Comparison of magnetic resonance feature tracking for systolic and diastolic strain and strain rate calculation with spatial modulation of magnetization imaging analysis

To compare cardiovascular magnetic resonance‐feature tracking (CMR‐FT) with spatial modulation of magnetization (SPAMM) tagged imaging for the calculation of short and long axis Lagrangian strain measures in systole and diastole.

MYOCARDIAL STRAIN is a sensitive measure of regional and global left ventricular (LV) contractile function (1). Recognizing abnormal strain allows the early detection of subtle LV dysfunction which precedes decreases in ejection fraction in conditions such as dilated cardiomyopathy (DCM) (2,3). Early identification of myocardial dysfunction is important for clinical risk stratification, prompt initiation of treatment, and guides therapeutic decision-making (4). We recently used dynamic tissue-tagging cardiac magnetic resonance (CMR) imaging to identify improvements in longitudinal strain parameters following treatment with spironolactone in patients with earlystage chronic kidney disease (5).
Myocardial tagging by cardiovascular magnetic resonance imaging (MRI) has been widely accepted as the reference standard noninvasive imaging technique for quantifying strain after validation against sonomicrometry in humans (6) and nonhomogenous strain phantoms (7). Most MR tagging techniques create a visible pattern of magnetization saturation in a grid or with parallel lines on the magnitude reconstructed images which are then analyzed, eg, spatial modulation of magnetization (SPAMM), or by extracting information about myocardial tags in k-space, eg, harmonic phase (HARP) (8). All tagging pulse sequences have at least one disadvantage which may include tag fading, low signal-to-noise ratio, long acquisition times requiring prolonged breath-holds, limited availability of validated postprocessing software, and protracted retrospective analysis. Moreover, each tagging technique requires the acquisition of additional sequences to those which are routinely performed, a factor limiting clinical applicability.
Cardiac magnetic resonance feature-tracking (CMR-FT) analysis offers a fast and practical method to calculate strain from routinely acquired steady-state free precession (SSFP) cine images without the need to perform additional tagged sequences (9). In a large study of boys with Duchenne muscular dystrophy, CMR-FT was proposed as an accurate method for measuring strain in a comparison with HARP, although quantification was limited to average circumferential systolic strain in the mid-LV short axis slice (10). Despite recent reports addressing interstudy (11), and inter-and intraobserver variability for CMR-FT (12,13), there has been no attempt to validate this technique for diastolic strain rate (SR) calculation against a reference standard myocardial tagging analysis (14). Furthermore, according to one previous study the CMR-FT framework cannot yet be reliably extended to assess long axis function (15). Therefore, the aim of the current study was to compare CMR-FT with SPAMM tissue tagging for the computation of longitudinal and circumferential systolic and diastolic strain measures in a group of healthy adult controls and patients with DCM.

This study was approved by the West Midlands
Research Ethics Committee and carried out in accordance with the principles of the Declaration of Helsinki. All patients provided informed written consent.

Control Subjects
Normal healthy adults were identified from an ongoing prospective, observational research study examining the effects of living kidney donation on cardiovascular structure and function (REC: 10/H1207/70). The current UK exclusion criteria for living kidney donation include: diabetes mellitus, any history of cardiovascular or pulmonary disease, evidence of hypertensive end organ damage, LV systolic dysfunc-tion, and atrial fibrillation. Prior to nephrectomy, all potential kidney donors who underwent normal baseline cardiac MR studies from March 2011 to June 2012 were included as healthy controls. Control subjects also had normal 12-lead electrocardiography, stress echocardiography, and routine hematology and biochemistry profiles.

DCM Subjects
Patients with DCM were prospectively identified as part of a detailed pathophysiological study assessing the effects of myocardial fibrosis on cardiac mechanics (REC: 12/WM/0157). The diagnosis of idiopathic DCM was made on the basis of left ventricular dilatation and systolic dysfunction in the absence of valvular heart disease, congenital heart disease, and ischemic heart disease sufficient to cause ventricular impairment, following gadolinium contrast-enhanced CMR (16), and normal coronary angiography.

MR Acquisition
MR studies were conducted using a 1.5-T scanner (Magnetom Avanto, Siemens, Erlangen, Germany). The time taken to acquire images for each patient was recorded.

