal regularization parameters used to set the tradeoff between data consistency ‖E(W, F) − K‖ 2 2 and minimization of undersampling artifacts. E is the forward operator, transforming the water images W and fat images F to k-space, and it contains the gridding operator, the FFT, the six-peak fat model, coil sensitivity maps, 56 MFs, and the ΔB 0 maps. The regularization parameters were optimized on the first volunteer and then retained for all volunteers. λ t is set to zero for non-resolved motion phase reconstruction, and the MFs are only used for MC reconstruction. Reconstruction of FW images with 1.4 mm 3 isotropic resolution took 4 h for FW NCB NRR and up to 10 h for FW CB MC images.

| Data analysis
For the retrospective assessment of the actual FA, channelwise relative B + 1 maps were calibrated with an absolute AFI map as outlined in Ref. [41] to obtain estimated, absolute B + 1 maps. This retrospective evaluation was performed after the measurement due to tight scan time restrictions. Using the absolute B + 1 maps, the mean and the SD of the actual FA for the entire heart (ROI B1 ) were determined. The determined ROI B1 was manually defined for each transversal slice along the heart-lung boundary, with the coefficient of variation (CV) defined as in Ref. [41].
ΔB 0 maps were then analyzed within a 3D ROI (ROI B0 ) covering the entire heart for each volunteer and each respiratory phase. Therefore, manually defined 2D ROIs covering the heart in six transversal slices for every respiratory phase were created with a 10 -18 mm spacing between slices in the head-foot (HF) direction. ROI B0 with full heart coverages was then generated by interpolating the 2D ROIs in HF direction between the slices in MATLAB R2019b (function: INTERPMASK, MathWorks, Natick, MA, USA). ΔB 0 maps are given in Hz and subsequently, the SD of each ΔB 0 distribution within ROI B0 was determined. To compare two masked ΔB 0 maps in different respiratory phases, the center frequency difference Δf 0 and the SD of the voxel-wise difference between the two ΔB 0 maps was determined. The ECG trigger efficiency was retrospectively assessed for all subjects and the entire triple-echo GRE-RPE data acquisition. This was performed by visual comparison of the locations of the detected trigger points by the scanner's trigger routine and the locations of the R-peaks in the ECG curve that are identified by the operator. The trigger efficiency was then defined as the ratio of R-peaks aligned with the detected trigger points to the total number of identified R-peaks in the ECG signal.
Fat-fraction (FF) images were calculated from W and F magnitude images as follows 57 : Figure 2 shows the estimated absolute B + 1 map and corresponding GRE-RPE image for volunteer V9 (BMI: 22.3 kg/m 2 ) with default and homogenous B + 1 shim after optimization. Qualitatively, the estimated B + 1 maps match the magnitude of the localizer images and predict the resulting signal intensities. Note, that receive profiles were not corrected. Quantitatively, the CV (mean ± SD) across the entire heart (ROI heart ) and all volunteers was reduced from 42 ± 4% for 0 , to 24 ± 4% for homogenous B + 1 shim while the average FA for all volunteers remained unchanged (5 ± 2° for 0 , 5 ± 1° for homogenous B + 1 shim). The impact of respiration on the resulting ΔB 0 maps in one representative volunteer V2 (BMI: 21.0 kg/m 2 ) is shown in Figure 3. Note, that the tune-up B 0 shim setting FF = |F| ∕ (|W| + |F|) . F I G U R E 3 A, Coronal and transversal view of ΔB 0 distributions for NRR and RR ΔB 0,RR maps from exhale to inhale motion state. Note ΔB 0,RR are filtered. The dotted white line serves as a guide for the eye to illustrate the movement of the liver through the respiratory phases. Mean (±SD) displacement of 9.8 ± 1.4 mm of right hemidiaphragm position between exhale and inhale was determined. NRR ΔB 0 range across the heart amounts to 450 Hz for the coronal slice and 400 Hz for the transversal slice. B,C, Line plots for foot-to-head (I) and posterior-to-anterior (II) direction are indicated in (A) as black dotted lines. In the foot-to-head direction, a maximum difference between inhale and exhale motion state of −90 Hz, for NRR and exhale of −70 Hz, a mean Δf 0 (±SD) between exhale and inhale of 29 ± 25 Hz, and between NRR and exhale\inhale of −13 ± 12\19 ± 19 Hz in is calculated. In posterior-to-anterior direction (II) a maximum difference between inhale and exhale motion state of −91 Hz, for NRR and inhale of 50 Hz, a mean Δf 0 (±SD) between exhale and inhale of 26 ± 24 Hz, and between NRR and exhale\inhale of −16 ± 18\9 ± 8 Hz is calculated. Images were taken from volunteer V2 was used throughout this work. Filtered RR ΔB 0 maps (ΔB 0,RR ) maps for four different respiratory phases are shown in Figure 3A in coronal and transversal orientations. Qualitatively, no undersampling or motion artifacts were observed in the different respiratory phases, yielding high-quality ΔB 0 maps that reflect only subtle changes of ΔB 0,RR over the different respiratory phases. Furthermore, similar maps were obtained for the NRR ΔB 0 maps (ΔB 0,NRR ) compared with ΔB 0,RR with the largest differences of 70 Hz between the inhale and the NRR phase, which is particularly reflected in the line plots ( Figure 3B) along the dotted lines indicated in Figure 3A. For ΔB 0,NRR , the ΔB 0 range across the heart was 450 Hz for the coronal slice and 400 Hz for the transversal slice (cf., Supporting Information Figure S1, which is available online, for nonfiltered ΔB 0,RR maps).

