Background
Cardiac mechanics, such as strain, torsion, and dyssynchrony, are important indicators of cardiac function and independent predictors of serious outcomes, even when accounting for traditional measures such as ejection fraction [
1,
2]. Several advanced cardiovascular magnetic resonance (CMR) sequences have been developed to assess cardiac mechanics including tagging [
3,
4], displacement encoding with stimulated echoes (DENSE) [
5‐
7], strain encoding (SENC) [
8], and tissue phase mapping (TPM) [
9]. While these techniques can provide reference standard measurements of myocardial motion and deformation, their use is often clinically impractical. Furthermore, because they are specialized non-clinical techniques, there are few large datasets available that could be used to guide the clinical use of these techniques. As such, there has been growing interest in the use of feature tracking software to approximate the mechanics produced by reference standard techniques [
10‐
12]. While feature tracking is simple to use and requires only standard anatomical cine sequences that are widely available, it is important to assess how well measures of cardiac mechanics such as left ventricular (LV) strain, torsion, and dyssynchrony derived from feature tracking agree with those derived from reference standard techniques.
While results from feature tracking have been compared to those from tissue tagging [
10,
13‐
16] and TPM [
17], many of these studies have been limited in scope. The largest study [
10], with 191 patients with Duchenne’s Muscular Dystrophy and 42 healthy controls, surveyed only mid-ventricular short-axis images, while other studies have had limited sample sizes (
n = 18 [
16],
n = 20 [
13]). Such studies have suggested that feature tracking may have poor reproducibility and poor agreement with reference standard techniques for some measures of cardiac mechanics [
12,
13] due to the following potential limitations: 1) feature tracking derives displacement fields by propagating myocardial borders from frame to frame, which relies on quantification of local changes in signal intensity and therefore is likely to fail when quantifying motion parallel to or inside the myocardium where there are no features; 2) feature tracking only captures in-plane displacement, and through-plane motion of the myocardium violates the assumptions required to track pixel data.
DENSE, an advanced CMR technique, encodes a component of tissue displacement into the phase of the CMR image [
5]. The reconstructed phase image measures displacements directly at the pixel-level with higher spatial resolution (2–3 mm) than myocardial tagging, which is limited by the number of tag lines that can be reliably tracked in the image. Therefore, motion within the myocardium can be accurately captured in all directions. Indeed, data from a deforming phantom demonstrated that DENSE has equal or better performance than tagged CMR, depending on the measured cardiac mechanic [
18]. Several advancements in DENSE acquisition since its introduction, such as complementary spatial modulation of magnetization (CSPAMM) artifact suppression [
6] and efficient spiral readouts [
7], make it an ideal, highly reproducible and validated technique for reference standard measurements of myocardial motion and deformation used by numerous previous studies [
18‐
24]. However, none of the feature tracking validation studies have been performed with DENSE. Indeed, a recent study [
25] included data from both DENSE and feature tracking, but no direct comparisons were made.
Additionally, a literature review including 62 CMR feature tracking studies found that slice-wise strains (i.e. the average strain over an entire image, such as basal, mid-ventricular, or apical short-axis slices) are the most commonly reported measures derived from feature tracking (Table
1 and Additional file
1). In total, 39 studies (63%) reported either circumferential, longitudinal, or radial slice-wise strain, and 13 studies (21%) reported only those strains. However, slice-wise strains, which are reflective of the change in length of an entire contour between just two frames, end-diastole (ED) and end-systole (ES), should not require segmental motion tracking [
26]. This suggests that the most commonly reported results from feature tracking could be easily assessed without performing tracking, by simply using the ED and ES contours which are already generated during most clinical CMR scans.
