A comparison of both DENSE and feature tracking techniques with tagging for the cardiovascular magnetic resonance assessment of myocardial strain

Myocardial strain is increasingly recognized as an important myocardial performance index [1, 2]. It can detect the decline of myocardial function preceding the reduction of left ventricular (LV) ejection fraction [3]. A growing body of literature largely from echocardiography has demonstrated that reduced longitudinal strain is associated with adverse clinical outcome with and without reduced ejection fraction [4]. As a result, the assessment of longitudinal strain is now recommended in the ACC/AHA guideline to be routinely performed in clinical echocardiographic evaluation [5].

Strain can be analyzed using a number of techniques by cardiovascular magnetic resonance (CMR) [6]. Tagged myocardial deformation is the first noninvasive technique that was successfully used in strain evaluation [7‐9]. By adding grids or lines to the imaging plane myocardial deformation can be quantitatively analyzed. Displacement encoding with stimulated echoes (DENSE) is a technique that encodes the tissue displacement directly into the phase of the CMR signal. The development of DENSE has contributed to improved spatial resolution compared with harmonic phase magnetic resonance (HARP) tagging in assessing myocardial deformation, dependent on the acquisition matrix and the size of the filter used [10, 11]. More recently, feature tracking (FT) of the cine images has provided a new dimension for strain analysis by directly tracking the features on routine CMR cine images [12‐14]. While each technique has its merit, the difference of physics in image acquisition and of algorithm in post processing may inevitably contribute to differences in strain values. In this study we sought to evaluate 3 techniques including tagging, DENSE and FT for LV strain evaluation. In addition, we compared three commercially available FT programs. HARP tagging was chosen as the reference because this is the most commonly used reference technique in the literature for strain comparison.

The study protocol was approved by the Institutional Review Board. All participants were recruited prospectively after obtaining written informed consent. Healthy subjects were recruited if they did not have cardiovascular history or risk factors and had normal electrocardiogram (ECG) and transthoracic echocardiogram. Patients with arrhythmia at the time of CMR were excluded from the analysis. Additional exclusion criteria were any contraindication to CMR (claustrophobia, CMR incompatible metals), allergic reaction to gadolinium based contrast agent or impaired renal function (estimated glomerular filtration rate < 45 mL/min/1.7m²). All study participants completed a questionnaire for demographic information and medical history.

There were 95 subjects recruited. All had acceptable image quality for FT analysis. However, the suboptimal quality was seen in tagging (N = 8) and DENSE (N = 3) images. The possible causes included poor breath-hold, body motion and arrhythmia. Therefore, the final analysis cohort consisted of 87 subjects with ten healthy controls and 77 patients with a cardiomyopathy of diverse etiology (Table 1). Mean LV ejection fraction was 50% ranging from 32 to 62%. An infarct pattern LGE was found in 19% (n = 17) and 29% (n = 27) had a non-infarct pattern. Regional LV wall motion abnormalities were present in 48% (N = 52) of the images analyzed including those due to left bundle branch block.

Representative strain curves are shown in Fig. 2, where peak strain was assessed using all 5 techniques from the same patient. There were fewer phases of tagging images due to degradation of tags in diastole. Similarly, there were fewer phases in DENSE as a result of prospectively gated imaging acquisition. The peak Ecc was comparable across the techniques with the highest strain value by Tomtec. The peak Ell was also comparable with the exception of significantly lower values by DENSE. In contrast, the peak Err values differed substantially and lacked agreement amongst the techniques.

The average strain value for the normal controls and for the whole cohort is shown in Table 2. In the whole cohort analysis, DENSE showed the best agreement with tagging in Ecc comparison with ICC 0.778 and mean bias of 0.155% in Bland Altman limits of agreement analysis. This was followed by Circle with ICC 0.652 and mean bias of 1.079% (Table 3). Overall the Ecc difference was relatively small across all techniques (Fig. 3a). When compared with tagging there was a small overestimation using all methods except CIM. In contrast, Err values highly varied from technique to technique (Table 2). The agreement was poor with tagging from all techniques (Table 3) as shown in the plots of Bland Altman limits of agreement (Fig. 3b) due largely to overestimation except Circle. In Ell assessment, FT by Tomtec and Circle had the best agreement with tagging rendering a mean bias less than 1.0% in both comparisons. Ell by DENSE, on the other hand, had a large mean bias of 4.9% due to underestimation when compared to tagging (Fig. 3c).

In the reproducibility analysis, the intra-observer agreement was good for all strain analyses (Table 4). While the inter-observer agreement was good for all Ecc evaluation, the agreement on Ell and Err evaluation was less consistent. We also analyzed coefficient of variations of the intra- and inter-observer agreement (Table 5). Of all the techniques, HARP had the largest data dispersion shown by the highest coefficient of variations in all strain analyses. In contrast, DENSE had small data dispersion with the exception of Err from inter-observer comparison. Overall, FT showed relatively small data dispersion for all comparisons consistently.

