FULL PAPER Magnetic Resonance in Medicine 73:995–1004 (2015)
Accelerated Human Cardiac Diffusion Tensor Imaging Using Simultaneous Multislice Imaging Angus Z. Lau,1,2* Elizabeth M. Tunnicliffe,1 Robert Frost,3 Peter J. Koopmans,3 Damian J. Tyler,1,2 and Matthew D. Robson1 INTRODUCTION
Purpose: To demonstrate the feasibility of accelerating measurements of cardiac fiber structure using simultaneous multislice (SMS) imaging. Methods: SMS excitation with a blipped controlled aliasing (CAIPI) readout was incorporated into a diffusion-encoded stimulated echo pulse sequence to obtain diffusion measurements in three separate slices of the heart (8-mm thickness, 12-mm gap). A novel image entropy-based method for removing image ghosts in blipped CAIPI acquisitions is also introduced that enables SMS imaging of closely spaced slices in the heart. Results: The average retained signal-to-noise ratio (SNR) using this acquisition scheme is 70% 6 5%, higher than the pﬃﬃﬃ standard 1= 3 ¼ 57% SNR penalty with three-fold acceleration. No significant difference was observed in the apparent diffusion coefficient and helix angle diffusion parameters between a time-equivalent conventional single-slice scan and the three-fold accelerated SMS acquisition. A 10% mean bias was observed in fractional anisotropy between single-slice and SMS acquisitions. Conclusion: The new sequence was used to obtain highquality diffusion measurements in three closely spaced cardiac slices in a clinically feasible nine breath-hold examination. The accelerated multiband sequence is anticipated to improve quantitative measurements of cardiac microstructure by reducing the number of breath-holds required for the scan, making it practical to incorporate diffusion tensor measurements within a comprehensive clinical examination. C Magn Reson Med 73:995–1004, 2015. V 2014 Wiley Periodicals, Inc. Key words: diffusion; cardiac; parallel imaging; simultaneous multislice; blipped CAIPI; multiband
1 Department of Cardiovascular Medicine, Oxford Centre for Clinical Magnetic Resonance Research, John Radcliffe Hospital, University of Oxford, UK. 2 Department of Physiology, Anatomy, and Genetics, University of Oxford, UK. 3 Nuffield Department of Clinical Neurosciences, Oxford Centre for Functional MRI of the Brain, University of Oxford, UK. Grant sponsor: National Institute for Health Research; Grant sponsor: Oxford Biomedical Research Centre Programme; Grant sponsor: British Heart Foundation Fellowship; Grant number: FS/10/002/28078, Programme Grant RG/11/9/28921; Grant sponsor: The Wellcome Trust; Grant number: WT100092MA. *Correspondence to: Angus Z. Lau, Department of Cardiovascular Medicine, OCMR, John Radcliffe Hospital, Oxford, OX3 9DU. E-mail: [email protected]
Received 7 November 2013; revised 6 February 2014; accepted 13 February 2014 DOI 10.1002/mrm.25200 Published online 21 March 2014 in Wiley Online Library (wileyonlinelibrary. com). C 2014 Wiley Periodicals, Inc. V
Characterization of cardiac microstructure using diffusion tensor MRI enables noninvasive assessment of myocardial fiber structure. In healthy hearts, myocardial fibers transition from a subepicardial left-handed helix to a subendocardial right-handed helix (1,2). Diffusionweighted imaging enables detection of cardiac fiber structure in both ex vivo (3–7) and in vivo studies (8– 13). This structure is disrupted in human heart disease, including hypertrophic cardiomyopathy (14) and myocardial infarction (15,16). In the diseased state, remodeling of myocardial tissue contributes to impaired cardiac function, as well as alterations in electrophysiology, leading to development of life-threatening arrhythmias and sudden cardiac death (17). Diffusion encoding using stimulated echoes (STEAM) is a robust method for reducing sensitivity of fiber structure measurements to bulk cardiac motion (8–11,13,18). In this method, diffusion-prepared magnetization after one R-wave beat is stored along the B0-axis and is imaged using a single-shot echo-planar imaging (EPI) readout after a second R-wave. However, the method suffers from inherently low signal-to-noise ratio (SNR) due to the stimulated echo encoding and the long diffusion times required. Technical advances that include the introduction of 3T scanners and 32-channel array coils have improved the situation, but multiple signal averages over several breath-holds are still required even for a single slice. Moreover, in the clinic, volumetric (multislice) coverage is useful when studying and diagnosing cardiac disease where the exact spatial location of disease is unknown. Volumetric acquisitions inherently allow for the comparison of measured parameters in diseased to remote regions, providing an intrasubject baseline measurement with each scan. Assessment of cardiac microstructure in basal, mid-, and apical slices requires 24 breath-holds (15 min), limiting the utility of this approach for routine clinical scanning. An exciting possibility to accelerate diffusion MRI in the heart is the simultaneous acquisition of multiple slices (SMS) (19). In SMS, multiple slices are excited and separated by parallel imaging principles using information from multiple receiver coils. In cardiac MRI, the limited field of view (FOV) in the slice direction, however, leads to a challenging situation with insufficient variation in coil sensitivity between slices. Recently, controlled aliasing (CAIPI) methods have been shown to reduce the noise amplification in the SMS experiment. The CAIPI strategy uses phase modulation of simultaneously excited slices (20), either by RF pulse design or by through-plane phase encoding (blipped CAIPI) (21). This
Lau et al.