Table 1
Reported mechanics from 62 CMR feature tracking studies
Mechanics |
Circumferential Strain – slice-wise | 36 |
Longitudinal Strain – slice-wise | 28 |
Radial Strain – slice-wise | 21 |
Circumferential Strain – segmental | 18 |
Longitudinal Strain – segmental | 12 |
Radial Strain – segmental | 12 |
Systolic Strain Rate | 5 |
Diastolic Strain Rate | 6 |
Torsion | 8 |
Torsion Rate | 5 |
Synchrony | 6 |
Atrial Strain | 8 |
Right Ventricular Strain - any | 13 |
Right Ventricular Strain - segmental | 7 |
Othera | 3 |
We hypothesized that LV strains, torsion, and dyssynchrony estimated from feature tracking would not agree well with those measured by DENSE as a reference standard. We also hypothesized that slice-wise strains from measuring the change in length of entire contours between the ED and ES frames (“contour-based” strains) would agree well with strains reported by feature tracking.
Discussion
This study evaluated the utility of commercially available feature tracking software for quantifying measures of cardiac mechanics, including LV strains, torsion and dyssynchrony. Our primary findings included: 1) the only truly acceptable agreement between feature tracking and the reference standard (DENSE) was observed for mid-ventricular Ecc, 2) feature tracking significantly overestimated the magnitude of Ecc in basal and apical images, 3) feature tracking overestimated Ell in four-chamber images, 4) feature tracking significantly overestimated dyssynchrony and under- or over- estimated torsion depending on the vendor with unacceptable CoVs, and 5) slice-wise strains from the change in length of entire contours (contour-based strains) had excellent agreement with slice-wise strains reported by feature tracking.
Slice-wise strains from feature tracking and contour-based strains
Feature tracking has emerged as a simple and convenient tool for estimating cardiac mechanics from standard CMR imaging. However, we found that the most commonly-reported mechanics from feature tracking (slice-wise Ecc and Ell) can be reproduced by contour-based strains. Such agreement between feature tracking and contour-based strains has been previously reported along with the suggestion that manual border delineation could be a low-cost alternative to purchasing feature tracking software [
26]. Because of the excellent agreement between feature tracking and contour-based strain, regional tracking capabilities and the cost of the feature tracking software are not required to assess these metrics. Many of the insights from previous feature tracking studies could have been easily produced without the software by the manual delineation of borders at two time points, ED and ES, which is already routine for most clinical examinations in which LV volumes are reported.
However, the use of feature tracking or manual delineation to assess slice-wise strains is not beyond reproach. Only mid-ventricular Ecc had good agreement between feature tracking and DENSE with 95% limits of agreement of ±6.3% and a CoV of 10.9%. Previous studies that assessed the agreement between feature tracking and myocardial tagging have shown 95% limits ranging from ±3.3% [
10] to ±9.1% [
16], with several other studies in between [
13,
14,
31]. However, for Ecc in basal and apical images, we found significant biases and larger CoVs, which indicates that feature tracking and DENSE do not agree as well in those regions. In particular, apical Ecc had the largest 95% limits of agreement (±10.8%), which is consistent with a previous study that also observed that the apical region had the largest 95% limits (±12.8% at 1.5 T and ±9.2% at 3 T) [
16]. The largest bias (−4.0%) was observed in basal Ecc. This bias was large enough to disrupt the physiologic gradient in Ecc from base to apex that was observed in the DENSE results and has been documented extensively [
21,
34‐
36]. These inconsistencies between feature tracking and DENSE at the basal and apical levels are likely due to both through-plane motion, which is most prominent at the base and invalidates the fundamental assumption that a segment of tissue can be observed and tracked through the entire cardiac cycle in a single 2D image plane, and the difficulty in tracking the true endocardial contour, which may be more prominent at the apex due to papillary musculature and trabeculations. Inter-test variability in both techniques, while larger in feature tracking [
12,
20], also contributes to imperfect agreement between them. Among the slice-wise strains quantified in this study, Ell had the highest CoV (19.3%) along with high 95% limits of agreement compared to DENSE (±9.3%). This is consistent with a previous comparison between feature tracking and myocardial tagging which found 95% limits of agreement to be ±9.5% [
13].
To reiterate, for slice-wise strains in general, manual contour delineation at just ED and ES can replace results from feature tracking. However, neither agree strongly with a reference standard technique like DENSE with the exception of mid-ventricular Ecc.