In this study we observed good agreement in most of the comparisons for Ecc across all techniques with DENSE and FT with Circle software having the best agreement with tagging. Ecc by Tomtec software appeared to have higher values resulting in modest agreement with tagging. Unlike Ecc, Err agreement was poor across all techniques. In the Ell comparisons, Tomtec and Circle had good agreement with tagging. Ell by DENSE appeared to be consistently lower than Ell by other techniques. Overall, reproducibility was good for all techniques studied especially in the intra-observer agreement despite the differences present between techniques.

Tagged imaging is widely available and frequently used as the reference standard for CMR FT and for speckle tracking in echocardiography [17, 18]. The broad use of tagging should be partly credited to HARP, a commercially available program allowing for fast tagging analysis. DENSE, on the other hand, remains largely as a research tool since the sequence is not yet commercialized [11, 19]. DENSE encodes tissue displacement directly into the phase of the CMR signal [20]. When the two orthogonal displacement encodings are acquired in separate breath-hold the temporal resolution of DENSE can be as high as 17 ms [11]. However, this approach demands perfect breath-hold, exact body position and matched heart rate between the two acquisitions, which can be challenging in patients with advanced morbidity. This is why a single breath-hold approach was used in our study for combined image acquisition in orthogonal directions of displacement encoding trading reduced temporal resolution for consistent image quality. Both DENSE and tagging provide encoding to the myocardium thereby allowing direct tracking of the myocardial deformation. However, the dedicated imaging can potentially be a burden to patients. Arrhythmia, motion and breath-hold difficulty can lead to inadequate image quality. In this clinical cohort, we observed poor image quality in three DENSE and eight tagged images causing considerable data loss. In contrast, FT uses images from routine cine imaging that is typically acquired with shorter acquisition time than DENSE or tagging and therefore is usually better tolerated by patients. Using the arrhythmia rejection protocol balanced steady state free precession cine imaging can produce good image quality even with arrhythmia. That is why we observed 100% usable images. It should be noted that there are diverse algorithms within FT technology which may in part contribute to the differences of strain value as some algorithms track the endocardial border while others track the full myocardial property. For example, Tomtec software tracks the endocardial border thereby yielding higher values than those of other FT products. The spatial resolution of each technique can also vary. For example, the gradient echo for tagged imaging has an in-plane resolution of 1.4 × 1.4 mm. However, it is reduced when analyzed by HARP due to signal filtering. In the case of feature tracking, while the spatial resolution of cine image is maintained, the true spatial resolution is difficult to determine because the exact spacing of the features, likely different from software to software, is usually a proprietary knowledge and unknown to the user.

There are increasing publications that include CMR strain evaluation for a variety of clinical conditions including ischemic heart disease [21, 22], hypertrophic cardiomyopathy [23], amyloidosis [24, 25] and cardiomyopathy due to Duchene’s muscular dystrophy [26]. Limited but emerging data support the prognostic value of longitudinal strain by FT [27]. Normal strain value has also been published for reference purpose. Meanwhile, data from a small study shows that strain value by FT differs between vendors [28] thereby raising concerns for mixed use of different FT programs in clinical evaluation. To date, few studies have reported comprehensive comparisons of strain analysis that include tagging, DENSE, and FT. In this report we demonstrate that there is an important difference in strain value among different techniques and FT software products despite good reproducibility of each product. Therefore, we believe the difference in strain value is likely the result of systemic difference in technology whether at the level of image acquisition or at the level of image post processing. Despite multiple commercially available products the design of the FT algorithm remains unknown and largely proprietary, which makes it difficult to fully understand the difference observed in our study.

The variability of CMR derived strain value demonstrated by us and by others suggests that it is necessary to standardize strain evaluation in order to implement clinical strain assessment and to maintain the same software analysis in comparing serial patient data. Before a standardized approach is established it is essential to create reference value not only based on technique but also by vendor specific product.

Our study has several limitations. We chose one product sequence and the commercially available software for tagging analysis. It is plausible that other sequences and or analysis techniques may yield different strain findings. The same drawback may also be true for DENSE imaging as there are many other options available in imaging and in post processing. Similarly, there are growing numbers of vendors that offer FT analysis. It is beyond the scope of this study to include all of them or their various versions. Nonetheless, our findings provide important insight that the strain value is technique and vendor dependent. The choice of tagging by HARP as the reference standard was an empirical decision based on its broad use, which can be challenged especially in the case of Err and Ell evaluation, given their modest reproducibility. It should also be noted that the tags were largely analyzable during systole and subject to fading during diastole. As a result, the number of cardiac phases available for analysis varied from case to case. While in theory the same effect of signal fading is expected from DENSE we found DENSE consistently provided more analyzable phases than tagging. We also recognize that this is a single center experience. However, we believe that our findings can be generalized because our cohort size is relatively large and we have included diverse types of cardiomyopathy representing a typical clinical cohort for CMR evaluation.

Small but important differences are evident in Ecc and Ell assessment across techniques in addition to large differences seen in Err evaluations. Our findings suggest that it is essential to develop reference standard for each technique or software product for strain evaluation and that analysis software should be maintained consistent for serial patient evaluations. Future research should investigate standardization of strain by CMR.