FIG. 1. Multiband cardiac-gated diffusionencoded STEAM sequence with a singleshot EPI readout. In this sequence, the FOV along the phase encode direction is restricted using two slab-selective 90 pulses, which store diffusion-prepared magnetization along Mz during the mixing time TM. In the next cardiac cycle, the prepared stimulated echo contrast is excited using a multiband RF pulse along the slice direction. The simultaneously excited slices are given a unique phase cycle along ky by controlled aliasing.
phase modulation produces interslice image shifts that improve the conditioning of the parallel imaging reconstruction. In this study, we have investigated the feasibility of SMS imaging to accelerate in vivo cardiac diffusion MRI. We have tested whether the combination of blipped CAIPI with a restricted FOV-stimulated echo acquisition allows the expected three-fold acceleration over a conventional stimulated echo scan, and quantitatively investigated the SNR improvements that were gained. A novel image entropy-based method for removing image ghosts in blipped CAIPI acquisitions is also introduced that enables SMS imaging of closely spaced slices in the heart. The new sequence was used to obtain high-quality diffusion measurements in three cardiac slices in a clinically feasible nine breath-hold examination. METHODS Multiband Diffusion-Encoded Pulse Sequence Figure 1 shows the pulse sequence used in this study. Multiband excitation was incorporated into a diffusionencoded STEAM sequence with single-shot EPI readout (13) on a clinical 3T scanner (Trio, Siemens, Erlangen, Germany). In this sequence, the field of view (FOV) in the phase encode direction was restricted by using two slab-selective 90 pulses. The first two 90 pulses encoded a spatial region that was subsequently refocused by the third excitation RF pulse, generating a stimulated echo. The excitation RF pulse was designed using the Shinnar Le-Roux algorithm (22) using the following parameters: time-bandwidth ¼ 4; slice thickness ¼ 8 mm; duration ¼ 4 ms; 103 passband ripple; 102 stopband ripple; peak B1 ¼ 0.006 mT). Off-center slice excitation was performed by phase modulation of the RF waveform by the gradient waveform, and a multiband RF excitation pulse was constructed by summing together phasemodulated (single-band) waveforms in-phase exciting each desired spatial location. In this process, the peak B1 scaled linearly with the number of excited slices. The blipped CAIPI method was used to shift each slice relative to each other in the phase encode direction so as to reduce g-factor noise amplification. Diffusion encoding is achieved by changing the amplitude and direction of
the stimulated echo encoding gradients (labeled GD). This non-zero gradient pair is used to refocus the stored diffusion-prepared magnetization and spoil the FID from the final 90 pulse, resulting in a minimum achievable bvalue of 20 s/mm2. Data Acquisition A total of 10 healthy male volunteers were scanned (age range, 28–45 y), and all studies were approved by the local institutional ethics committee. Single-slice reference data for each of the three individual slices was obtained in a single separate breath-hold. The acquisition for the reference data consisted of two heartbeats for EPI phase correction and two heartbeats for a b ¼ 20 s/ mm2 image, repeated for each slice for a total of 12 heartbeats. Multiband diffusion-encoded data was obtained in a similar fashion, and the single breath-held acquisition lasted 16 heartbeats, consisting of two heartbeats for EPI phase correction, two heartbeats for a b ¼ 20 s/mm2 image, and 12 heartbeats for six b ¼ 350 s/mm2 images. At each spatial location, one b ¼ 20 s/mm2 image and six b ¼ 350 s/mm2 diffusion-encoded directions were used to calculate the diffusion tensor. The imaging acquisition was repeated using a single-slice excitation pulse for each slice to evaluate data quality. A 32-channel anterior/posterior cardiac receiver array with body coil transmission (3T Trio, Siemens, Erlangen, Germany) was used for imaging. Images were acquired in mid systole. The imaging parameters were as follows: echo time ¼ 13 ms; repetition time ¼ 2 RR intervals; matrix size ¼ 128 42; partial Fourier factor ¼ 5/8; FOV ¼ 360 180 mm2; through-plane slice thickness ¼ 8 mm; slice gap ¼ 12 mm; phase-encode slab thickness ¼ 120 mm; spatial resolution¼ 2.8 2.8 mm2; EPI bandwidth ¼ 2448 Hz/pixel. To minimize aliasing and reduce g-factor noise amplification, the phase encode direction was placed parallel to the chest wall, from the right ventricle to the left ventricle of the heart. An FOV/3 interslice image shift is produced by blipped CAIPI modulation. This corresponds to a 60-mm shift in the phase encode direction. The left ventricular shim volume and center frequency were kept constant for each acquisition to ensure that B0 image distortions remained the same for both multiband and reference data.
Multiband Human Cardiac Diffusion Tensor Imaging
FIG. 2. Multiband scan protocol. Following standard cardiac localizers (seven breath-holds), single-slice comparison data (15 breath-holds) was obtained in basal, mid-, and apical short axis views. Multiband diffusion-encoded data was obtained (nine breath-holds), with one breath-hold used to obtain b ¼ 20 s/mm2 data in each slice position for multiband reconstruction. Each breath-hold lasted a total of 16 heartbeats, containing a full diffusion-encoded set for diffusion tensor analysis.
The multiband diffusion-encoded scan protocol used to evaluate these methods is shown in Fig. 2, and the scan parameters for each of the scans are summarized in Table 1. The protocol design allows generation of three diffusion-encoded datasets: 1. midventricular slice, with nine averages requiring nine breath-holds; 2. three slices, each acquired separately, with three averages, each requiring nine breath-holds; 3. multiband three-slice acquisition, with eight averages requiring nine breath-holds (one used for reference data) These datasets enable a fair comparison of image quality for the midventricular slice, based on equivalent total scan time. An additional noise measurement was made with the same EPI readout but with RF pulses turned off for SNR characterization using the pseudo multiplereplica method (21,23).