Torsion, Dyssynchrony, and other mechanics
Less commonly-reported measures of mechanics from feature tracking include segmental strains, strain rates, torsion, and dyssynchrony. However, these are precisely the measurements for which accurate feature tracking would be most useful since none of these can be quantified by the manual delineation of contours at only ED and ES. Unfortunately, feature tracking has limited success in accurately and reproducibly quantifying these mechanics.
Previous studies of segmental strains and strain rates from feature tracking have demonstrated poor reproducibility and poor agreement with reference standards [
12,
16]. Similar results were observed for torsion in the present study where torsion from feature tracking (TomTec) significantly underestimated DENSE by 1.4 deg/cm on average, while torsion from Circle overestimated DENSE by 1.5 deg./cm. This large bias is consistent with the literature as the torsion found by DENSE (3.5 ± 0.9 deg/cm) is similar to previous results from DENSE (3.1 to 3.9 deg./cm) [
21] and myocardial tagging (3.4 to 3.7 deg/cm) [
37] while the torsion result from TomTec (2.1 ± 1.2 deg/cm) is similar to previous feature tracking studies (2.3 ± 0.8 deg./cm) [
33]. Furthermore, the CoV and 95% limits of agreement for comparing DENSE and feature tracking were high (41.1% and ± 2.4 deg/cm, respectively). Another previous study also found poor agreement and correlation between torsion derived from feature tracking and myocardial tagging as well as poor reproducibility from feature tracking [
15]. The poor results from feature tracking using bSSFP images are likely due to the difficulty in tracking myocardial motion in the circumferential direction. While a strong gradient between the blood pool and the myocardium exists for accurately tracking the location of the endocardial contour in bSSFP imaging, the gradients in the orthogonal direction, which are necessary for tracking twist along that contour, are much weaker. Therefore, it is nearly impossible to track motion parallel to the myocardial wall unless additional features (such as papillary muscles) are present for tracking.
In the present study, dyssynchrony was significantly overestimated compared to DENSE (42 ± 22 vs 16 ± 20 ms) while demonstrating large 95% limits of agreement (±56 ms) and CoV (76.3%). A previous study has also demonstrated poor reproducibility for the quantification of dyssynchrony from segmental Ecc (CoV: 37.5%) [
38]. The poor agreement between feature tracking and DENSE is largely due to the need for accurate segmental strains within the dyssynchrony calculation. Poor reproducibility of segmental strains, which has been demonstrated for feature tracking [
12], likely resulted in both erroneously high measured dyssynchrony within volunteers (i.e. a bias) and greater variability in general.
Limitations
While this study evaluated the agreement between measurements derived from feature tracking and those same measures derived from a reference standard DENSE sequence, we could not evaluate the prognostic utility of the measures. While we observed imperfect agreement between the techniques, it is still possible that feature tracking (or manual contour delineation) produces useful results. However, careful consideration is required before generalizing results from reference standard techniques to feature tracking. There may be cases where only a reference standard technique is sufficient (e.g., identifying a gradient in Ecc from base to apex). Measures of radial strain were not included in this study due to well-known poor reproducibility [
12,
13]. In addition, the current study did not evaluate all patient populations. Different populations will likely show different levels of agreement. In particular, populations with poor function and reduced through-plane motion would be expected to have better agreement between feature tracking and reference standard techniques. However, since changes in strains may precede changes in other functional measures, quantification of cardiac strains is likely most important in populations with healthy or nearly healthy function (e.g. pediatric obesity [
23]). While it is common to assess the agreement between feature tracking and reference standard techniques with only healthy participants [
13], we note that there was a preponderance of healthy subjects in the assessment of Ell compared to other strains in this study.
Conclusion
Compared to DENSE as a reference standard, feature tracking was inaccurate for quantification of apical and basal LV Ecc, Ell, torsion, and dyssynchrony. Feature tracking was only accurate for quantification of mid LV Ecc. Moreover, feature tracking is unnecessary for quantification of global peak strains, since simplified processing of only ED and ES contours yields very similar results to those derived from feature tracking. Current commercial feature tracking technology therefore has limited utility for quantification of cardiac mechanics.