Reproducibility and its confounders of CMR feature tracking myocardial strain analysis in patients with suspected myocarditis

A total of 125 patients were randomly selected from a dual-center database of 941 patients referred for a contrast-enhanced CMR exam with the primary suspicion of myocarditis. Exclusion criteria were documented refusal of consent, cardiovascular surgery or intervention within 90 days prior to CMR, and any prior evidence or CMR characteristics for coronary artery disease, and other cardiovascular comorbidities described previously [13, 14]. The study protocol was reviewed and approved by the local Institutional Review Boards (NCT03470571) at the Brigham and Women’s Hospital, Boston, and the Inselspital, Bern University Hospital, Bern (NCT04774549).

Images were obtained with a 3.0-T or a 1.5-T system (Magnetom Trio, Verio or Aera, Siemens Healthineers, GE Signa Series, GE Healthcare) [13]. Ciné images were based on clinical routine parameters used at the time of the exam, and acquisitions used in this analysis were retrospectively gated with a minimum of 25 phases (detailed in Supplemental Table 1). These cinés were acquired covering the left and right ventricle in a short-axis (SAX) stack without gap, from which biventricular circumferential and radial CMR-FT parameters were acquired along with ventricular volumes and mass. LV longitudinal strain was measured from three LV-centered long-axis (LAX) views (2-, 3-, and 4-chamber), with RV longitudinal and radial strain measured on the free-wall of the 4-chamber view only. After the placement of endocardial and epicardial contours excluding the papillaries and slices with outflow or inflow planes, deformation measurements were calculated by the feature tracking algorithms (Fig. 1, Circle Cardiovascular Imaging, version 5.9). Contours were adjusted by the reader if acquired. Two-dimensional (2D) data was acquired from the SAX and LAX planes individually, while 3D results were obtained as a result of a 3D construction by the software combining the different 2D planes. The key focus in the main text is on peak strain and early diastolic strain rate for the circumferential and longitudinal orientation. Detailed analyses for other FT parameters are provided in the supplemental information. CMR-FT datasets or individual measurement types were excluded for the following conditions: if a ciné was not acquired for the slice plane, poor angle plane, ventricular wall not fully visible, and if tracking was inadequate because of artifacts, extremely poor image quality as a result of arrhythmia, and or other tracking issues. Furthermore, the readers assessed if images were impacted by gating issues, likely caused by arrhythmia at the time of exam and for 3D models, readers categorized if the fit of 2D planes was acceptable. To assess variability of CMR-FT measurements, datasets were recoded, and the same reader re-analyzed all 125 patients blinded for intrareader assessments at least 7 days later, while a second blinded reader performed a third assessment for inter-reader analysis. CMR level III–certified readers performed for the clinical assessments of suspected myocarditis and quantified tissue characterization. Feature tracking analysis was performed by junior readers under the supervision of CMR level III analyzers.

Examinations also included the acquisition of late gadolinium enhancement (LGE) images in a short-axis stack which was quantified both by visual presence, and by the extent (% of myocardium) measured using a full-width half maximum threshold [15]. Edema was analyzed by the signal intensity ratio of the myocardium versus the major or minor pectoral skeletal muscle on T2-weighted images [13].

As the primary index of agreement, the mean absolute difference (|Δ|) of strain measures between two reads was calculated. This was performed for both inter-reader and intra-reader assessments, and also expressed as a relative difference, in comparison to the mean measurement. A repeated measures ANOVA compared 2D LV circumferential, LV longitudinal, and RV circumferential measures. A paired t-test compared 3D LV circumferential and longitudinal measures. Both patient- and imaging-related variables were investigated as potential confounders on reader reliability and agreement for CMR-FT analysis. To determine the impact of factors on the absolute disagreement (|Δ|) of CMR-FT analysis, univariable linear regression analysis was performed and factors and variables with p < 0.10 were then forwarded into a multivariable regression model to determine the strongest confounders. The forwarded variables were assessed for collinearity and removed from the multivariable model based on known relationships or by comparing models using Akaike information criterion. This was performed individually for each CMR-FT parameter presented in the manuscript.

Patient exams were performed between the years 2002–2019 (Table 1, group demographics are displayed in Supplemental Table 2). One-third of the patients had reduced left ventricular (34%) and right ventricular (37%) ejection fraction with 52% presenting with LGE. A 12-lead electrocardiogram performed prior to the CMR showed 66% had abnormal findings[14].

2D longitudinal measurements were acquired in the highest proportion of patients, with up to 124/125 patients analyzed (99%). Analysis of SAX slices was performed in 122 (98%) of patients. LV-3D measurements were performed in 120 (96%) patients, while inclusion was lowest for RV analysis. A maximum of 106 (85%) patients were analyzed for RV FT (reasons for exclusion detailed in Supplemental Fig. 1).