entropy of the resulting sum-of-squares image. This is a one-parameter model, and we obtain the phase step size that corresponds to the corrected image by fitting the entropy values to an empirical sine curve. Figure 3a–3c shows the automatic identification of this phase step, and removing the phase modulation from the k-space data removes the three-fold image ghost. For this particular case, we display the resulting phase staircase (Fig. 3d) and its corresponding point spread function with three-fold ghosting. This phase correction is repeated for each slice using the single-slice reference data, the results are averaged, and the average phase correction is applied to all acquired datasets. Single slice b ¼ 20 s/mm2 images at each spatial location were obtained using the single-band pulse and used as reference data for the multiband reconstruction. Image reconstruction was performed using the split-slice GRAPPA method (28). In this method, individual GRAPPA kernels are computed by finding kernel weights wi (dimensions [wx,wy]) that simultaneously best match the reference data and minimize signal from neighboring slices. This is formulated as a least-squares matrix inversion and optimization problem,
Table 1 Multiband Scan Protocol Scan
Description and Parameters
Three-plane scout, long axis and short axis views, long axis cine, 3D shimming Basal (three breath-holds), mid- (nine breath-holds), apical (three breath-holds) short axis views Imaging parameters: echo time ¼ 13 ms; repetition time ¼ 2 RRs; matrix size ¼ 128 42; partial Fourier factor ¼ 5/8; FOV ¼ 360 180 mm2; through-plane slice thickness ¼ 8 mm; slice spacing ¼ 12 mm; phase encode slab thickness ¼ 120 mm; spatial resolution ¼ 2.8 2.8 mm2; EPI bandwidth ¼ 2448 Hz/pixel Short axis views (eight breath-holds) Single-slice reference data (one breath-hold) Imaging parameters: same as single-slice scan, but with three simultaneously excited slices
Image Reconstruction Image reconstruction was performed off-line in MATLAB. Prior to image reconstruction, the acquired data were corrected for three sources of phase variation across the phase encode direction: 1. EPI phase correction from opposite polarity readout gradients (Nyquist ghosting); 2. off-isocenter phase correction similar to off-center three-dimensional (3D) slab excitation; 3. phase correction due to eddy currents induced by the blipped CAIPI kz blips. Internal EPI readout phase correction (24,25) and offisocenter phase correction in blipped CAIPI acquisitions (21,26) have been described in the literature. In Fig. 3, we corrected the phase error from the blipped CAIPI acquisition by introducing a simple deconvolution model and novel image entropy-based minimization method. Image entropy has previously been proposed as an image metric for Nyquist ghost removal (27). For the FOV/3 shift used in this study, the predominant contribution to the phase error comes from the largest kz blip. We model the phase error that accumulates across k-space while traversing the phase encode direction as being a staircase function with period 3. We step through the full range of phase errors as a function of the step size, and measure the image
Following standard cardiac localizers (seven breath-holds), singleslice comparison data (15 breath-holds) were obtained in basal, mid-, and apical short axis views. Multiband diffusion-encoded data was obtained (nine breath-holds), with one breath-hold used to obtain b ¼ 20 s/mm2 data in each slice position for multiband reconstruction. Each diffusion-encoded breath-hold lasted a total of 16 HBs, containing a full diffusion-encoded set for diffusion tensor analysis.
Lau et al.
FIG. 3. Image entropy-based approach to correct eddy current related phase errors in a blipped CAIPI acquisition. a: A phase staircase function is constructed and used to modulate the acquired k-space data across the phase encode direction. b,c: The phase step that corresponds to minimum image entropy is identified and used to remove the three-fold image ghost (b) resulting in the corrected image (c). The images are cropped to 180 180 mm2. d: Phase staircase with step size 1.3 rad, and corresponding magnitude point spread function with three-fold ghosting which corresponds to this phase error.
X wi ¼ argmin jjsi ½Ai wi jj2 þ l j6¼i jj Aj wi jj2
Smb;i ¼ ½Amb wi :
where Si is the reference k-space data for slice i, [Ai] represents the convolution matrix for slice i, and l is a regularization parameter that balances data consistency with slice leakage artifact. In this formalism, [Ai] is a reordered matrix (dimensions [wxwyNc, NxNyNc]) in which rows represent k-space neighborhoods corresponding to the fitted GRAPPA kernel weights. Nc is the number of coils, and (Nx,Ny) denote the dimensions of the acquired k-space signal. The set of kernels is then applied to each aliased multislice frame (here, denoted by the convolution matrix [Amb]) to recover unaliased images at each slice location (k-space data Smb,i) according to
Slice separation was performed for each multiband dataset, before averaging, to avoid phase cancellation due to motion-related phase errors. Following slice separation, each slice was shifted back to the center of the FOV by undoing the blipped CAIPI phase modulation across k-space. The reconstruction parameters used for multiband reconstruction were: l ¼ 1, slice-GRAPPA kernel size ¼ 11x5 k-space points. Projection onto convex sets (POCS) was used to fill in missing k-space lines in the partial Fourier reconstruction. Coil sensitivities were extracted using the b ¼ 20 s/mm2 images for each breathhold (29) and used for coil combination for the diffusion weighted images within the breath-hold. SNR Characterization
FIG. 4. Retained SNR maps for multiband reconstruction with no FOV shift (top) and with an FOV/3 shift (bottom) produced by controlled aliasing. pﬃﬃﬃ The color scale changes from colored to grayscale at 57% (1= 3), the break-even point for an SNR advantage between a three-fold accelerated acquisition and the corresponding single-slice acquisition. The images are cropped to 180 180 mm2.
