An in-vivo comparison of stimulated-echo and motion compensated spin-echo sequences for 3 T diffusion tensor cardiovascular magnetic resonance at multiple cardiac phases

There is a growing interest in diffusion tensor cardiovascular magnetic resonance (DT-CMR) due to its unique ability to noninvasively interrogate the myocardial microstructure, and recent publications have demonstrated potential clinical applications [1‐7]. However, the dynamic nature of the heart makes it a challenging target for DT-CMR. Initial in-vivo DT-CMR studies used a stimulated echo acquisition mode (STEAM) based echo-planar imaging (EPI) technique, with diffusion encoding gradients at identical trigger times in two successive cardiac cycles [8, 9]. Diffusion is encoded over one complete cardiac cycle (the diffusion time, Δ) and relatively large diffusion weightings (b-values) can be achieved with relatively small gradient pulses. This allows STEAM DT-CMR to be performed at time points throughout the cardiac cycle [6, 10], which has been used to demonstrate impairment of the rotation of the aggregates of cardiomyocytes known as sheetlets in hypertrophic and dilated cardiomyopathy patients [6, 7, 11]. A recent validation study [6] demonstrated excellent similarity between angular DT-CMR measures obtained in hearts imaged while beating and arrested in-vivo, ex-vivo and histologically in relaxed and contracted states. However, differences in myocardial position and conformation between the diffusion encoding gradients result in changes in image phase and signal intensity. This makes studies difficult during arrhythmia or during free-breathing [9]. The measured diffusion is also affected by the strain history of the myocardium between the encoding gradients [12] although the size of this effect is under debate [1, 6, 13].

To deal with the effects of strain on the diffusion tensor a correction model has been used with input from 3D strain data [12]. Alternatively, using the same model, times in the cardiac cycle where strain has no effect (so called strain “sweet-spots”) on measured diffusion can be identified from strain data [14]. These techniques also require a measurement of 3D strain throughout the cardiac cycle.

Another approach is to use a spin-echo (SE) sequence similar to neurological diffusion tensor techniques [15, 16]. The duration of the diffusion gradients when using SE sequences is usually too long to identify a matching quiescent cardiac phase, but motion compensation may be incorporated into the diffusion encoding gradients [17]. Recently, velocity and acceleration compensated (M2) [18, 19] gradients have been used to perform SE DT-CMR in vivo [20, 21]. A comparison between STEAM and M2-SE techniques [21] described improved SNR efficiency and a more linear transmural distribution of helix angles using M2-SE, as well as differences in fractional anisotropy (FA) and mean diffusivity (MD) between sequences. The short diffusion time of M2-SE techniques means that the technique does not require two cardiac cycles per image, strain effects are minimised and the 50% signal loss inherent to stimulated echoes is avoided. The long diffusion encoding gradient pulses for M2-SE avoid the T1 signal loss during the time between the second and third radiofrequency (RF) pulses (the mixing time) but result in long echo times (TE) and consequential signal loss. However, previously published in-vivo human demonstrations of M2-SE DT-CMR have predominantly used high performance gradient systems at 1.5 T [20, 21].

In this work we implement an M2-SE DT-CMR sequence on a 3 T clinical CMR scanner with standard gradients and directly compare to the STEAM sequence using matched protocols at three cardiac phases. We also consider the contribution of strain to the differences in DT-CMR results between the cardiac phases and sequences.

Fifteen healthy subjects (9 male, median age 24, range 20–36) were recruited with consent according to ethical approval. Images were acquired using a Siemens 3 T scanner (Skyra, Siemens Healthineers, Erlangen, Germany) with standard gradients (43mT⋅m⁻¹ and 180T⋅[m⋅s]⁻¹ per axis) using anterior (18-elements) and posterior (8–12 elements) RF coils. Balanced steady-state free-precession cine imaging was used to identify a mid-ventricular short-axis plane and the end-systolic and end-diastolic periods for subsequent acquisitions.

In this work we have implemented a 1st and 2nd order motion compensated (M2) SE DT-CMR sequence and compared results to those obtained from a monopolar STEAM sequence at 3 points in the cardiac cycle. To the best of our knowledge, this is the first published work to compare these sequences at multiple points in the cardiac cycle. While standard performance gradients have been used for velocity and acceleration compensated (M2) diffusion prepared techniques at 3 T [5, 19] and M2-SE has been demonstrated with a range of gradients strengths at 1.5 T during systolic contraction [20, 21, 42], this is the first study to demonstrate M2-SE with standard performance gradients at 3 T. While we were able to perform both sequences at all cardiac phases in most subjects, there were clear differences in reliability, image quality and the DT-CMR results.

