Reproducibility assessment of rapid strains in cardiac MRI: Insights and recommendations for clinical application

Purpose: Studies have shown the incremental value of strain imaging in various cardiac diseases. However, reproducibility and generalizability has remained an issue of concern. To overcome this, simplified algorithms such as rapid atrioventricular strains have been proposed. This multicenter study aimed to assess the reproducibility of rapid strains in a real-world setting and identify potential predictors for higher interobserver variation. Methods: A total of 4 sites retrospectively identified 80 patients and 80 healthy controls who had undergone cardiac magnetic resonance imaging (CMR) at their respective centers using locally available scanners with respective field strengths and imaging protocols. Strain and volumetric parameters were measured at each site and then independently re-evaluated by a blinded core lab. Intraclass correlation coefficients (ICC) and Bland-Altman plots were used to assess inter-observer agreement. In addition, backward multiple linear regression analysis was performed to identify predictors for higher inter-observer variation. Results: There was excellent agreement between sites in feature-tracking and rapid strain values (ICC ≥ 0.96). Bland-Altman plots showed no significant bias. Bi-atrial feature-tracking and rapid strains showed equally excellent agreement (ICC ≥ 0.96) but broader limits of agreement ( ≤ 18.0 % vs. ≤ 3.5 %). Regression analysis showed that higher field strength and lower temporal resolution ( > 30 ms) independently predicted reduced interobserver agreement for bi-atrial strain parameters ( ß = 0.38, p = 0.02 for field strength and ß = 0.34, p = 0.02 for temporal resolution).


Introduction
In the last decade, multiple studies have proven the incremental value of strain imaging in comparison to long-established parameters such as ejection fraction (EF) or the presence of focal fibrosis regarding prognosis in various cardiac diseases [1][2][3][4][5].However, reproducibility and generalizability of strain imaging has remained an issue of concern.
Since the implementation of post-processing approaches for echocardiography [6] and later feature-tracking strain for cardiac magnetic resonance imaging (cardiac MRI) [7][8][9], multiple studies have evaluated differences between modalities as well as software solutions [10][11][12].In addition, research has also aimed at identifying potential confounding factors in specific diseases [13] and at defining optimal temporal and spatial resolution [14] for feature-tracking strain measurements.However, contrary to these efforts, commercial vendors still blind their users to the details of their algorithms, virtually creating black-box models [7].This has prompted efforts to simplify the algorithms to make them more comprehensible.Recently, rapid strain approaches have been introduced as alternatives to conventional feature-tracking strains for the assessment of deformation patterns in different cardiomyopathies.
Rather than tracking the tissue of the myocardial wall over time, this method calculates longitudinal shortening of both atria and the left ventricle (LV) in relation to the atrioventricular valve planes.To do so, it does not require specialized post-processing software nor does it require significant amounts of processing power, making it a simple and fast alternative for cardiac function assessment of the LV and atria.
After these were initially only used as research tools, they were more recently adopted in clinical software solutions.While previous studies have made efforts to evaluate reproducibility, they often focused on imaging protocols in a research context [15,16], thereby not taking the potential impact of different scanner types and acquisition protocols into account.Therefore, this study aimed to assess the reproducibility of simplified ventricular and bi-atrial cardiac MRI strain parameters in a real-world, multicenter setting and to identify factors causing measuring variations.

Materials and methods
The agreement of cardiac MRI rapid strain measurements was assessed in a multicenter, multi-vendor trial.Locally responsible institutional ethics committees approved the study protocol, which was in line with the Declaration of Helsinki with a waiver for informed consent.

Study population
Each of the four participating sites retrospectively identified 20 patients and 20 healthy controls (HC) of a local examined/assessable cohort.Inclusion criteria for the patient populations were clinical indication for cardiac MRI with pathological findings.Clinical records, including additional diagnostics tests, subsequently verified the diagnosis.For cross-center analyses, patients were grouped by three morphological disease phenotypes: hypertrophic, dilated and indistinct.
Inclusion criteria for the HC were no history, signs or symptoms of cardiac diseases, no cardiovascular risk factors (e.g., hypertension or diabetes), and normal left and right ventricular function according to institutional reference ranges.