SNR performance was characterized using the pseudomultiple replica method (21,23). In this method, noise samples are drawn from a measured noise distribution, added to a low-noise dataset, and the resulting series of synthesized images are used to compute the signal-tonoise reduction due to multiband reconstruction. For each slice, four b ¼ 20 s/mm2 images were averaged to obtain a low-noise dataset. Multiband (slicealiased) data were synthesized by summing together the k-space data from each slice. Different relative FOV shifts were simulated (no shift and FOV/3) by applying the correct phase modulation across k-space. A sliceGRAPPA kernel was calibrated on each of the reference images and was used for subsequent multiband reconstruction. Noise was added to the simulated multiband k-space data according to the noise covariance matrix across coils. SNR was computed in a voxel-wise fashion
Multiband Human Cardiac Diffusion Tensor Imaging
was used to obtain the diffusion tensor and the signal intensity at b ¼ 0. Apparent diffusion coefficient (ADC), fractional anisotropy (FA), and helix angle (HA) maps were then computed as described previously from the diffusion tensor (3,4). The measured ADC and FA values were measured in a region of interest covering the entire left ventricle in the middle slice, enabling a statistical comparison between multiband data requiring nine breath-holds, reference data requiring three breath-holds (the multislice time equivalent scan), and reference data requiring nine breath-holds (the single-slice time equivalent scan). The helix angle maps were analyzed by fitting a linear model across radial projections over the entire left ventricle in the middle slice. The radial helix angle gradient from this fit reflects the linear evolution of helix angle through the myocardial wall, which has previously been observed ex vivo (30) and in vivo (9). A nonparametric Friedman test was performed to detect differences in the measured ADC, FA, and radial HA gradient values across subjects between each method. Statistical significance was considered to be P < 0.05. RESULTS
FIG. 5. Diffusion-weighted short axis images of the heart using multiband reconstruction and single-slice acquisition. The field of view of the images is 360 180 mm2. The phase encode direction is parallel to the chest wall, from the right ventricle to left ventricle. The acquired spatial resolution of the images is 2.8 2.8 8 mm3.
by dividing mean magnitude signal by the standard deviation over 128 pseudoreplicas. The g-factor associated with the multiband reconstruction was defined by dividing the SNR in the reference acquisition by the SNR in the corresponding multiband acquisition. Using this procedure, we obtained maps of 1/g (retained SNR) for each controlled aliasing pattern. The average 1/g value was then computed over a region of interest containing the entire myocardium. Data Analysis We assumed that the only source of motion was between breath-holds and that acquired images within a single breath-hold were consistent. Under this assumption, rigid registration (in-plane translation and rotation) was used to align both multiband and reference b ¼ 20 s/mm2 images from each breath-hold to a single reference b ¼ 20 s/mm2 image. The registration parameters were then used to align the remaining diffusion-weighted images to the initial b ¼ 20 s/mm2 image in each breath-hold. The images were interpolated by an in-plane factor of 2 (resultant voxel dimensions ¼ 1.4 1.4 8 mm3). Voxelwise fitting of the image data (seven acquired diffusionweighted images: one b ¼ 20 s/mm2 direction and six b ¼ 350 s/mm2 directions) to an elliptical tensor model
Figure 4 shows the effects of the interslice FOV shift on SNR in the multiband images compared with a singleslice pﬃﬃﬃacquisition. When the retained SNR figure is >57% (1= 3), the multiband acquisition is superior (in terms of maximising SNR over a fixed time period) to the corresponding time-equivalent single slice acquisition with the same spatial coverage. The retained SNR over the left ventricle (mean 6 SD over subjects) was significantly different with controlled aliasing, improving from 36% 6 8% with no FOV shift to 70% 6 5% with the FOV/3 shift (P < 104). This demonstrates superior SNR in the multiband acquisition relative to a time-equivalent single-slice acquisition with the same spatial coverage. Figure 5 shows unaliased multiband images and the corresponding single-slice reference images, demonstrating the close correspondence between the two scans. Figure 6 shows multiband b ¼ 350 s/mm2 images in the midventricular slice in 10 scanned volunteers to demonstrate reconstructed image quality across subjects. Figure 7 shows apparent diffusion coefficient (ADC), fractional anisotropy (FA), and helix angle (HA) maps in the heart derived from diffusion-weighted images obtained using a nine-breath-hold multiband acquisition, with comparison to single-slice data obtained in the same views using either three breath-holds per slice (equivalent scan time to the full multiband dataset) or
FIG. 6. Diffusion-weighted b ¼ 350 s/mm2 images in 10 subjects, using multiband reconstruction and single-slice acquisition. The acquired spatial resolution of the images is 2.8 2.8 8 mm3. The images are cropped to 120 120 mm2.
Lau et al.
FIG. 7. ADC, FA, and HA maps in the heart for multiband reconstructed images, with comparison to single-slice data obtained in the same views using either three breath-holds per slice (equivalent scan time to the full multiband dataset) or nine breath-holds for the midventricular slice.
nine breath-holds for the midventricular slice. Qualitatively, the maps generated by each technique appear similar. The helix angle maps clearly demonstrate the transition from subendocardial left-handed helix angle to subepicardial right-handed helix in the healthy myocardium. We analyzed the mean ADC, FA, and radial HA gradient values in the midventricular slice across subjects using multiband acquisition and the corresponding single-slice acquisitions (Fig. 8). A nonparametric Friedman test revealed a significant difference in FA between multiband and reference data with a mean bias of 0.05 (10% difference), and no significant difference in ADC and radial HA gradient values. Table 2 shows individual ADC, FA, and radial HA gradient values from each subject. Furthermore, Table 2 shows that the standard deviation of ADC and FA across subjects is lower in the multiband acquisition compared with the timeequivalent multislice acquisition, presumably due to the SNR increase revealed in Figure 4.
It is anticipated that this development will improve quantitative measurements of cardiac microstructure by reducing the number of breath-holds required for the scan. Accelerated cardiac DTI will enable studies to
DISCUSSION In vivo measurements of cardiac microstructure are challenging because large bulk physiological motions restrict available imaging time to quiescent periods of the cardiac cycle. Furthermore, existing methods are limited to relatively healthy patient populations who can tolerate scans requiring many breath-holds, limiting translation of diffusion MRI into the clinic. In this study, we demonstrated the feasibility of using simultaneous multislice excitation to accelerate a clinical diffusion tensor imaging (DTI) protocol by three-fold. When compared with a conventional three-slice acquisition without multiband excitation, the diffusion parameter maps (ADC, FA, and HA) are visually similar. A statistical analysis reveals no significant difference in ADC and radial HA gradient values using the accelerated sequence, and a slight bias (10%) in FA values from the accelerated sequence relative to a time-equivalent single-slice scan. A noise-induced positive bias in fractional anisotropy values is well known, and is likely due to the scan time reduction per slice from multiband acceleration (31).
