Reduction of breathing artifacts in multifrequency magnetic resonance elastography of the abdomen

With abdominal magnetic resonance elastography (MRE) often suffering from breathing artifacts, it is recommended to perform MRE during breath‐hold. However, breath‐hold acquisition prohibits extended multifrequency MRE examinations and yields inconsistent results when patients cannot hold their breath. The purpose of this work was to analyze free‐breathing strategies in multifrequency MRE of abdominal organs.


| INTRODUCTION
Abdominal magnetic resonance elastography (MRE) is gaining in importance for the clinical diagnosis of a large variety of diseases that alter the mechanical properties of tissues such as liver fibrosis, 1-5 portal hypertension, 6,7 renal dysfunction, [8][9][10] and tumors. [11][12][13][14] Since the introduction of MRE in 1995 by Muthupillai et al, 15 various MRE methods tailored for clinical examinations of the abdomen have been proposed. 16 A general challenge for abdominal MRE is to cope with respiratory motion during data acquisition, which potentially degrades the consistency of MRE-encoded vibration data. 17 Therefore, the Quantitative Imaging Biomarkers Alliance of RSNA (Radiological Society of North America) recommends MRE of the liver to be performed during breath-holds after expiration. 18 Standard MRE of the liver consists of acquisitions of four slices squeezed into multiple breath-holds of 12 to 20 seconds 2 or into a single breath-hold using rapid single-shot sequences or simultaneous multislice MRE. 19,20 However, there are several reasons to avoid breath-holds in abdominal MRE. First, patients with chronic liver diseases often suffer from comorbidities that limit their ability to hold their breath after expiration during the acquisition of MRE data. 17 Second, patients cannot always follow breathing commands, resulting in motion during scans or inconsistent organ positions. 21 Furthermore, breath-holds might alter abdominal pressure and hepatic perfusion, which in turn affect liver stiffness. 22,23 Finally, paused data acquisition to permit intermittent respiration prolongs total acquisition time beyond the clinically acceptable limit if rich data are acquired as in multifrequency, 3D MRE.
Published studies report multifrequency 3D MRE of abdominal organs with up to eight wave dynamics consecutively acquired at up to seven frequencies, resulting in total acquisition times exceeding 5 minutes, during which patients are allowed to breathe freely. [8][9][10]24,25 The same protocol with breath-hold acquisition would exceed 20 minutes. 26 Nevertheless, there are good reasons for multifrequency MRE of abdominal organs, including measurement of viscoelastic dispersion or improved detail resolution. 25,27,28 For example, tomoelastography, a multifrequency MRE technique for abdominal organs, was applied during free breathing for detecting small hepatic lesions 14 or fine renal structures, 25 despite the risk of breathing artifacts. Although individual images from rapid single-shot MRE scans may have minimal motion artifacts in the organs of interest, intrascan misregistration may degrade the resulting elastograms. This study addresses strategies to avoid and correct intrascan misregistration in single-shot MRE of abdominal organs. Our hypothesis is that detail resolution in multifrequency MRE can be further improved by suppression of motion artifacts either during image acquisition using respiratory navigators or with postprocessing using image registration before MRE wave inversion.
Experimental strategies include one-dimensional breathing navigators for either defining acquisition windows or for adapting slice positions following inspiration and expiration. Additionally, motion is corrected in raw MRI signals during postprocessing. These strategies will be compared with the current gold standard, which is breath-hold MRE. The overall aim of our study is to recommend strategies for efficiently suppressing breathing artifacts in abdominal multifrequency MRE toward time-efficient, highly resolved stiffness mapping of the liver, spleen, pancreas, and kidneys without the need for restricting breathing during the examination.  -array surface coil and  the spine-array coils integrated into the examination table. For anatomical orientation, localizer, T 1 -weighted and T 2 -weighted axial, as well as T 2 -weighted coronal MRI sequences were acquired before MRE. The MRE sequences and driver setups were similar to Shahryari et al. 14 In brief, the driver setups consisted of four pressurized air drivers that were attached to the body surface with a Velcro belt at the position of the transpyloric plane. Two of the drivers were attached ventrally on the midclavicular line, while the remaining two were placed dorsally on the scapular line. Mechanical vibrations of 30, 40, 50, and 60 Hz frequency were induced in the abdomen with 0.4 bar air pressure for the ventral drivers and 0.6 bar air pressure for the dorsal drivers. Neither position nor air pressure of the actuators was individually adapted to the body shape of the volunteer. A single-shot spin-echo EPI sequence with flow-compensated motion-encoding gradients was used to acquire the 3D wave field. 29 Eight phase offsets over a full vibration period were recorded for all three Cartesian motion field directions. Nine contiguous slices of 5 mm thickness and placed in a coronal view through the liver, spleen, pancreas, and both kidneys were acquired, covering a FOV of 284 × 350 mm 2 . Further imaging parameters were as follows: matrix size = 104 × 128, pixel resolution = 2.7 × 2.7 × 5 mm 3 , TR = 1200 ms, TE = 55 ms, parallel imaging with GRAPPA factor 2, motion-encoding gradient frequency = 43.48 Hz for 30, 40, and 50 Hz vibration frequencies and 44.88 Hz for 60 Hz vibration frequency, and motion-encoding gradient amplitude = 30 mT/m. For BH-MRE (i), single-component, single-frequency MRE data of each of the three Cartesian wave field components and four drive frequencies were acquired block-wise. Therefore, separate scans were performed during 12 single breath-holds in expiration of approximately 10 seconds duration. Data were merged afterward to full six-dimensional data arrays of sizes 104 × 128 × 9 × 8 × 3 × 4, depicting 3D spatial-resolved displacement data (104 × 128 × 9) at multiple time steps (8), components (3), and frequencies (4) resulting in N = 96 consecutively acquired slice blocks. The magnitude images of the 96 acquired slice blocks were averaged to a mean magnitude image (M). For FB-MRE (ii), all data were continuously acquired without account of respiratory motion.