SSFP
Vertical long axis (VLA) and horizontal long axis (HLA) SSFP cine imaging of the left and right ventricles was performed. These images were then used to pilot the LV short axis stack acquired using serial contiguous short axis cines ( (17).

Myocardial Tagging
Three short axis tagged images at the LV base (mitral valve), middle (papillary muscle) and apex, as well as an HLA image were acquired using prospective electrocardiographic gating. A uniform tag grid was created on the images using SPAMM with a tag separation of 8 mm using a segmented k-space fast field echo multishot sequence (TR 3.9 msec, TE 1.7 msec, voxel size 1.99 Â 2.04 Â 8.00 mm 3 , FA 5 , tag grid angle 45 with slice thickness 8 mm, temporal resolution 30 msec, minimum 15 phases per cardiac cycle), as previously described (5). For DCM patients, tagging was performed prior to administration of gadolinium.

Myocardial Strain and Strain Rate Analysis
A timed offline analysis was performed on tagged and SSFP images at identical slice positions by two independent blinded observers (R.J.T and W.E.M.; 3 and 4 years experience, respectively). Both tagged and SSFP slices were reviewed by two independent experts (R.P.S and P.L.; 12 and 15 years experience, respectively) and only high-quality MR studies were included for analysis.

CIMTag Dynamic Tissue-Tagging
Tagged images were analyzed with CIMTag2D software (Cardiac Image Modelling, University of Auckland, Auckland, NZ). An overview of the generalized analysis framework is depicted in Fig. 1. The model geometry was initialized in the first frame (end-diastole) using guide-point modeling (18). Briefly, guide points placed by the user on the endocardial and epicardial border of the LV in the end-diastolic frame were fitted by the model using linear least-squares optimization, resulting in an initial segmentation of the LV with minimal user interaction (Fig. 1a). The reference model was then automatically warped to fit the tissue displacement map given by the SPAMM Figure 1. Overview of the generalized CIMTag2D analysis framework. a: Guide points placed by the user on the endocardial and epicardial border of the LV in the first frame (end-diastole) were fitted by the model using linear least squares optimization, resulting in an initial segmentation of the LV with minimal user interaction and subsequent initialization of the finite element model in the first frame of the SPAMM sequence. b: Visual depiction of the tissue displacement map provided by non-rigid registration image tracking process at end-diastole. c: User corrected texture map overlay as seen after placing guide points in endsystole, thereby interactively warping the model to provide a best fit between image tags and model stripes. Figure 2. Acquisition of circumferential strain with FT software in a normal subject. The endocardial contour of a midventricular SSFP image is drawn manually. The first segment is always set in the anterior septum for consistency.
images. The tissue displacement map was given from a nonrigid registration tracking procedure as previously described (8,19,20). Points were tracked from frame to frame using the incremental displacement maps. A texture map of model stripes was overlaid on the display (Fig. 1b) to provide a graphical representation of the tracking result. The initial tag locations, spacing, and orientation were automatically determined by interrogating the location of the harmonic peaks in the k-space data. The user corrected any tracking errors by placing guide points on the texture map overlay, thereby interactively warping the model to provide a best fit between image tags and model stripes (Fig. 1c). The HLA image sequence was used to determine LV longitudinal strain and SR. Left ventricular circumferential strain and SR were measured from the three short axis views. Whole wall global peak systolic strain and SR values were obtained and subdivided according to wall thickness into respective thirds: subepicardium, midwall, and subendocardium.

CMR-FT
Diogenes CMR-FT software (TomTec Imaging Systems, Munich, Germany), a vector-based analysis tool, was   used to perform subendocardial strain analysis in the corresponding SSFP images (Fig. 2). Endocardial borders were drawn manually in the end-diastolic frame for each image. The CMR-FT software automatically propagated the contour and followed its features (brightness gradient at the tissue-cavity interface, dishomogeneities of the tissue, spatial coherence) throughout the remainder of the cardiac cycle. Global measures of subendocardial longitudinal strain were derived from the HLA view. Global subendocardial circumferential strain parameters were derived from the three short axis views. As for tagging analysis, global diastolic SR signals were recorded during early filling (Fig. 3). CMR-FT segmental strain parameters were not extracted for comparison with tagging because a series of reports including our own substudy of patients with ischemic cardiomyopathy (Supporting Materials, Appendix A) demonstrate high intra-and interobserver variability for regional data (11)(12)(13)21).

