Towards standardized postprocessing of global longitudinal strain by feature tracking – OptiStrain CMR-FT study

Global longitudinal strain (GLS) is an important prognostic biomarker in the evaluation of the left ventricular (LV) function [1‐3]. Visual assessment of wall motion abnormalities is fast but dependent on the experience of the observer and short of objective quantification [4, 5]. In echocardiography, GLS derived by automated speckle-tracking has been shown superior in detecting and quantifying subtle impairment of LV systolic function [2], as well as to provide higher predictive value for mortality than LV ejection fraction (LV-EF) in the presence of regional wall motion abnormalities [1], or isolated GLS reduction in the presence of global (diffuse) systolic impairment [3].

Feature tracking (FT) by cardiac magnetic resonance (CMR), or CMR-FT, has conceptual similarities with speckle tracking, in providing quantitative assessment of myocardial deformation [6]. Despite the methodological differences in image acquisition and postprocessing, the similarity extends to the use of routinely acquired cine (steady-state free precession, SSFP) images, avoiding the need for additional dedicated sequences, such as tagging [7, 8]. Numerous studies reported on validation and agreement with other deformation techniques (reviewed elsewhere ([8, 9]), which, in summary, reveal that the measurements derived by CMR-FT are not easily transferrable, nor in scale or precision. One of the reasons for the differences is a number of methodologically different software solutions, pertaining numerous approaches to automatic contour-placement and tracking, as well as underlying algorithms of strain calculation [7]. Moreover, despite high reproducibility GLS in some single centre studies, there remains a remarkable inter-centre difference despite the use of same vendor, reflecting an important source of operator induced, or random variability [10, 11]. We hypothesized that standardization of user postprocessing may reduce the random component of variability. Improved precision of measurements may support transferability of CMR-FT and allow comparability of intervendor and intercentre results, despite the systematic differences, owing to the different image processing algorithms used by vendors. In this study, we undertook a systematic analysis and elimination of potential errors to guide the development of standardized operating procedure for local reads by comparison of two different vendors.

Anonymized datasets were sourced from the prospective longitudinal observational multicentre investigator-led study [12‐14]. Groups of unrelated subjects with either known ischaemic heart disease (IHD) or non-ischaemic dilated cardiomyopathy (NIDCM), were composed to examine the influence of regional wall motion abnormalities (RWMA) and diffuse myocardial impairment on postprocessing of GLS, respectively. The control group consists of subjects with normal blood pressure, low-pretest likelihood of cardiomyopathy, normal LV mass, volumes, and global systolic function, as well as absence of myocardial scar on late gadolinium enhancement (LGE) imaging and no regular medication. Clinical meta-data, including systolic/diastolic blood pressure (BP), body mass index (BMI) were recorded. Exclusion criteria for all subjects were the generally accepted contraindications to CMR (implantable non-MR safe devices, cerebral aneurysm clips, cochlear implants).

Baseline characteristics are displayed in Table 1. The final analysis included 44 subject cases: 15 patients with IHD and regional wall motion abnormalities, 19 patients with NIDCM and 10 healthy controls. Quantitative analysis was not diagnostic in 12 subjects, whereas in 17(39%) subjects 1 ROI had to be discarded. The results of GLS measurements before and after standardization are provided in Table 2 for both vendors. There were significant differences between controls and all patients, as well as patient subgroups, in most CMR measurements, as well as GLS (p < 0.005) (summarised in Table 2); these differences were accentuated through standardization procedures, as shown by significant increase in Cohen D’s and AUCs (Fig. 3, Table 3). The benefit of standardization was also evident through improved CV and ICC agreements between observers and the different vendors. Results of Bland Altman analyses for inter-and intraobserver reproducibility for both vendors prior to and after standardization are displayed in Table 2 (plots Figs. 4, 5 and 6). Limits of agreement and the coefficient of variation were reduced after standardization within each vendor, albeit more strongly for interobserver than intraobserver agreement (p = 0.03 vs <0.001, respectively) and more for MEDIS than CVI42 (Medis: p < 0.000; CVI42: p < 0.028).

The common sources of poor reproducibility included:

Mean GLS values for all subjects were on average 2 percentage points different between the two vendors (p = 0.023), also on the subgroup level (p < 0.05), despite relatively high agreement in overall measurements (r = 0.85, p < 0.001). Although the inter-vendor difference persisted post-standardization, it was smaller compared to initial results, substantiating the fixed element of systematic difference in derived measurements with two vendors.

Results of our study reveal that post-processing approach to GLS matters with respect to the reproducibility of measurements and detection of effective difference in GLS between controls and patients. We demonstrate that standardization of GLS post-processing helps to reduce the random component of variability, introduced by inconsistencies between and within observers, while fixed systematic inter-vendor bias due to vendor related differences in image processing algorithms remained. We further show that the greater precision of measurements affords greater effect size, and as thus, improved discrimination between controls and subgroups of patients, which was not vendor-dependent. Results of our study provide a proof of concept that standardization of GLS measurements is an essential step in ensuring the reliable quantification of myocardial deformation, between different observers, users and across vendors.