| Subjects and paradigms
For G-MRE (iii) and GF-MRE (iv), we integrated a navigator echo into the EPI-MRE sequence before slice block acquisition, to monitor the respiratory position based on the method described in Henningsson and Botnar. 30 In brief, a pencil beam RF pulse with 30° flip angle and 20 mm diameter was positioned craniocaudally through the right diaphragm with the readout axis parallel to the long axis of the beam. The acquired navigator data were then Fouriertransformed in real time, generating a one-dimensional line image illustrating respiratory motion. An acceptance window of 5 mm has been defined for G-MRE to interrupt the acquisition of data outside this window. In GF-MRE, the acceptance window was 20 mm, within which the position of the full image slice block was continuously moved along the head-to-foot axis, to follow the position of the right diaphragm depicted by the navigator signal. As in G-MRE, data acquisition was interrupted if a very deep breathing motion exceeded the 20 mm window detected by the navigator. Of note, image slices were continuously excited in G-MRE and GF-MRE, similar to FB-MRE but with paused data acquisition whenever the navigator signal was outside the acceptance window. Because TR in all protocols was identical (1.2 seconds), a similar steady state of magnetization was established during FB-MRE, G-MRE, and GF-MRE. Only in BH-MRE did the steady-state magnetization have to be re-established for each single breath-hold, potentially inducing MRI intensity fluctuations. However, as analyzed later, these fluctuations did not affect the sharpness of averaged magnitude images M. All protocols are illustrated in Figure 1.

| Magnetic resonance elastography data processing
Organ displacement and motion artifacts were quantified by 2D rigid-body image registration using the open-source Elastix toolbox. 31 Therefore, a presegmentation region of interest (ROI) was manually delineated for each organ, covering the full organ with a tolerance margin of approximately 5 pixels that covers organ displacements caused by motion. Presegmentation ROIs for image registration are demarcated in Supporting Information Figure S1. The mean MRE magnitude images (M) displaying a mixed T 2 and T * 2 contrast were registered to the first of 8 × 3 × 4 = 96 slice blocks acquired during a multifrequency MRE scan using (1) advanced Mattes mutual information with 32 histogram bins as a similarity metric 32 and (2) stochastic gradient descent as an optimizer implemented in Elastix. The registration was performed with a pyramid resampling scheme at three different resolutions. The optimizer was set to a maximum of 100 iterations, and at each iteration, 2000 random coordinates were sampled for the computation of the similarity metric. As a result, N = 95 in-plane relative displacements x (horizontally) and y (vertically) were generated for each organ. used to apply the transformations to the real and imaginary parts of the complex MRI data, the phases of which were further processed for MRE parameter recovery. Although no explicit smoothing step was implemented in the registration process, a slight smoothing may have occurred due to the B-spline interpolation. Figure 2 shows representative 40 Hz waves (left-toright component) in a slice covering the liver, kidneys, spleen, and pancreas acquired with all MRE protocols, uncorrected and corrected. More material on animated waves is presented in Supporting Information Video S1. All MRE wave data were processed using wave number-based multifrequency dual elasto-visco inversion (k-MDEV), which provides maps of the shear wave speed (SWS) in meters per second). 33 Because SWS is derived from the real part of wave numbers, it is virtually not influenced by viscosity and can be considered a surrogate of stiffness. The inversion pipeline is publicly available at https://bioqi c-apps.chari te.de. 34 For parameter quantification, organ-specific ROIs were manually drawn based on time-averaged magnitude MRE (M) images and accounting for whole-organ boundaries, as illustrated in Figure  2. Furthermore, these ROIs were refined by empirical thresholds of 1 m/s for softer organs (liver and pancreas) and 1.5 m/s for stiffer organs (spleen and kidneys), to remove blood vessels that appear enlarged in SWS maps. 24,35 Without motion, both M (MRE magnitude averaged over 96 consecutively acquired image slice blocks) and SWS maps are expected to display sharp edges at tissue interfaces, which is reflected in the Laplacian Δ of the images. Therefore, image sharpness was quantified by computing the variance (σ) of the Laplacian Δ of MRE images as described in Pech-Pacheco et al. 36 A Laplacian derivative kernel of size 3 × 3 was applied to both the mean magnitude M (averaged over 96 image slice blocks) and SWS maps. The value of σ was then derived by computing the variance of ΔM and ΔSWS within presegmented ROIs of each organ and protocol.