Left Ventricular Function, Volumes, Mass, and Wall Thickness
Analysis of LV function, volume and LV mass was performed offline (Argus Software, Siemens, Erlangen, Germany) in accordance with previously validated methodologies (17). Left ventricular mass was indexed to body surface area using the Mosteller formula: BSA (m 2 ) ¼ ͱ((weight (kg)Âheight (cm))/3600). Left ventricular wall thickness was measured in the short axis in each of the six segments of the mid-LV according to the AHA standardized model, and at the identical slice position from which mid LV circumferential strain parameters were derived.

Statistical Analysis
Data are presented as mean 6 standard deviation, median (interquartile range), or frequency (percent-age). Data distribution for continuous variables was assessed using normality plots and the Kolmogorov-Smirnov test. Nonparametric data were logtransformed prior to analysis to achieve normality. Individual strain and SR values obtained using the two different methods were compared using the Bland-Altman technique and Pearson's correlation. A Spearman's rank correlation of the differences with  the means of parameters derived from FT and CIMTag was performed. The mean strain parameters derived from the two techniques were compared using paired t-tests if normally distributed. Continuous variables from controls and DCM patients were compared using independent t-tests. An analysis of variance (ANOVA) with repeated measures with a Greenhouse-Geisser correction was used to compare differences in tagging derived strain parameters across the three myocardial layers. Statistical analysis was performed using SPSS v. 21 (Chicago, IL). A type I error rate <5% (P < 0.05) was considered statistically significant.

Variability of CIMTag2D and CMR-FT Strain Measurements
Interobserver and intraobserver variability assessments were performed using a paired t-test and reported as a bias (mean difference) and standard deviation (SD). The coefficient of variability, defined as the SD of the differences divided by their mean, and intraclass correlation coefficient (ICC) for absolute agreement were also calculated (22).

Study Population
A total of 45 subjects were identified (35 controls, 10 DCM); age was 44 6 14 years. Baseline characteristics are shown in Table 1. Compared with healthy controls, DCM patients were older (58 6 14 vs. 41 6 12 years, P < 0.01) and had higher body weight (87 6 17 vs. 77 6 11 kg, P < 0.01). No other demographic data were significantly different. All participants completed the full imaging protocol. Two controls, however, were excluded from the imaging analysis because of poor tagging imaging quality due to breathing artifacts (n ¼ 1) and electrocardiogram gating issues (n ¼ 1); these participants were therefore only included in volumetric and left ventricular mass assessments. All SSFP images were of excellent image quality and compatible with CMR-FT software.

Myocardial Strain and Strain Rate Analysis
A detailed summary of the statistical comparison of FT versus tagging for all global strain and strain rate parameters is presented in Table 2.

Circumferential Strain (E cc )
Across all subjects, CMR-FT derived peak systolic global circumferential strain measurements at the mid LV slice were not significantly different from those calculated via tagging (Fig. 9). As with the long axis analysis, FTÀE cc values correlated most strongly with CIMTag-E cc values derived from the subendocardium (r ¼ 0.83, P < 0.001; Fig. 10a and 11). A Bland-Altman plot shows close agreement between the two techniques across the entire cohort with neither systemic overestimation nor underestimation and a bias of only 0.2 6 4.0% (À22.7 6 6.2% vs. À22.5 6 6.9%, P ¼ 0.80; Fig. 10b).
Measures of circumferential strain from the mid-LV slice showed better agreement between CMR-FT and CIMTag compared with measures derived from the LV base and apex (Supporting Materials, Appendix B). Tagged imaging also showed a graded increase in  circumferential shortening from the base towards the apex.