| RESULTS
Similar results were found across all volunteers, as shown in Figure 4, where ΔB 0 histograms for the entire 3D ROI B0 are illustrated for each volunteer with Δf 0 and SD values for exhale and inhale (cf., Supporting Information Figure S2 showing ROI heart in exhale motion phase). While the shape of the distributions barely changes between exhale and inhale, each distribution appears to be shifted between inhale and exhale. Quantitatively, most of the ΔB 0 values, that is, 95%, are within a range of 460 ± 50 Hz for the whole heart. The SD averaged across all volunteers is 117 Hz\122 Hz for ex-\inhale phase with a mean center frequency shift Δf 0 between inhale and exhale of −15 Hz. Figure 5 illustrates voxel-wise differences of ΔB 0,RR after applying the MFs, transforming each respiratory phase to the exhale phase, and subtracting the exhale ΔB 0,RR map. Maximum differences are seen for the motion-corrected F I G U R E 4 Pixelwise main magnetic field deviation (ΔB 0 ) values are displayed in histograms, whereby the data are taken from different ROI heart for every respiratory phase and all 11 volunteers (cf. Supporting Information Figure S2 for ROI heart in exhale motion phase for each volunteer). With center frequency shift Δf 0 between exhale and inhale, and SD of exhale and inhale motion phase given in right corner of each histogram. Mean Δf 0 of −15 Hz between inhale and exhale, and mean SD of 120 Hz (min\max: 97\150 Hz) for inhale and exhale motion state across all volunteers are calculated F I G U R E 5 ΔB 0 differences maps of different RR and MC phases are shown for coronal slices of volunteer V7. The first four images show the difference between the ΔB 0 map in the exhale phase (ΔB 0,RR(exhale) ) and the four RR and subsequently MC ΔB 0 maps (MC ΔB 0,RR(i) ) from exhale to inhale phase. The mean center frequency shift (Δf 0 ) between ΔB 0,RR(exhale) and MC inhale ΔB 0,RR map with −14 Hz and a mean SD of 35 Hz is calculated. The last image shows the difference between ΔB 0,RR(exhale) , and ΔB 0,MC with a mean Δf 0 of −4 Hz, and a mean SD of 16 Hz inhale maps (fourth column) with maximum and mean Δf 0 (±SD) of −90 and −14 ± 35 Hz across the whole heart for volunteer V7 (BMI: 20.0 kg/m 2 ). The last column shows differences in ΔB 0 between the ΔB 0,MC map taking all respiratory phases into account and the exhale ΔB 0,RR map. Here, only marginal differences are observed with a maximum and mean Δf 0 (±SD) of −40 and −4 ± 16 Hz. Similar values were found for the other volunteers. Overall, these findings demonstrate that the local respiration-induced frequency shift in the order of 10 Hz on average is small compared with the chemical shifts used in the multi-peak model in which more than 90% of the signal has a >700 Hz frequency shift with respect to the water peak. The resulting frequency shift in the ΔB 0,MC map compared to the exhale ΔB 0,RR map amounts to less than 2% of this value, as well as roughly 10% to the fat peak at −115 Hz, which accounts for 4% of the fat signal. Figure 6 shows the resulting FW images of V2 in a coronal slice for FW NCB NRR images using 100% of the acquired data, for FW CB NRR images in the diastolic phase using only 30% of the data throughout the cardiac cycle, and for FW CB MC images in the diastolic phase (also 30% of the data) with ΔB 0, NRR, and ΔB 0,MC maps. Qualitatively, cardiac binning decreases the fat signal in F CB NRR and F CB MC compared with F NCB NRR in the blood pool ( Figure 6, blue arrows), due to the reduced impact of blood flow that alters the signal phase especially during systole, likely due to a first gradient moment induced phase accumulation of moving spins in the systolic phase (cf., Supporting Information Video S1). This flow-induced effect is as well visible in the line plot shown in Figure 6B denoted by the blue arrow. Due to the reduced flow in the diastolic phase of FW CB MC and FW CB NRR , this artifact is reduced by approximately 50% compared with FW NCB NRR . Motion blurring is reduced due to the applied MF in FW CB MC compared with the NRR images as denoted by the yellow arrows in Figure 6. This is also reflected in the increased fat signal in FW CB MC (cf., Figure 6 B, red arrow) for both ΔB 0,NRR, and ΔB 0,MC reconstruction. Here, the compensation of the underlying respiratory motion leads to a 60% and 90% higher signal and a reduction of the FWHM of 11% and 27% compared with F CB NRR . The resulting FW images of the remaining volunteers are shown in Supporting Information Figure S3.