For the healthy cohort in this work, the sequences performed similarly in systole, but STEAM was more reliable and image quality scores and quantitative measures were suggestive of better quality data in the sweet spot and diastole. A substantial number of M2-SE acquisitions were deemed to have failed (7/15 in diastole). In contrast, STEAM was successful in all but 1 acquisition and the visual image quality scores were significantly higher for the STEAM sequence at the sweet-spot (p = 0.002) and diastole (p = 0.001).

Another factor contributing to the differences in results is the differing sensitivity to cardiac motion of the two sequences. The STEAM sequence uses short diffusion encoding gradients, which minimises the detrimental effects of motion during the gradients (although inter-gradient motion is problematic). In contrast, the spin-echo sequence requires a much larger gradient area and the gradients are further extended in time in the M2-SE sequence in order to compensate for velocity and acceleration. While the M2 encoding gradients have a reduced sensitivity to motion, it requires that higher order motion components are minimal while they are run.

The higher failure rate of the M2-SE sequence at the sweet-spot and in diastasis could be attributed to uncompensated cardiac motion during the long diffusion encoding gradients, which is perhaps also consistent with the primary eigenvalues substantially exceeding that of free water in some pixels when using this sequence (Fig. 8). The largest number of studies failed in diastole (7/15) and the diastolic M2-SE image quality score significantly increased with lengthening RR-intervals. The magnitude of cardiac motion is low in diastasis, but the motion trajectories may be complex. Welsh et al. [18] considered the effects of velocity, acceleration and jerk compensation on diffusion encoding gradients and concluded that 2nd order compensation was sufficient in rats. However, the increasing gradient duration as the motion compensation order increases, also makes it increasingly likely that more complex motion will occur during them. Recent work demonstrated increased cardiac motion and eddy current related signal loss in M2-SE sequences than in STEAM sequences [43], although very little of the difference in MD between sequences could be attributed to this. High performance gradient systems, such as those now commercially available with maximum gradient strength of 80mT⋅m⁻¹ at slew rates of 200T⋅ [m⋅s]⁻¹ may partially address the residual motion sensitivity of M2-SE. Stoeck et al. [20] used a maximum gradient amplitude/slew rate of 80mT⋅m⁻¹/100T⋅[m⋅s]⁻¹ to reduce the total M2 diffusion encoding time to 50 ms, versus 62 ms achieved here (TE = 73 ms versus TE = 76 ms here). In a study of the sensitivity of the M2-SE technique to gradient strength, 2/6 systolic acquisitions failed when maximum gradient strengths were only 30mT⋅m⁻¹ [42]. In order to fully understand the effects of cardiac motion on the results and success rates of both sequences in various points throughout the cardiac cycle, future work should consider computational simulations, extending the work of Mekkaoui et al. [44]. One of the difficulties with such simulations is that they will require in-vivo measures of 3D motion trajectories covering the whole cardiac cycle, accurate to at least 3rd order motion (jerk). Myocardial phase velocity mapping, with 3-direction encoding and retrospective cardiac gating [45] may be suited to acquiring such data.

Investigations of the optimum timing of the M2-SE sequence [20, 46] have found that triggering between 15% and 77% of the time from the R-wave to peak-systole (during systolic contraction) resulted in an acceptable level of motion related signal loss (although time points beyond peak-systole were not analysed). In this work, we found the highest success rate when the centre of k-space was timed to peak systole, equivalent to triggering at 75% of peak systole (TE = 75 ms, time to peak systole 300 ms). In comparison to other work, we found lower success rates at the systolic sweet-spot, which we defined as a constant 150 ms from the R-wave based on previous work considering the predicted effects of strain on STEAM data [27]. However, the three M2-SE acquisitions which failed at the sweet-spot correspond to subjects with the maximum (100 bpm) and minimum (50 and 55 bpm) mean heart rates, suggesting that a fixed trigger time may not be best suited for M2-SE imaging.

A further difficulty with acquiring M2-SE data in diastasis is that the longer acquisition window frequently caused the scanner to miss the next R-wave. Based on DICOM timestamp information we estimate the proportion of missed triggers as 7% for diastolic M2-SE and 0.1–0.6% in other acquisitions. These missed triggers extend breath-hold durations and perturb the longitudinal magnetisation’s steady-state, which is likely to result in errors in the calculated diffusion tensor. In earlier pilot work [24], we achieved a higher success rate for M2-SE imaging in diastasis (15/20 acquisitions successful) by triggering M2-SE to alternate R-waves (TR = 2RR-intervals). Principal component analysis – temporal maximum intensity projection (PCA-TMIP) post-processing technique [47] and related methods [48] could potentially be combined with M2-SE sequences to palliate its motion sensitivity. Alternatively, pharmacological methods of increasing the length of diastasis may be useful [49].