Agreement Between CMR-FT and Tagging as a Function of Ventricular Wall Thickness
There was no significant difference between the mid-LV ventricular wall thickness of DCM patients and healthy controls (7.2 6 1.1 mm vs. 7.3 6 1.0 mm, P ¼ 0.7). The limits of agreement (LoA) for calculation of peak systolic circumferential strain in patients with DCM versus healthy controls were comparable (LoA À7.76 to 7.92% vs. À5.85 to 9.05%). For calculation of peak systolic circumferential strain across the overall cohort, there was no association between ventricular wall thickness and the size of the bias relative to its mean value (r ¼ À0.15, P ¼ 0.33).

DISCUSSION
In this study we principally compared a SSFP FTbased algorithm against a reference standard tagged image analysis (SPAMM) for the assessment of Lagrangian strain and SR. This report builds upon recent validation work by including comparisons of 2D longitudinal and circumferential strain and SR during systole and diastole; only systolic strain parameters have previously been validated (10,14,15). We compared the ability of the two techniques to accurately measure diastolic SR during early filling, a sensitive marker of LV diastolic dysfunction which is an important precursor of incident heart failure (23).  On the basis of our analysis performed in the subendocardium and in the circumferential direction, CMR-FT could realistically be extended to the computation of early diastolic strain rate. From a technical viewpoint, the current study benefits from having performed all sequences on a 1.5T MR scanner; a previous validation utilized a mix of acquisitions from 1.5T and 3T MR scanners (10). The validation described herein was also undertaken on corresponding SSFP and tagging images acquired from identical slice positions, a method which has not always been adopted (10). Finally, by performing a timed analysis this report highlights that CMR-FT can generate strain data over four times more rapidly than myocar-dial tagging and this has obvious clinical implications.
Our results contrast with a recent report from Augustine et al. (15) which demonstrated that CMR-FT measurements of longitudinally directed strain showed poor agreement with tagging. There are a number of differences between the studies which could account for this discrepancy. In our analysis, strain parameters derived from CMR-FT versus tagging were compared for all patients entered into the study, which included healthy subjects as well as those with DCM; the validation by Augustine et al. was more modest and included only 20 healthy volunteers out of a total of 145 subjects (13.8%). Moreover, Augustine et al. compared measurements of longitudinal strain derived from tagging across the whole myocardial wall with CMR-FT measures derived from the blood-tissue interface, which effectively select subendocardial deformation information. In order to perform a valid comparison between the two techniques, it is imperative to measure strain from the equivalent myocardial layer, namely, the subendocardium. Indeed, the longitudinal myocardial fibers are principally located in the subendocardium (24). For this reason, in our Bland-Altman analyses we made the a priori decision to compare CMR-FT strain parameters with tagged subendocardial values, which likely explains the improved agreement between the two techniques reported in the current study.
Results of the two methods for calculation of subendocardial strain measures correlate and, with the exception of early diastolic longitudinal SR, Bland-Altman assessments display good agreement, although there remain small differences in the measurements between techniques. Some of the variability in CIMTag strain measurements may relate to the requirement to manually contour both endocardial and epicardial borders, together with the need to make a visual assessment of the tissue displacement map before making corrections to provide a best fit between image tags and model stripes. In contrast, only endocardial contours were constructed in the CMR-FT platform in this study, after which the process was fully automated without an option to modify tracking. These  Bland-Altman plots demonstrating agreement for peak systolic global circumferential strain rate calculation using FT versus tagging. Spearman's rank correlation of the differences and the means was nonsignificant (r ¼ À0.034, P ¼ 0.833). differences may account for CIMTag measurements having poorer interobserver variability compared with CMR-FT as well as tagging postprocessing taking considerably longer. Another disadvantage of SPAMM tagging, which may also contribute to increased variability in strain outputs, is the potential for image quality to be affected by changes in heart rate. Furthermore, measurement of strain throughout the cardiac cycle to include diastolic parameters is not always achievable with 1.5T CMR scanners because of loss of tags caused by T 1 relaxation (25). CMR-FT offers a potentially more robust calculation of diastolic strain data because it relies on standard cine images whose spatial resolution is not adversely affected by T 1 relaxation.