The trigger efficiency was retrospectively investigated and showed a mean value of 99.7% across 10 out of 11 volunteers. Figure 7A shows an exemplary ECG signal curve of channel I obtained from volunteer V7 for 90 s and 25 s. As can be seen, the R-peak is detected correctly by the scanner in all but one case; for the entire acquisition, 4 of 438 were not correctly detected, resulting in a trigger efficiency of 99.1%. In one volunteer (V6), the efficiency with ΔB 0,NRR (purple) and ΔB 0,MC maps (orange), the position is indicated by the dotted line in FW NCB NRR . MC led to a reduction of FWHM of 11% (purple) and 27% (orange) compared with NRR and CB fat image, indicated by the black arrows was only 80%, since every third to sixth R-peak in the ECG curve appeared to be missing (Supporting Information Figure S4). To investigate the impact on the images, subfigure (B) shows resulting images after retrospectively removing every 5th and 10th trigger point, resulting in a trigger efficiency of 80%, 90%, and 99.1% for V7. The resulting images in Figure 7B, are qualitatively similar; however, the edges and small details become blurrier with decreasing number of correct trigger points. Figure 8 shows reformatted images for V2 and V5 (BMI: 34.0 kg/m 2 ) shown in Supporting Information Figure S3, in short axis (SAX), long axis (LAX), and four-chamber view (4CV). A coronary artery can be seen in the water image of FW CB MC and in the corresponding FF images, denoted by the red arrow. FF for subcutaneous fat is calculated to 93 ± 3%. Figure 9 shows water images FW CB MC in diastole of V11 (BMI: 20.6 kg/m 2 ) with 1.4 mm 3 and 1.1 mm 3 isotropic resolution in two oblique views with depicted coronary arteries (blue and yellow box), which are surrounded by fat, showing no residual problems related to FA heterogeneity or residual motion. Note that, for the 1.1 mm 3 scan, not only has the resolution been increased by 27%, but also the reduction factor R has been increased by 26% compared to the 1.4 mm 3 scan. F I G U R E 7 A, Illustration of the 90 s and 25 s of the entire ECG signal curve. As can be seen, the R-peak is detected by the scanner in all but one RR-interval, where the ECG signal appears to be corrupted for a short period of approximately half an RR-interval. For the entire acquisition, 4 of 438 were not correctly depicted in volunteer V9, resulting in 99.1% trigger efficiency. B, Transversal images showing 80%, 90%, and 99.1% trigger efficiency and the impact of the resulting fat and water images in diastole. Qualitatively, the edges and detailed structures such as the coronary artery become increasingly blurry with decreasing number of correct trigger points

| DISCUSSION
This work demonstrates for the first time successful 3D cardiac FW imaging at UHF for a wide range of BMI and body shapes allowing cardiac phase-resolved fat and water images in arbitrary orientations. In addition to FW imaging, this work investigated the impact of shallow respiration and MC on the ΔB 0 distribution in the heart. Related to the FW frequency shifts only small deviations are caused by MC in the resulting ΔB 0,MC maps compared with the reference end-exhale ΔB 0,RR map. CB MC FW reconstruction reduced respiratory motion-induced blurring in the FW images and avoided flow artifacts in the end-diastolic FW images.