Multivariable analysis showed that for the majority of 2D and 3D LV peak strain and diastolic measures, an increase in slices independently improved agreement (Tables 3 and 4). Agreement was better for ventricles with higher end-diastolic volumes and reduced ventricular function shown by attenuated feature tracking measures and ejection fractions. For LV 3D early diastolic strain rate, the presence of pericardial effusion compromised agreement both circumferentially and longitudinally. Sixteen (13%) of the images were deemed by the readers to be impacted by poor gating, likely due to arrhythmia, and this was the only significant confounder for the agreement of RV GLS in the multivariable model. For the RV diastolic parameters, independent confounding factors differed between inter-reader and intra-reader models, with agreement better at 1.5 T versus 3 T, at lower heart rates and at reduced early diastolic strain rates.

For both 2D and 3D systolic and diastolic measures, circumferential measures showed the highest reliability and agreement in comparison to longitudinal [6, 18]. An explanation for a higher reproducibility with GCS in CMR studies might be related to our observation of improved agreement in SAX stacks with numerous slices. A full SAX stack acquired in CMR will often incorporate 3–4 fold more slices than a typical 3-planar LAX acquisition. Thus, in a clinical setting where arrhythmia, poor gating, and rapid heart rate may be present, the utilization of more slices can allow any errors to be averaged out better whereas the fewer slices used for LAX analysis may be more exposed to these issues. Consequently, it would be advisable to perform analysis on the maximum data as possible and artificial intelligence techniques now allow this to be performed without a significant increase in workload. It is important to note in our study the reliability for longitudinal markers was still excellent, and GLS in particular is still often most commonly used for diagnostic and prognostic reasons [3, 5]. As circumferential and longitudinal fibers compose different regions of the myocardium, assessment of these orientations may be used in the future to investigate different disease processes.

Impaired diastolic function can lead to the onset of symptoms and cardiovascular events [19], especially in cohorts where systolic dysfunction is not overtly impaired. We also recently showed that diastolic strain rate was a significant prognostic marker as well, especially in patients with an LVEF > 40% [3]. While awareness for the clinical significance of this marker is rising, little is known about its reproducibility with CMR-FT. In 20 patients with a myocardial infarction, Nazir et al. showed 2D early diastolic strain rate measurements were reproducible by CMR-FT at both 1.5 and 3 T, comparable to MRI tagging [7]. Here we observed for 2D and 3D left ventricular measurements, mostly good to excellent agreement for early diastolic strain rate. When looking at patient factors, 2D agreement was improved in patients with LGE enhancement and 3D agreement was improved in patients with poor diastolic dysfunction. However, in the presence of pericardial effusion or a poor alignment of the 3D construction, diastolic strain rates were less reproducible.

With advancing technology, 3D imaging is available with the goal of providing better coverage of the heart. In an independent acute myocarditis cohort, Gatti et al. reported that peak strain and systolic and diastolic strain rates were impaired in acute myocarditis for both 2D and 3D CMR-FT analysis [20]. We also observed a moderate to good ICC for many 3D parameters in our patient cohort; however, these were slightly poorer in comparison to the 2D measurements with a larger relative disagreement. This finding is contradictory to previous findings by Liu et al. in 100 healthy controls, where 3D FT-CMR yielded more reproducible analysis [9]. Unlike true 3D block acquisitions that are acquired within a single measurement, the 3D models used for CMR-FT analysis are only constructions of multiple 2D short-axis and 2D long-axis planes. Thus, the 3D construction is dependent on accurate fitting of multiple images acquired in different acquisitions. Ciné acquisitions were spread out across the exam, extending to 59 min in some cases. Especially for the 3D analysis, this is not ideal as it is likely there is minor patient movement over the course of an hour. Furthermore, patients are more likely to have issues with maintaining breath-holds in the same end-expiratory position, and arrhythmia and poor gating can impact reconstruction as well. Consequently, we observed that a mismatched fit did compromise 3D GCS and GLS inter-reader agreement. 3D assessments would likely be improved if SAX and LAX acquisitions are acquired in sequence.

RV strain is increasingly incorporated as a marker of RV diagnostics [21, 22]. Generally, studies with samples of under 20 controls or patients investigating RV reproducibility have shown fair results for free wall peak strain and diastolic strain rate [8, 22‐24]. With a sample of 125 patients, we observed good reliability for RV peak strain and diastolic strain rate, demonstrating both systolic and diastolic functions of the RV can be reliably measured in a patient population. However, we excluded a higher proportion of RV FT measurements compared to the LV. During imaging exams, the LV may be prioritized over the RV and we observed minor wrap or artifact in the RV that may not significantly hamper visual assessments of function but did create tracking errors. Similarly, multiple images were excluded because of pulsatile flow artifacts over the RV, especially at the base of the RV free wall. Tricuspid valve excursion is higher than the mitral valve, resulting in rapid regional movement in the RV. In combination with these pulsatile artifacts, tracking was impacted. Moreover, only RV agreement was compromised by the 3 T magnetic strength in comparison to 1.5 T, and it is known pulsatile flow artifacts can be worse at the higher field because of more susceptibility. The other factor that impacted RV analysis was arrhythmia and poor gating. As RV analysis was restricted to shortening in the LAX and thus only conducted in one plane, if this image has poor image quality or tracking, there are not other slices or planes to compensate as observed in the LV.