FIG. 8. Bland-Altman plots comparing mean ADC (a), FA (b), and radial HA gradient (c) values in the midventricle across 10 subjects for multiband reconstructed images (nine breath-holds) versus single-slice data obtained in the same view (nine breathholds). A nonparametric Friedman test revealed a significant difference in FA between multiband and reference data (mean bias 0.05), and no significant difference in ADC and radial HA gradient.
Multiband Human Cardiac Diffusion Tensor Imaging
Table 2 Diffusion Parameters Across Subjects in the Midventricular Slice Measured Using Multiband and Single-Slice Reference Acquisitions ADC (103 mm2/s) Subject 1 2 3 4 5 6 7 8 9 10 Mean SD
Radial HA gradient ( /mm)
1.11 6 0.28 1.37 6 0.35 1.24 6 0.27 1.09 6 0.27 1.10 6 0.19 0.96 6 0.18 1.31 6 0.27 1.46 6 0.24 1.03 6 0.21 1.21 6 0.32 1.19 0.16
1.55 6 0.32 1.31 6 0.20 1.57 6 0.55 1.11 6 0.24 1.23 6 0.15 1.06 6 0.14 1.16 6 0.21 1.27 6 0.30 1.03 6 0.18 1.32 6 0.37 1.26 0.19
1.17 6 0.23 1.21 6 0.16 1.34 6 0.28 1.10 6 0.20 1.22 6 0.16 1.07 6 0.11 1.17 6 0.21 1.30 6 0.22 1.03 6 0.14 1.29 6 0.28 1.19 0.10
0.55 6 0.15 0.37 6 0.12 0.50 6 0.13 0.48 6 0.13 0.45 6 0.11 0.48 6 0.11 0.53 6 0.14 0.55 6 0.13 0.49 6 0.14 0.47 6 0.13 0.48 0.05
0.58 6 0.13 0.36 6 0.12 0.59 6 0.13 0.50 6 0.14 0.39 6 0.08 0.43 6 0.11 0.50 6 0.12 0.54 6 0.13 0.50 6 0.12 0.43 6 0.12 0.48 0.08
0.50 6 0.13 0.38 6 0.11 0.54 6 0.12 0.44 6 0.13 0.39 6 0.09 0.41 6 0.08 0.45 6 0.11 0.48 6 0.10 0.44 6 0.11 0.38 6 0.09 0.44 0.05
3.6 6 2.0 5.5 6 3.5 2.6 6 2.5 3.9 6 2.2 4.5 6 2.2 4.0 6 1.3 7.9 6 6.6 2.4 6 1.8 5.2 6 1.7 3.9 6 1.8 4.4 1.6
4.1 6 4.2 3.9 6 3.1 2.6 6 3.9 4.3 6 2.5 4.3 6 1.9 4.7 6 1.7 8.2 6 10.8 3.0 6 2.0 5.3 6 5.8 4.3 6 2.1 4.5 1.5
3.8 6 3.8 4.2 6 3.3 1.8 6 2.4 4.5 6 2.5 4.2 6 1.9 4.5 6 1.2 7.6 6 11.6 3.0 6 1.6 5.3 6 1.7 4.3 6 1.5 4.3 1.5
The numbers in parentheses indicate the number of breath-holds required for each acquisition. The mean and standard deviation over the myocardium of each diffusion parameter is listed for each subject, and the group mean and standard deviation of the mean is listed as a measure of intersubject variability. Abbreviations: MB, multiband; Ref, reference.
assess the potential of the technique to better diagnose, monitor, and treat cardiac disease in a larger patient population. Cardiac DTI measurements are SNR-limited, and the minimum number of breath-holds per slice (ie, minimum number of signal averages) reported previously has been at least eight (13). Assessment of cardiac microstructure in three slices thus requires 24 breath-holds. Although the nine breath-holds used in this study are far from ideal, this is strong clinical motivation to implement and refine techniques that accelerate DTI measurements in the heart. Furthermore, although three slices is not optimal in terms of full heart coverage, this is equal to the highest number reported previously (13,18). We envision that whole heart coverage (six slices, 8-mm slice thickness, 2-mm slice gap) could be obtained by interleaving two thee-slice SMS scans in a total scan time of 18 breath-holds, which would enable whole-heart clinically feasible DTI in patients. A free-breathing acquisition with respiratory navigation or synchronization would have clinical utility, especially in patients who cannot tolerate breath-holding. We note that navigator-based stimulated echo encoding requires that the heart remain stationary over the complete cardiac cycle, leading to very low navigator efficiencies on the order of 40% (32). Also, even with modified acceptance parameters, there is some bulkmotion corruption that contributes to reduced reproducibility of diffusion parameters in a free-breathing acquisition (13). Despite these obstacles, SMS scan time reduction would improve relative navigator efficiency by obtaining more data when inside the acceptance window, and is another strong motivator for the development of SMS for cardiac DTI. Care was taken to keep both shim volume and center frequency constant between multiband and reference acquisitions. This ensures that the multiband acquisition has the same B0 image distortion and shift as the corresponding single-slice references. B0 image distortion artifacts were present in the acquired data, but they did not prevent data analysis. Off-resonance image distortion is a
result of several factors. First, the shim volume is chosen to shim the left ventricle in all three slices for the multiband scan, reducing performance compared with sliceby-slice shimming. Second, the 13-ms echo time leads to increased image distortion due to the 180-mm phaseencode FOV used. For the multiband accelerated acquisition to become clinically useful, robust B0 distortion correction will be required as regions of the myocardium affected by distortion become nondiagnostic. To address this, B0 distortion correction can be implemented, either by using the B0 field map from the 3D shimming acquisition or by self-calibrating methods such as phase-encode polarity reversal in alternate breath-holds (33). Our current protocol provides neither a B0 map nor the blip-up/ blip-down data necessary for distortion correction, but this would be a useful feature for clinical implementation. The minimum slice thickness for the multiband RF excitation used in this study was 1 mm. In principle, the performance of blipped CAIPIRINHA improves as the ratio between slice thickness and slice gap decreases, due to reduced dephasing across slices by the throughplane CAIPI blips. In practice, the minimum slice thickness is limited by