The 3D cardiac FW imaging at 7T requires proper B + 1 management to avoid signal cancellations within the relatively large target region. Applications targeting the human body at 7T using either fixed volunteer-unspecific B + 1 shimming, 21,23 tailored B + 1 shimming, 58 or dynamic pTx techniques. 24,30 Fixed volunteer-unspecific approaches do not require additional RF management, but may lead to large B + 1 variations within the target region 23 or can lead to an extended acquisition time if the TIAMO method is used. 32 In this work, tailored B + 1 shimming was performed for three transversal slices covering the heart using channel-wise relative 3D B + 1 maps obtained within the entire thorax. This approach was necessary because the default B + 1 shim resulted in signal voids in the heart in nine volunteers. Although to our experience the size of the heart is at the limit of what can be homogenized with B + 1 shimming on an eight-Tx channel system, the phaseonly shimming approach yielded B + 1 distributions without signal voids in all but one volunteer. For the one volunteer, magnitude and phase shimming was used to obtain a B + 1 distribution without signal voids. This is reflected by the mean CV of the B + 1 maps across all volunteers in the human heart with an unacceptable CV of 42% for the default shim, which was reduced to a CV of 24% with B + 1 shimming. More homogeneous FA patterns may be obtained using dynamic pTx, such as kT point pulses as shown recently 24 ; however, here the different frequencies need to be considered during RF pulse design. 59 Nevertheless, the residual FA variations seem not to affect the proposed FW approach that is relying on magnitude and phase information. Within the 3D volume, sufficient signal was available in all volunteers to successfully demonstrate the feasibility of cardiac FW imaging at 7T.
The 3D triple-echo GRE-RPE sequence allows the acquisition of high-resolution images under free-breathing and, therefore, is highly beneficial for UHFs. This RPE trajectory in combination with MC has been previously applied in a patient study to improve the image quality of cardiac PET\MR images for localization of coronary plaques 38 and in healthy volunteers for more accurate peak flow determination 42 at 3T. In contrast to the applied self-navigation approach, which has proven valuable for estimating respiratory motion not only at lower fields, there are several approaches to track respiratory motion, including image-based navigators. 60 The latter is more difficult to acquire at UHF because the used B + 1 shim setting has a strong influence on navigator performance, which can lead to unpredictable scan times and scan aborts. The estimated motion information from self-navigation did not suffer from local B + 1 changes, and accurate motion minimization was ensured.
In order to keep the scan time short, MC is applied to use 100% of the acquired GRE-RPE data and to reduce the blurring caused by respiration and, thus, increase image resolution. In contrast to the retrospective respiratorygated acquisition, the entire data set is available for cardiac binning, which reduces undersampling in the resulting cardiac bins. MC remaps each tissue voxel obtained at different respiratory motion phases into a single reference phase using motion vector fields. Thus, if ΔB 0 were to change over the respiratory cycle, a respirationdependent signal phase term would be included in the resulting images that could affect the applied FW image reconstruction. Specifically, this may result in crosstalk between water and fat peaks, leading to swaps or incorrect FW values, as has been shown previously. 16 To investigate the extend of ΔB 0 across the respiratory cycle, 3D ΔB 0 maps have been acquired as a function of the respiratory phase, and changes in ΔB 0 have been investigated. While fairly strong spatial peak-to-peak ΔB 0 variations of 460 ± 50 Hz have been found across the heart, which match with values reported by other groups for a global shim, 61 the results demonstrate only limited ΔB 0 variations of −90 Hz at maximum and 35 Hz on average between expiration and inspiration for shallow-breathing, which confirm previous observations made for 2D transversal cardiac slices obtained at different respiratory phases. 48 Resulting MC ΔB 0 maps that incorporate signals from all respiratory phases and map them to the exhale phase show even fewer differences of 10 Hz on average to the exhale ΔB 0 map and such values can be regarded as small compared with the chemical shift between water and fat. It was found that the shape of the ΔB 0 distribution barely changes while its center frequency shifts as a function of respiration. This may be linked to the predominately translational motion of the heart in the HF direction, but further investigations are needed to fully characterize ΔB 0 changes across the respiratory cycle. Thus, for the present case of shallowbreathing, we show that it is sufficient to use NRR maps instead of the RR and MC ΔB 0 maps for further reconstruction. In addition, it is also possible to use ΔB 0 maps in the reference motion phase for MC, CB, FW separation, but volunteer-specific parameter adjustment is required to estimate artifact-free ΔB 0 maps in the entire FOV, which was not necessary for MC or NRR ΔB 0 map estimation.