Cardiac MRI acquisition
The minimum requirement for all imaging protocols was the use of a retrospectively gated standard balanced steady-state free precession (bSSFP) sequence for cine imaging.Short-axis stacks completely covering both the left and right ventricle and standardized cardiac long axis views (2-, 3-and 4-chamber) were mandatory.Compressed-sensing sequences as well as studies with artefacts (wrap around, respiratory/ cardiac ghosting, image blurring/mistriggering, metallic artefacts, shimming artefacts or signal loss [17]) were excluded from this study according to the study's prospectively drafted standard operating procedures (SOP).

Image analysis
Prior to the segmentation of the datasets, all sites were given an SOP for volumetric and strain measurements, the core lab had drafted.Then, one observer from each site performed LV, left atrial (LA), and right atrial (RA) volumetric analysis on all 40 subjects by manually correcting the AI-derived end-diastolic and end-systolic contours in the 2-and 4chamber long-axis views as well as the short-axis stacks, carefully excluding the left ventricular outflow tract and papillary muscles.All observers had a minimum of 5 years of experience in cardiovascular imaging (site 1: 5 years, site 2: 23 years of experience, site 3: 8 years, site 4: 6 years).All measurements were performed using dedicated software (cvi42, version 5.13.2,Circle Cardiovascular Imaging Inc., Calgary, Canada).
Rapid LV strain was calculated by measuring the relative longitudinal shortening of the distance between the atrioventricular junction and the apex.This method involved assessing the changes in length along the longitudinal axis of the left ventricle during the cardiac cycle.Similarly, rapid bi-atrial strains were measured by assessing the relative longitudinal shortening and extension of the distance between the atrioventricular junction and the apex of both atria.LV and LA results were averaged from both long axes according to the biplane-area length method [18], while RA results were derived only from the 4-chamber view.Additionally, feature-tracking global longitudinal strain (GLS) analysis was performed using all long-axis sequences, while the previously manually corrected AI-derived contours were not modified.A clinical example of both feature-tracking and rapid strain analysis can be seen in Fig. 1.
After analysis by the local site observers and contours were deleted, the core lab observer (***blinded for peer review***, 6 years of experience) then re-evaluated all 160 studies by either traveling to the site (sites 2 and 4) or remotely accessing the studies (sites 1 and 3).The core lab did not provide data for this study to prevent bias.

Statistical analysis
Statistical analysis was performed using dedicated software (SPSS Statistics for Windows; v23.0;IBM Corp Armonk, NY, USA).
The Shapiro-Wilk test was used to assess the normal distribution of continuous data.Continuous data are either expressed as mean ± standard deviation (SD) or as with median with interquartile range (Q1, Q3), as appropriate.Categorical data are expressed as numbers and proportions.As strain values relate to an end-diastolic reference frame, the values for atrial strain are positive (atrial expansion during systole) while ventricular strain values are negative (ventricular shortening during systole).To evaluate the interobserver agreement between the site and the core lab, two-way random-effects model intraclass correlation coefficients (ICC) and Bland-Altman plots were used.For the Bland-Altman analyses were performed with the core lab measurements as the reference standard, differences of the means were calculated by subtracting the sites' means from the core lab's means and limits of agreement (LoA) were found by 95 % confidence intervals of the mean difference (1.96 × SD).Levels of agreement were defined as follows: poor, <0.5; moderate, 0.5-0.75;strong, 0.76-0.9;excellent, >0.9 [19].
Differences between site observers' and core lab results were evaluated by paired t-tests.The interobserver variation was determined per subject by dividing the mean absolute difference of the strain values (pooled sites-core lab) by the mean value of all observers for the respective parameters.The results were classified as: low (<10 %); intermediate (11-20 %); high (21-30 %); and very high (>30 %) [20].
The predictive value of influencing factors on the interobserver variation was assessed by multiple regression analysis with backward elimination.Variables entered into the regression analysis were the MRI scanner vendor, magnetic field strength, phases per cardiac cycle, type of core lab evaluation (on-site visit/remote access), heart rate, and temporal resolution.The latter was calculated per subject by dividing the RR-Interval by the acquired phases per cardiac cycle.
To control for the different diagnoses as potential confounding variables, subjects were classified into four morphological groups as follows: (I) hypertrophic disease phenotype (hypertrophic cardiomyopathy (HCM), amyloidosis), (II) dilated disease phenotype (dilated cardiomyopathy (DCM)), (III) other disease phenotypes (e.g.myocarditis), (IV) healthy controls.The analyses were performed per group and results were compared between pooled site observers and the core lab.A twosided p-value <0.05 was considered statistically significant.