SNR considerations, and we chose the 8-mm slice thickness used here based on literature values (13). The number of simultaneously excited slices is limited by the number of receiver channels available. Here, a 32channel cardiac receive array was sufficient to unalias three simultaneously excited slices. The blipped CAIPI technique reduces the parallel imaging g-factor penalty by creating an interslice image shift in the phase encoding direction. The g-factor simulations using the pseudo multiple replica method reported an average retained SNR of 70% when an FOV/3 shift pﬃﬃﬃwas used. This figure is higher than the standard 1= 3 ¼ 57% SNR penalty when reducing scan time by a factor of 3, indicating that the use of multiband is more SNR-efficient than acquiring the single slices in sequential breath-holds. The benefit of this SNR gain is presumably indicated in Table 2, which shows reduced standard deviation in ADC and
FA across subjects using the multiband acquisition. In the neuroimaging literature, slice-acceleration factors from 3 to 12 have been reported using commercially available 32-channel head coil arrays, and investigating the limits of this technique in the heart is the subject of future work. In this study, the phase encoding direction was chosen to be parallel to the chest wall to minimize g-factor noise amplification. We note that the anterior–posterior choice for phase encoding direction, perpendicular to the chest wall, leads to aliasing of the bright wall on top of the myocardium. This is a challenging image reconstruction scenario, likely because the 32-channel cardiac receive array elements are distributed along the chest wall. This can be avoided and the g-factor can be reduced by phase encoding along an axis with the most coil sensitivity variation. Image reconstruction was performed using the splitslice GRAPPA method. This method estimates individual GRAPPA kernels from single-slice reference data and uses these kernels to synthesize separated slice data from the collapsed multiband k-space data. The approach assumes that the sum of the reference data is a good representation of the multiband data. In particular, the phase relationship between each slice needs to be consistent between reference and multiband data. We used b ¼ 20 s/mm2 images as reference data and obtained all reference data in a single breath-hold to align the heart with the multiband acquired data. This ensures that the phase relationship between slices in the reference images is consistent with the multiband data, which enables an accurate reconstruction. Robust phase correction is required to obtain accurate image reconstruction of echo-planar SMS data. When additional phase encode blips are introduced, as in a blipped CAIPI readout, image ghosting can result from imperfections in gradient waveforms with opposite polarity. The effect is especially pronounced when the FOV in the through-plane direction is small (ie, small slice gap to slice thickness ratio). For an FOV/3 image shift, the large gradient blips in the through-plane direction contribute a phase error that can be modeled by a phase staircase function across the phase encode direction. The effect of this phase error convolves the true image with a three-fold image ghost, which overlaps with the desired multiband aliasing pattern. Furthermore, the periodic symmetry of the Fourier transform is broken by this phase modulation, which can lead to blurring when uncorrected. We introduced a simple deconvolution model and novel image entropy-based approach to remove this image ghost. Minimization of the image entropy focuses the image energy into a single region. The black-blood diffusion-weighted images with reduced FOV are especially suited to this approach because there are large regions in the image with no signal. In this study, an empirical sine curve was used to characterize and minimize the image ghost, but a model-free approach may also be feasible by simply choosing the image with minimum image entropy. The method is similar to imagebased, data-driven approaches in multishot EPI (34) in which minimizing image intensity in signal-free
Lau et al.
background regions calibrates for motion-related inconsistencies between shots. The main advantage of this approach is that a separate phase-encoded reference scan is not required, reducing scan time. Two-dimensional echo-planar phase correction methods may also be useful for full characterization of the phase error across the phase encode direction (35,36). Simultaneous multislice acquisition can be viewed as sampling a three-dimensional k-space in which the kz (through-plane) direction is traversed by the FOV shifting blips (26,37). Unaliasing of multiband data can therefore be achieved using 3D SENSE (38) or by obtaining fully sampled (autocalibrating) central k-space data for GRAPPA reconstruction (26,39). SENSE reconstructions require accurate knowledge of the coil sensitivity profile for each slice. We found that using the coil sensitivity maps estimated from individual channels (29) in a SENSE reconstruction produced similar results to the split-slice GRAPPA method. Obtaining autocalibrating single-shot data leads to much longer readout duration, however, increasing B0-induced EPI image distortion. The multiband excitation pulse was constructed by summing in-phase RF pulses exciting each spatial location. The peak B1 for the pulse is proportional to N, the number of slices. In this study, there was sufficient RF transmitter power for up to three simultaneously