The strain history of the imaged tissue during the diffusion time affects the measured diffusion [12] and the long diffusion time of STEAM means it is more affected than M2-SE. We found significant correlations between peak radial strain and the difference between systolic and diastolic FA and E2A using STEAM (p < 0.05). We did not perform a full strain correction [10, 50]. Currently the only model [12] describing the modification of the diffusion tensor in response to strain assumes that the myocardium can be modelled as a homogeneous elastic medium. While it is clear that diffusion measured with a monopolar STEAM sequence over 2 cardiac cycles should be affected by cyclical strain, the complex multi-scale structure of the heart means that the exact effects of strain are more complex than those described by the existing model. We have recently highlighted issues with the results provided by the model by comparing ex-vivo and in-vivo STEAM DT-CMR in pigs [13]. The correlation between E2A and peak radial strain could be due to the strain effect, but E2A mobility (systole-diastole) obtained using the strain insensitive M2-SE technique also correlates with peak radial strain (p = 0.01). Furthermore, it is logical to expect a correlation between E2A mobility and radial strain even in the absence of the effects of strain on the DT-CMR acquisition. The change in sheetlet angle represents the primary mechanism behind wall thickening [1] and we have recently validated these changes in a pre-clinical model [6]. However, we do expect some contribution of strain to the E2A mobility measured with STEAM and this may be reflected in the correlation (p < 0.01) of differences in E2A between sequences with strain in both diastole and systole. Recent work comparing arrested and beating heart measures of E2A in relaxed and contracted states, estimated the maximum component of E2A mobility attributable to strain [6] as ~17%. The median E2A mobility measured in this study using STEAM of 40^o compared to 11^o measured with M2-SE (Supporting Table S1) suggests that strain is not the only contributor to this difference. A component of the difference between parameters other than E2A may also be the result of strain effects. While an ex-vivo study has demonstrated changes in eigenvalues and FA between contracted and relaxed states, MD was maintained [51]. As a result, we can estimate the maximum possible contribution of strain to MD measured with STEAM as the difference between the sweet-spot and the diastolic or systolic values. These differences of 0.05 × 10⁻³mm²⋅s⁻¹ and 0.06 × 10⁻³mm²⋅s⁻¹ (calculated from values in Supporting Table S1) at systole and diastole respectively, suggest that strain is not a major contributor to the much larger difference in MD observed between the two sequences. Indeed, even at the sweet-spot, where the existing model predicts minimal strain effects for STEAM data, there are large differences between the results from the two sequences.

The difference in diffusion time between these two sequences is expected to result in changes in the DT-CMR results [21, 52]. During the diffusion time of the STEAM sequence (~1 cardiac cycle, ~1000 ms), free water molecules would diffuse a root mean square distance of around 75μm (D = 3 × 10⁻³mm²⋅s⁻¹ at 37 °C (41)), while the equivalent distance for the M2-SE sequence (diffusion time ~20 ms) is only around 10μm, which is less than typical cardiomyocyte dimensions (see Fig. 11). This limits the number of diffusing water molecules interacting with cellular structures and, while meaningful theoretical predictions of the strength of this effect are impossible without complex numerical simulations, previous work in ex-vivo tissue can provide some insights. Kim et al. [52] studied the diffusion time dependence in fresh samples of calf heart. Increasing the diffusion time from 33 ms to 412 ms resulted in an increase of FA from 0.40 ± 0.07 to 0.59 ± 0.06 and a reduction in MD from 1.59 ± 0.12 × 10⁻³mm²⋅s⁻¹ to 1.33 ± 0.08 × 10⁻³mm²⋅s⁻¹. Despite the obvious limitations of comparing with ex-vivo tissue, these results suggest that the changes we observe here between sequences are predominantly a consequence of diffusion time effects. A recent review by Kiselev [53] suggests that probing cellular muscle structure is only possible with the longer diffusion times available with stimulated echoes and we might expect reductions in the uncertainty of angular DT-CMR measures with such sequences.

A further consequence of using a full cardiac cycle as the diffusion time in the STEAM sequence is that the b-value is modulated by variations in the RR-interval. In order to account for this b-value variation we corrected the b-values used in the tensor reconstruction based on the time stamps of the DICOM images. However, the time dependence of the measured diffusion is not accounted for in this correction, but should be small for typical variations in the RR-interval (the mean of the intra-subject standard deviation of the RR-interval was 5 ms) [52].