| Statistical analysis
The mean U, SWS, and sharpness (σ) were calculated for all four breathing paradigms and organs. All values are tabulated as mean ± SD, unless otherwise stated. Statistical significance of differences in displacement and stiffness across all abdominal organs and motion-reduction strategies (protocols) were analyzed using linear mixed-effect models. Linear mixed-effect models accounted for U and SWS as dependent variables, whereas organs and protocols, as well as their interactions, were taken as independent variables. Statistical differences of the dependent variables were tested by Tukey's post hoc test with Bonferroni correction for multiple comparisons. A paired Student t-test was used to test for differences between uncorrected and motion-corrected MRE data based on SWS values. A Wilcoxon signed-rank test was used to test for differences between uncorrected and motion-corrected σ in all organs and protocols. Furthermore, the same test was applied to test for differences between uncorrected BH-MRE and motion-corrected FB-MRE, as well as motion-corrected BH-MRE and motion-corrected FB-MRE in all organs.

| Sharpness
Sharpness was improved by image registration as revealed to the naked eye when looking at both MRE magnitude images (M) and SWS maps (Figure 3-6). Accordingly, σ of ΔM showed a significant reduction of blurring after registration in all organs investigated for FB-MRE and G-MRE (all P < .05). Registration improved the GF-MRE data obtained in the kidneys, spleen, and pancreas (all P < .05), whereas the liver, analyzed based on the first slice, had lower displacement amplitudes in GF-MRE and did not benefit from registration. Interestingly, sharpness even improved in BH-MRE examinations for the kidneys and pancreas, indicating the susceptibility of these organs to unconscious motion during breath-holds ( Figure 9). After motion correction of FB-MRE, σ was indistinguishable from uncorrected BH-MRE in the liver, spleen, and pancreas (all P > .05), while corrected FB-MRE in the kidneys had still lower σ-values than uncorrected BH-MRE (P < .01) despite visual improvements. Motion-corrected BH-MRE shows significantly higher F I G U R E 7 Group statistical plots of displacement amplitudes U (in millimeters and in pixels) of the liver, kidneys, spleen, and pancreas for different MRE protocols (*P < .05, **P < .01, ***P < .001) σ values compared with motion-corrected FB-MRE in the liver, kidneys, and pancreas (all P < .05), whereas no such difference to the spleen was observed. Representative ΔM and ΔSWS images are shown in Supporting Information Figure S1. It is apparent from the examples shown that ΔSWS improved primarily at organ boundaries, indicating sharper edges while within organs, and higher-frequency ΔSWS patterns are seemingly unaffected by the image registration.
All relevant parameters are summarized in Table 1. MRE is performed during free breathing or breath-holds. Finally, we show that unnavigated free-breathing MRE combined with motion correction by image registration represents the optimal trade-off in terms of scan time and detail resolution. In the following, we will briefly discuss each of these three points.