Previous studies leading up to the present work have highlighted the differing normal values as well as the results for intra- and interobserver reproducibility, as well as between vendors and centres (summarised in Table 1S from PMID:19789193) [16]. In summary, the reported intraobserver CV ranges between 5.2–12.3%, whereas interobserver CV 10.9–15.4%, and our initial results agree well with these previous reports. However, we have shown that by following the standardization protocol considerably improves reproducibility of measurements, in the study group as a whole, as well as subgroups. By employing a standardization protocol, CMR-FT can become an objective and reproducible method for the quantification of LV deformation. Whereas the burden of contour manipulation may at first appear substantial, we have narrowed this down to a few essential and systematic steps, predictable of failure of tracking, as well as vendor-specific contour placement, which has proactively served to considerable improvement. This information is important as it may guide the necessary optimization steps of CMR-FT softwares, in order to adequately serve the clinical routine. Diversity of normal values is often noted limitation of CMR-FT, yet the range of the thus-far reported mean values is admittedly narrow (19–21.3%, Table 1S), also reproduced by the recent metanalysis by Vo et al., 20.1% [11]. Majority of the previous studies used TomTec based software [6, 7], a product using the same tracking algorithm as the MEDIS software, and our MEDIS derived values in controls reproduce these previous reports. The systematically higher measurements with CVI42 signal a very different image-processing approach; yet the high inter-vendor agreement of measurements suggests that although the softwares may be employing different algorithms, they track similar features of myocardial deformation. The benefits of standardization can further be seen through marked improvement of CV and ICC and Bland Altman plots, reflecting the effect of harmonization for both intra- and inter-observer variability. Reduction of the mean differences and limits of agreement translate into smaller dispersion of the GLS measurements, which is greater for interobserver reproducibility. Applying the standardized steps improved results for both vendors; vendor-specific steps clearly helped to reduce intervendor bias, again communicating a random variability component or, in other words, the many ways in which different observers could potentially use the different softwares. Our findings emphasize the role of clear and documented instructions and their unconditional implementation, in support of multi-user transferability in routine clinical practice.

The detection of early disease relies on precision in the technique, which can control for misclassification from healthy subjects. The patients and groups in our study were selected to be representative of the common clinical scenarios, where employment of GLS is known to be complementary to the assessment of global LV function, e.g. the mid-range LV-EF 30–50% [17]. Compared to controls, both patient groups were older and had significantly but similarly reduced global systolic function. In both groups, GLS values were significantly lower in comparison to controls. Of note, comparative GLS measurements between IHD and DCM group revealed significantly lower GLS in the DCM group (p < 0.001). This is an important observation, which is in part explained by considerably higher LV volumes in DCM group, indicating the presence of global remodelling and consequently, operation at much higher loading. In the IHD group, GLS is reduced considerably, but only owing to severe regional impairment, whereas the preserved myocardium at first compensates with hypertrophic response and not change of loading [1, 18]. Given the rather homogenous presentation of cases within the model disease groups, the AUC for separation of patient groups from healthy controls were excellent before and after standardization, although additional improvement remains notable.

The introduction of CMR-FT was long hailed as a much-needed clinical application that reuses the routine cine acquisitions, while reducing the need for additional imaging that encode the changes with myocardial deformation, such as tagging. The overall viability of this technology appeared to depend on the availably of a quick, sleek and foremost accurate offline postprocessing, which resulted in offspring of several dedicated software products for CMR-FT. Yet the results of CMR-FT analyses vary from vendor to vendor and remain highly observer dependent. Several solutions were proposed, foremost the averaging of results of repeated analyses for increasing intra-vendor reproducibility [11]. Our results reveal that benefit of such solution is likely dubious, the doubling or tripling of analysis time notwithstanding, as the source of high variability primarily arise from the tracking failure of automatically detected (auto-) contours, which cannot be improved by repetition, but manipulation of contour placement on post-processed SSFP images. In our study, this approach turned out to influence most strongly the accuracy of CMR-FT and several reasons underlie this observation. There are many independent variables that cannot be improved by repetitive tracking including image quality, frame rate, slice geometry (e.g. cutting through the papillary muscles), observer and centre experience. Image quality will suffer with poor breath-capacity, mitral annular calcification, pericardial effusion, mis-triggering, low frame rate and imperfect slice positioning, and will result in poor auto-tracking due to difficult endo- and epicardial border definition and misallocation of placed boundary points. Institutional structures mandating standardised approaches and providing adequate training will have high impact on reproducibility and precision.

Standardization of GLS postprocessing helps to reduce the random component of variability, introduced by inconsistencies of and between observers, and to some extent also intervendor variability. There remains fixed systematic inter-vendor bias due to vendor related differences in image processing algorithms. Greater precision of measurements affords an improved effect size, and as thus, discrimination between controls and subgroups of patients, irrespective of the choice of postprocessing software or underlying pathophysiology. Results of our study provide a proof of concept that standardization of GLS measurements is an essential step in ensuring the reliable quantification of myocardial deformation, between different observers, users and across vendors, and for implementation of CMR-FT in clinical routine.