This study also has limitations. First, the tissue's relaxation times T 1 and T * 2 impact the FW quantification. In this work, a T 1 bias, that is, overestimation of FF, 6 is expected to remain by applying a TR of 6.1 ms and an actual FA of approximately 5° that is close to the expected Ernst angle. Thus, further corrections may be necessary. 62 Additionally, a T * 2 bias, that is, underestimation of the fat content, 6 was observed in the phantom evaluation of test tubes with different FW ratios (100%, 53%, and 26%, cf., Supporting Information Figure S5). A total of 8-9% lower fat content was measured for 100% and 26% FW ratio, and 21% lower fat content for 53% FW ratio. This bias could be reduced by T * 2 estimation, which has been realized in studies of fat quantification in the liver or other body parts with an increased number of echoes at lower field strength. 20,63 However, such deviations found in the 7T phantom data may not reflect the in vivo behavior because the relaxation times (fat: T 1 = 225 ms; T * 2 = 90 ms and water/agarose T 1 = 150 ms; T * 2 = 30 ms) differ from in vivo values. To quantify the FF and estimate its accuracy in vivo, further studies are needed to (1) compensate for the confounding factors and (2) validate the measurements with spectroscopy data, as has been done at lower fields. 64,65 Second, limitations arise with the 7T system in terms of the increased complexity of the overall measurement process. A contributing factor is the tailored calibration of the B + 1 field, which is necessary to avoid complete signal cancellation. Despite a tight reconstruction and shimming procedure, an additional calibration time of about 10 minutes is required. Third, despite careful placement, 37 acquisition of an acceptable ECG signal is error-prone and remains difficult compared with lower field strength. Despite other findings, 66 a reliable ECG signal was obtained in 10 of 11 volunteers in this study as in Ref. [67]. We speculate that the poor efficiency found in one volunteer may be related to the respiratory motion, but a separate analysis is necessary. Fourth, the contrast between blood and myocardium is not very pronounced in T 1 -weighted GRE sequences at UHF because of the similar T 1 values of blood and myocardium. At lower field strengths, T 2 preparation pulses are used to increase contrast. At UHF, the use of T 2 preparation pulses is challenging, and tailored approaches like VERSE pulses need to be used. 68 Fifth, the present work does not yet apply a resolution higher than currently used at clinical field strength, which is about (1.05-1.5 mm) 3 17,45,69 isotropic. Increasing the resolution would require a detailed study of the echo spacing, which is currently 1.28 ms, thus pushing the current gradient system to its limits and requiring an increase in the echo spacing to non-noise-optimal spacings. 70 The aim of this work was rather to investigate the feasibility of the technique and to identify potential challenges that are associated with the high field in combination with the present techniques.

| CONCLUSIONS
Despite the known challenges in UHF body MRI posed by B + 1 heterogeneity, spatial and temporal B 0 variations, and respiratory and cardiac motion, this work demonstrates the feasibility of 3D FW imaging at UHF for the whole heart. Based on this work, we intend to take advantage of UHF, such as higher SNR, by further developing the method to achieve improved spatial resolution with the aim to separate epicardial fat from small fat infiltrations in the right ventricle in the future. In addition, the workflow applied in this work can be used in other applications and push the field of UHF body MRI forward.

SUPPORTING INFORMATION
Additional supporting information may be found in the online version of the article at the publisher's website. FIGURE S1 Shows Figure 3 with unfiltered RR ∆B 0 maps FIGURE S2 ROI heart for exhale motion phase as an overlay on GRE image for three orthogonal views of all eleven volunteers FIGURE S3B + 1 shimmed, motion-compensated, and cardiac binned fat-water images in the end-diastolic phase are shown for all eleven volunteers in 3 orthogonal views with BMI ranging from 19 to 34 kg/m 2 . As can be clearly seen, the fat and water content could be robustly separated across different body geometries around the heart. Furthermore, signal voids caused by destructive B + 1 interference in the heart could be avoided in all volunteers. Note that images were not corrected by receive profiles (B − 1 ) FIGURE S4 ECG curve of volunteer V6 with 80% trigger efficiency. (A) First 90 and 20 seconds and (B) interval from 330 to 420 seconds with zoom between 338 and 355 seconds with corresponding trigger points (red asterisk)