Using phantom data we demonstrate a 1.75 factor increase in SNR per image using the M2-SE sequence over STEAM. Due to the longer T1 and shorter T2 of myocardial tissue, the theoretical SNR ratio is lower in vivo than in the phantom. Based on Eq. 1 the theoretical SNR ratio in vivo is SNR_SE/SNR_STEAM = 1.37, assuming myocardial T2 = 45 ± 4 ms (T2-prepared method [54], measured in 20 subjects in earlier work [24]), T1 = 1327 ± 59 ms (modified Look Locker 5(3)3 [39, 40], in the same 20 subjects) and an RR interval of 970 ms. In the “b0” data we measured a median [interquartile range] in-vivo SNR ratio of SNR_SE/SNR_STEAM 0.78 [0.38], 0.68 [0.23] and 0.59 [0.38] at systole, sweet-spot and diastole. These ratios are lower than the results of von Deuster et al. [21], who measured SNR efficiency and less than theoretically predicted. SNR per image measured using the multiple repetitions method is affected not only by true image noise but also by variations in signal intensity due to motion related signal loss, imperfect registration and residual blood signal. In this work, EPI readouts, receive gain, receive coils, reconstruction schemes and image scaling were matched between sequences and it is therefore reasonable to assume that image noise is similar (and we have confirmed this is so in phantom experiments, not included here). Therefore, we use the ratio between mean signal intensities in the two sequences as a surrogate for SNR ratio with fewer confounding effects than the multiple repetitions method. The median [interquartile range] signal ratios (M2-SE/STEAM) were 1.01[0.21], 0.96[0.12] and 1.19[0.34] at systole, sweet-spot and diastole; suggesting that a substantial fraction of the reduction in M2-SE SNR per image measured using the multiple repetitions method relative to theory is due to the confounding factors discussed above. The remaining difference between mean signal ratio and the theoretical SNR ratio could be a consequence of sequence differences in the in-plane excitation profiles. Small M2 gradients are used as crushers around the 180^o pulse in the “b0” acquisition and higher order motion during these gradients may also contribute to the reduction in measured mean signal ratio relative to theory. In practice the SNR per unit acquisition time (as calculated by von Deuster et al. (21)) of the M2-SE sequence is boosted beyond that measured here by the additional averages that can be acquired due to its shorter TR compared to STEAM. However, STEAM allows much higher b-values to be used with little penalty in TE [33] and therefore a potential increase in diffusion contrast to noise relative to the M2-SE sequence.

Our more comprehensive findings differ from the smaller studies of both von Deuster et al. [21] and Stoeck et al. [20], who did not report any unsuccessful DT-CMR acquisitions. At the sweet-spot we found similar mean FA using both sequences and MD using STEAM, but higher MD values using M2-SE (1.69 ± 0.15 × 10⁻³mm²⋅s⁻¹ vs. 1.43 ± 0.06 × 10⁻³mm²⋅s⁻¹ (21)). There were a number of key differences between our work and these previous studies. While we imaged at 3 T with a widely available gradient strength Stoeck et al. [20] and von Deuster et al. [21] used a high strength gradient system (80mT⋅m⁻¹, 100 T⋅[m⋅s]⁻¹) at 1.5 T. We used parallel imaging to shorten the EPI readout and used optimal diffusion encoding directions for gradient performance to achieve a similar TE for M2-SE (76 ms vs. 73 ms (20) vs. 70 ms [21]) and a shorter STEAM TE (25 ms vs. 31 ms(21)). We imaged at three cardiac phases rather than only at the systolic sweet spot. We acquired data during multiple breath holds, as did von Deuster et al. [21], which we have found to be a more reliable method of acquiring STEAM data, although M2-SE acquisitions may be well suited to navigator gated free-breathing acquisitions [20, 21]. There is also the potential confounding issue of experience. While our group has ~6 years experience acquiring and processing STEAM DT-CMR data and one experienced operator (ADS: 11 years CMR experience, 5 years DT-CMR experience, >40 subjects imaged with M2-SE) imaged all 15 subjects, M2-SE is a relatively new and evolving technique [55]. Therefore, we biased our study to the advantage of M2-SE acquisitions by selecting trigger times for all systolic and diastolic acquisitions based on the optimal timing for M2-SE acquisitions.

Using a clinical 3 T scanner with standard gradients, both DT-CMR sequences perform equally well at the systolic pause and are successful in the majority of subjects. STEAM can also be used reliably in diastole, but 2nd order motion compensation of M2-SE did not appear to adequately cope with the low magnitude but complex trajectory of the motion at this cardiac phase. There were clear differences between the results obtained using each of these methods and only a few of these differences correlated with measures of strain. These differences are predominantly the result of the 2 orders of magnitude difference in diffusion time between these two techniques. While M2-SE may be more suited to free-breathing studies or imaging subjects with arrhythmia, STEAM is a more robust acquisition when diastolic imaging is required. It is, therefore, vital to understand the differences between these two methods and obtain normal values for the sequence of choice when planning future patient studies.