excited slices, and SAR limits were not reached due to the low duty cycle of the scan. Improved coil array designs may enable higher slice acceleration factors, motivating power-optimized multiband pulse designs. Poweroptimized pulses via phase pﬃﬃﬃﬃﬃ modulation (40) approach peak B1 proportional to N for large N, leading to potentially shorter pulse duration and reduced B0 distortion by reducing echo time. The peak B1 requirement can also be reduced by using the VERSE algorithm (41), but off-resonance slice profile degradation may influence image quality. Improved multiband designs using the Shinnar Le-Roux algorithm (42,43) may also enable adaptive resolution and imaging of closely spaced slices in the heart. The acquisition is compatible with in-plane undersampling (21). When this is done, the interslice image shift is reduced by the degree of in-plane acceleration. In particular, 2 in-plane acceleration combined with 3 multiband excitation results in an FOV/6 interslice image shift. Image quality depends on whether sufficient coil sensitivity variations exist in the phase encode direction. In this study, six diffusion directions were used to compute the diffusion tensor at each location in the myocardium. Although the use of six directions results in reproducible ADC and FA values (13,18), additional diffusion directions will provide a more accurate description of cardiac microstructure (44). In vivo studies with more than six diffusion directions are challenging because of the long scan times required. The threefold SMS scan time reduction demonstrated in this study can be used to obtain additional diffusion directions compared with a conventional three-slice, non–multiband-accelerated cardiac DTI scan. Interestingly, we note that the fractional anisotropy of the myocardium is not expected to approach values in the brain because of the relatively free diffusion of water in the tissue, and this
Multiband Human Cardiac Diffusion Tensor Imaging
may influence the optimal number of directions in the heart required for a robust description of fiber structure. CONCLUSIONS The feasibility of accelerating measurements of cardiac fiber structure using simultaneous multislice imaging was demonstrated and quantitatively evaluated. A multiband excitation pulse, along with a blipped controlled aliasing (CAIPI) readout, was incorporated into a diffusion-encoded stimulated echo pulse sequence to obtain diffusion measurements in three separate slices of the heart. A novel image entropy-based method for removing image ghosts in blipped CAIPI acquisitions is also introduced that enables SMS imaging of closely spaced slices in the heart. The average retained SNR using this acquisition scheme is 70% 6 5%, higher than pﬃﬃﬃ the standard 1= 3 ¼ 57% SNR penalty when reducing scan time by a factor of 3. The accelerated multiband sequence is anticipated to improve quantitative measurements of cardiac microstructure by reducing the number of breath-holds required for the scan, making it practical to incorporate diffusion tensor measurements within a comprehensive clinical examination. REFERENCES 1. Streeter DD Jr, Spotnitz HM, Patel DP, Ross J Jr, Sonnenblick EH. Fiber orientation in the canine left ventricle during diastole and systole. Circ Res 1969;24:339–347. 2. Streeter DD Jr, Hanna WT. Engineering mechanics for successive states in canine left ventricular myocardium. II. Fiber angle and sarcomere length. Circ Res 1973;33:656–664. 3. Scollan DF, Holmes A, Winslow R, Forder J. Histological validation of myocardial microstructure obtained from diffusion tensor magnetic resonance imaging. Am J Physiol 1998;275:H2308–H2318. 4. Holmes AA, Scollan DF, Winslow RL. Direct histological validation of diffusion tensor MRI in formaldehyde-fixed myocardium. Magn Reson Med 2000;44:157–161. 5. Garrido L, Wedeen VJ, Kwong KK, Spencer UM, Kantor HL. Anisotropy of water diffusion in the myocardium of the rat. Circ Res 1994; 74:789–793. 6. Hsu EW, Muzikant AL, Matulevicius SA, Penland RC, Henriquez CS. Magnetic resonance myocardial fiber-orientation mapping with direct histological correlation. Am J Physiol 1998;274:H1627–1634. 7. Pop M, Ghugre NR, Ramanan V, Morikawa L, Stanisz G, Dick AJ, Wright GA. Quantification of fibrosis in infarcted swine hearts by ex vivo late gadolinium-enhancement and diffusion-weighted MRI methods. Phys Med Biol 2013;58:5009–5028. 8. Edelman RR, Gaa J, Wedeen VJ, Loh E, Hare JM, Prasad P, Li W. In vivo measurement of water diffusion in the human heart. Magn Reson Med 1994;32:423–428. 9. Reese TG, Weisskoff RM, Smith RN, Rosen BR, Dinsmore RE, Wedeen VJ. Imaging myocardial fiber architecture in vivo with magnetic resonance. Magn Reson Med 1995;34:786–791. 10. Tseng WY, Reese TG, Weisskoff RM, Wedeen VJ. Cardiac diffusion tensor MRI in vivo without strain correction. Magn Reson Med 1999; 42:393–403. 11. Dou J, Reese TG, Tseng WY, Wedeen VJ. Cardiac diffusion MRI without motion effects. Magn Reson Med 2002;48:105–114. 12. Gamper U, Boesiger P, Kozerke S. Diffusion imaging of the in vivo heart using spin echoes—considerations on bulk motion sensitivity. Magn Reson Med 2007;57:331–337. 13. Nielles-Vallespin S, Mekkaoui C, Gatehouse P, et al. In vivo diffusion tensor MRI of the human heart: reproducibility of breath-hold and navigator-based approaches. Magn Reson Med 2013;70:454–465. 14. Tseng WY, Dou J, Reese TG, Wedeen VJ. Imaging myocardial fiber disarray and intramural strain hypokinesis in hypertrophic cardiomyopathy with MRI. J Magn Reson Imaging 2006;23:1–8.