| DISCUSSION AND CONCLUSIONS
(i) Breath-hold MRE results in the smallest displacement amplitudes of abdominal organs compared with free-breathing MRE and navigator-triggered gating MRE techniques. The liver and pancreas are the abdominal organs most severely affected by breathing, whereas the kidneys show the smallest displacement amplitudes. An earlier study of our group shows that multifrequency MRE of the kidneys is improved when omitting any breathing command compared with breath-hold MRE. 25 However, this case-specific finding from a single subject was never reproduced or compared with other abdominal organs, as done in the present study. (ii) All protocols provided values in the range of published values of abdominal MRE (converting SWS to shear modulus by µ = SWS 2 · 1000 kg/m 3 ) 37-39 or directly agreed with reported SWS values in that region. 25,40 The visible blurring of the magnitude images in free-breathing MRE (see  suggests that stiffness values are strongly affected by breathing motion, which, however, is not the case. This observation is remarkable because it indicates the intrinsic stability of MRE to respiration. 41,42 We attribute this stability to two main effects. First, MRE analyzes parameters that refer to local wavelengths, which in turn are unaffected by boundary conditions to a certain extent. Although actuator position, wave amplitudes, and wave patterns can change with breathing, the resulting parameter maps do not necessarily reflect these changes. Second, local wavelengths are calculated from time-harmonic tissue displacements selected from the fundamental frequency in the Fourier spectrum. 16 This is temporal filtering, which efficiently suppresses breathingrelated signals as long as they do not overlap with the timing of the fundamental oscillation. For these reasons, MRE might be more robust to breathing-induced artifact than other MRI techniques. Because the MRE magnitude signals presented in Figures 3-6 are timeaveraged across 96 consecutive acquisitions, similar blurring would occur in any MRI data acquired over the same time during free breathing. (iii) Nonetheless, blurring can sufficiently be corrected by image registration. The resulting improvement in sharpness σ was seen primarily in magnitude images, and to a lesser extent in stiffness maps. The SWS maps predominantly suffered from distorted boundaries, which became sharper after correction, although without significant changes in σ. Therefore, we may conclude that σ of SWS maps tends to exaggerate degradation of images by possible artifacts. For example, breathing-related organ displacement causes highfrequency SWS changes, which are spuriously interpreted by σ as sharpness. Despite the inappropriateness of σ to reflect improvements in stiffness maps, the facts that MRE magnitude images are significantly improved (as correctly quantified by σ) and that anatomical organ boundaries are more conspicuous in SWS maps after image registration (as revealed to the eye when looking at the SWS maps in Figures 3-6) suggest image registration to be a feasible way of reducing motion artifacts in multifrequency MRE.
In essence, we recommend free-breathing multifrequency MRE in conjunction with retrospective rigid-body image registration for abdominal tomoelastography, as it combines the shortest acquisition times with preserved sharpness of organ boundaries and tissue interfaces.
Our study has limitations. First, our displacement analysis was essentially 2D. Given that major breathing motion occurred along the craniocaudal axis, 43 which is covered by coronal view, our 2D in-plane motion correction was effective. However, through-plane motion occurring in transverse orientation needs to be addressed by rigid 3D image registration. Second, our navigator-based motion-correction technique uses single navigators in conjunction with subsequent full-slice block acquisition. For that reason, early slices acquired immediately after the navigators are more reliable than late slices due to the latency between navigator signal and slice acquisition. In our setup, liver motion was evaluated in the first slice, which was well addressed by GF-MRE. However, GF-MRE was less efficient in organs other than the liver, which were covered by images slices acquired at later time points. Although adaptation of slice positions is possible to focus on other organs, we do not recommend GF-MRE but rather FB-MRE combined with image registration. Third, as mentioned previously, Laplacian Δ and variance σ are limited in assessing sharpness of stiffness (SWS) maps. A suitable parameter for the future assessment of detail resolution in MRE should quantify the sharpness of tissue boundaries rather than higher frequency features in SWS maps. Future research in MRE is warranted to identify such a measure of sharpness of elastograms. Ideally, such a measure should afford quantitative assessment of image quality in MRE and ultimately allow automated detection of reliable maps 44 and reduction of technical failures. 45 Finally, we focused here on SWS as a surrogate of stiffness but did not analyze viscosity-related parameters such as wave penetration, 33 damping ratio, 46 loss modulus, 11 and loss angle or fluidity. 14 However, we assume that our analysis of breathing artifacts can also be applied to other mechanical parameters attainable by MRE.
In summary, we have introduced and compared different motion-artifact reduction strategies in multifrequency MRE with the aims of testing the susceptibility of MRE to breathing motion and proposing a time-efficient multifrequency MRE protocol that generates reliable stiffness maps. Our findings indicate that larger displacement amplitudes due to respiratory motion do not significantly affect averaged MRE values and that breathing motion can readily be corrected by image registration, which improves both the sharpness and detail resolution of MRE magnitude images and parameter maps. Consequently, the shortest free-breathing protocol is recommended for 3D multifrequency MRE of the abdomen and allows stiffness mapping without breathing commands in approximately 2 minutes.