In general, the CMR-FT derived longitudinal strain and SR data were more variable than the circumferential data calculated from the short axis. This may be due to difficulty tracking at the blood-tissue interface and a tendency to track the mitral valve apparatus. In keeping with previous reports, circumferential strain showed the strongest agreement between the two techniques and was particularly robust for the mid-LV slice (10,15). At the LV apex, however, less muscle is available for creating the tag grid upon which the guide point modeling for CIMTag is based, which may have led to increased variability. CMR-FT may lose some of its ability to track the features in each voxel at the tissue-cavity interface at the apex (where there is a potential for cavity obliteration in end-systole) which may also account for increased error. The variability in circumferential strain at the LV base could be explained by a relative increase in the throughplane motion typically observed at this level. The extent to which this degree of variability in CMR-FT measurements might relate to clinical use requires further exploration.
This study confirmed a transmural strain gradient which exists across the myocardial wall (26); both longitudinal and circumferential strain values from tagging increased from subepicardial through to subendocardial layers. This likely explains the better agreement of CMR-FT with tagging values from the subendocardium. It also offers a potential explanation for the reduced sensitivity and slightly higher strain values derived from CMR-FT. The CMR-FT algorithm only tracked deformation at the endocardial border, therefore potentially losing important transmural information captured in tag grids that span the entire thickness of the ventricular wall. Mechanistically, this is in keeping with the pathology of dilated cardiomyopathy, which is characterized by epicardial injury on ex vivo histopathology and demonstrable in vivo with late postgadolinium imaging (27). In this study, ventricular wall thickness was not significantly different between DCM and controls. In DCM, the ventricles are dilated but often with normal ventricular wall thickness, imparting an appearance of thin ventricular walls (28). Nonetheless, in an analysis which included all study subjects, wall thickness did not appear to affect the agreement between strain measures calculated using CMR-FT and myocardial tagging.
There is already evidence that regional assessment by CMR-FT may not be as robust as existing tagging modalities but the potential advantages in ease of acquisition and reduced time for analysis suggest that this method of assessing global deformation could still be of clinical utility. By design, this study did not attempt to validate regional measures because of ongoing concerns over the reproducibility of CMR-FT segmental data (11)(12)(13). CMR-FT applies 2D B-mode tracking such that the motion components parallel to Figure 13. (a) Pearson correlation and (b) Bland-Altman plots demonstrating agreement for early diastolic global circumferential strain rate calculation using FT versus tagging. Spearman's rank correlation of the differences and the means was significant (r ¼ À0.315, P ¼ 0.045), suggesting a proportional error with a downwards trend. tissue boundaries responsible for segmental information are much more affected by noise compared to the perpendicular components from which global strain measures are derived. This relates to the gradients in backscatter amplitude being much higher between tissue and blood than within the myocardium. A number of limitations to this study deserve mention. There were a comparatively low number of subjects included with pathology; however, this cohort enabled a validated assessment of the CMR-FT based technique over a broad range of LV function. Only subendocardial global measures of deformation could be measured on CMR-FT and a comprehensive assessment of SR during late diastolic filling was not possible because of constraints over temporal resolution and loss of tags. Assessment of myocardial deformation by strain and strain rate is sensitive to differences in sampling rate. Even though considerable effort was made to ensure all tagging sequences were acquired with more than 15 phases (and many over 20 phases), there will almost inevitably be a difference in temporal resolution between a prospectively gated sequence and retrospectively gated sequence, specifically when confined by the resting heart rate and breath-holding of the patient. The issue of temporal resolution is such that while the results may correlate, values recorded are unlikely to be the samein a clinical situation, it would be important to compare results using the same method in any single patient. All scans were performed on a 1.5T scanner for consistency, although employing a 3T scanner may have resulted in improved tag persistence. Finally, differences in breath-hold requirements may have contributed to variability between acquisitions.
In conclusion, in a study population with a wide range of LV function, CMR-FT systolic and diastolic global circumferential strain measures and systolic global longitudinal strain measures showed satisfactory agreement and correlated with corresponding values from tagged imaging. The CMR-FT global algorithm has potential for clinical utility, for it can be performed without the need for additional imaging and lengthy postprocessing. Segmental analysis, however, may benefit from further development to improve variability in regional deformation measures. In this regard, CMR-FT cannot yet be used as a robust alternative to myocardial tagging.