1003 15. Wu MT, Tseng WY, Su MY, Liu CP, Chiou KR, Wedeen VJ, Reese TG, Yang CF. Diffusion tensor magnetic resonance imaging mapping the fiber architecture remodeling in human myocardium after infarction: correlation with viability and wall motion. Circulation 2006; 114:1036–1045. 16. Wu MT, Su MY, Huang YL, Chiou KR, Yang P, Pan HB, Reese TG, Wedeen VJ, Tseng WY. Sequential changes of myocardial microstructure in patients postmyocardial infarction by diffusion-tensor cardiac MR: correlation with left ventricular structure and function. Circ Cardiovasc Imaging 2009;2:32–40, 36 p following 40. 17. Kawara T, Derksen R, de Groot JR, Coronel R, Tasseron S, Linnenbank AC, Hauer RN, Kirkels H, Janse MJ, de Bakker JM. Activation delay after premature stimulation in chronically diseased human myocardium relates to the architecture of interstitial fibrosis. Circulation 2001;104:3069–3075. 18. McGill LA, Ismail TF, Nielles-Vallespin S, et al. Reproducibility of in-vivo diffusion tensor cardiovascular magnetic resonance in hypertrophic cardiomyopathy. J Cardiovasc Magn Reson 2012;14:86. 19. Larkman DJ, Hajnal JV, Herlihy AH, Coutts GA, Young IR, Ehnholm G. Use of multicoil arrays for separation of signal from multiple slices simultaneously excited. J Magn Reson Imaging 2001;13:313–317. 20. Breuer FA, Blaimer M, Heidemann RM, Mueller MF, Griswold MA, Jakob PM. Controlled aliasing in parallel imaging results in higher acceleration (CAIPIRINHA) for multi-slice imaging. Magn Reson Med 2005;53:684–691. 21. Setsompop K, Gagoski BA, Polimeni JR, Witzel T, Wedeen VJ, Wald LL. Blipped-controlled aliasing in parallel imaging for simultaneous multislice echo planar imaging with reduced g-factor penalty. Magn Reson Med 2012;67:1210–1224. 22. Pauly J, Le Roux P, Nishimura D, Macovski A. Parameter relations for the Shinnar-Le Roux selective excitation pulse design algorithm [NMR imaging]. IEEE Trans Med Imaging 1991;10:53–65. 23. Robson PM, Grant AK, Madhuranthakam AJ, Lattanzi R, Sodickson DK, McKenzie CA. Comprehensive quantification of signal-to-noise ratio and g-factor for image-based and k-space-based parallel imaging reconstructions. Magn Reson Med 2008;60:895–907. 24. Bruder H, Fischer H, Reinfelder HE, Schmitt F. Image reconstruction for echo planar imaging with nonequidistant k-space sampling. Magn Reson Med 1992;23:311–323. 25. Hu X, Le TH. Artifact reduction in EPI with phase-encoded reference scan. Magn Reson Med 1996;36:166–171. 26. Zahneisen B, Poser BA, Ernst T, Stenger VA. Three-dimensional Fourier encoding of simultaneously excited slices: generalized acquisition and reconstruction framework. Magn Reson Med 2014;71:2071– 2081. 27. Clare S. Iterative Nyquist Ghost Correction for Single and Multi-shot EPI using an Entropy Measure. In Proceedings of the 11th Annual Meeting of ISMRM, Toronto, Ontario, Canada, 2003. p. 1041. 28. Cauley SF, Polimeni JR, Bhat H, Wald LL, Setsompop K. Interslice leakage artifact reduction technique for simultaneous multislice acquisitions. Magn Reson Med 2014;72:93–102. 29. McKenzie CA, Yeh EN, Ohliger MA, Price MD, Sodickson DK. Selfcalibrating parallel imaging with automatic coil sensitivity extraction. Magn Reson Med 2002;47:529–538. 30. Lombaert H, Peyrat JM, Croisille P, Rapacchi S, Fanton L, Cheriet F, Clarysse P, Magnin I, Delingette H, Ayache N. Human atlas of the cardiac fiber architecture: study on a healthy population. IEEE Trans Med Imaging 2012;31:1436–1447. 31. Jones DK, Cercignani M. Twenty-five pitfalls in the analysis of diffusion MRI data. NMR Biomed 2010;23:803–820. 32. Nguyen C, Fan Z, Sharif B, He Y, Dharmakumar R, Berman DS, Li D. In vivo three-dimensional high resolution cardiac diffusion-weighted MRI: a motion compensated diffusion-prepared balanced steady-state free precession approach. Magn Reson Med 2014;72:1257–1267. 33. Gallichan D, Andersson JL, Jenkinson M, Robson MD, Miller KL. Reducing distortions in diffusion-weighted echo planar imaging with a dual-echo blip-reversed sequence. Magn Reson Med 2010;64:382– 390. 34. Robson MD, Anderson AW, Gore JC. Diffusion-weighted multiple shot echo planar imaging of humans without navigation. Magn Reson Med 1997;38:82–88. 35. Chen NK, Wyrwicz AM. Removal of EPI Nyquist ghost artifacts with two-dimensional phase correction. Magn Reson Med 2004;51:1247– 1253.
1004 36. Zur Y. Two-dimensional phase correction method for single and multi-shot echo planar imaging. Magn Reson Med 2011;66:1616– 1626. 37. Zhu K KA, Pauly JM. Autocalibrating caipirinha: reformulating caipirinha as a 3D problem. In Proceedings of the 20th Annual Meeting of ISMRM, Melbourne, Australia, 2012. p. 518. 38. Pruessmann KP, Weiger M, Scheidegger MB, Boesiger P. SENSE: sensitivity encoding for fast MRI. Magn Reson Med 1999;42:952–962. 39. Griswold MA, Jakob PM, Heidemann RM, Nittka M, Jellus V, Wang J, Kiefer B, Haase A. Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magn Reson Med 2002;47:1202–1210. 40. Wong E. Optimized Phase Schedules for minimizing peak RF power in simultaneous multi-slice RF excitation pulses. In Proceedings of
Lau et al.
the 20th Annual Meeting of ISMRM, Melbourne, Australia, 2012. p. 2209. Conolly S, Nishimura D, Macovski A, Glover G. Variable-rate selective excitation. J Magn Reson 1988;78:440–458. Cunningham CH, Wood ML. Method for improved multiband excitation profiles using the Shinnar-Le Roux transform. Magn Reson Med 1999;42:577–584. Cunningham CH, Wright GA, Wood ML. High-order multiband encoding in the heart. Magn Reson Med 2002;48:689–698. Sosnovik DE, Wang R, Dai G, Wang T, Aikawa E, Novikov M, Rosenzweig A, Gilbert RJ, Wedeen VJ. Diffusion spectrum MRI tractography reveals the presence of a complex network of residual myofibers in infarcted myocardium. Circ Cardiovasc Imaging 2009;2:206–212.