Baseline characteristics and acquisition parameters
Each participating site contributed data from 20 patients and 20 HC, resulting in a total of 160 subjects (median 49 (18 to 85) years; 64 (40 %) females).The patient cohort included 38 (48 %) patients with a hypertrophic disease phenotype (11 patients with amyloidosis, 27 patients with HCM), 30 (38 %) patients with a dilated disease phenotype, and 12 (14 %) with an indistinct disease phenotype (7 with ischemic cardiomyopathy, 5 patients with acute myocarditis).Detailed baseline characteristics for all sites are provided in Table 1.
All patients and HC had undergone cardiac MRI at their respective centers using locally available scanners from two manufacturers (Siemens Healthcare; Philips Healthcare) with either 1.5 T (sites 1, 3 and 4) or 3 T (sites 2 and 4).Key acquisition parameters for all sites can be found in Table 2.

Interobserver agreement
Overall, there was excellent agreement for LV parameters, including rapid strain between site observers and the core lab (ICC ≥ 0.96).In addition, there were no significant differences between the site observers' results and the core lab evaluation for both LV volumetric and LV strain parameters (Fig. 2).
In addition, a strong-to-excellent agreement was also observed for biatrial volumetric and junctional strain parameters (LA: ICC ≥ 0.93; RA: ICC ≥ 0.89) and there were no significant differences between the site observers' results and the core lab (Fig. 3).In further analysis, Bland-Altman plots (Fig. 4) revealed no systematic bias for ventricular or atrial strain parameters (mean differences all ≤±1.2 %), but the LoA were broader for atrial rapid strains compared to ventricular rapid strains and feature-tracking GLS (LoA ≤ 18.0 % vs. ≤3.5 %).Detailed results of the interobserver agreement between pooled site observers and the core lab are given in Table 3.

Analysis per site
In the site-specific analysis, the overall results were mostly confirmed.All volumetric LV parameters showed excellent agreement between site observers and the core lab (ICC ≥ 0.91), except for EF at site 2, which showed strong agreement (ICC 0.79).Both feature-tracking GLS and LV rapid strain showed excellent agreement (ICC ≥ 0.92).Detailed results of the interobserver agreement of LV parameters per site are given in Supplemental Table S1.
LA volumetric parameters showed excellent agreement between site observers and the core lab (ICC ≥ 0.95), while RA volumetric parameters showed strong-to-excellent agreement (ICC ≥ 0.83).Both LA and RA feature-tracking and rapid strains demonstrated strong-to-excellent agreement (ICC ≥ 0.82).Detailed results of the interobserver agreement of bi-atrial parameters per site can be found in Supplemental Table S2.

Analysis per cohort
In the analysis based on morphological groups, the pooled results of the site observers and the core lab revealed strong-to-excellent agreement for LV volumetric values and all strain parameters for the 38 patients with a hypertrophic disease phenotype (ICC ≥ 0.89), excellent agreement for the 30 patients with a dilated disease phenotype (ICC ≥ 0.90) and excellent agreement for the remaining 12 patients with other morphological disease phenotypes (ICC ≥ 0.90).The cohort of 80 healthy controls showed strong-to-excellent interobserver agreement (ICC ≥ 0.81).Detailed results of the interobserver agreement of all LV parameters per cohort can be found in Supplemental Table S3.
Further, the bi-atrial measurements showed excellent agreement for volumetric and strain parameters within the patients with a hypertrophic disease phenotype (ICC ≥ 0.96), strong-to-excellent agreement in patients with a dilated disease phenotype (ICC ≥ 0.88), and strong-toexcellent agreement for the patients with an indistinct morphological disease phenotype (ICC ≥ 0.81).In the healthy controls, the agreement for RA volumetric and strain parameters was moderate-to-strong (ICC ≥ 0.68) while LA parameters showed strong-to-excellent interobserver agreement (ICC ≥ 0.89).Detailed results of the interobserver agreement of bi-atrial parameters per cohort can be found in Supplemental Table S4.

Interobserver variation
Overall, the interobserver variation of ventricular strain measurements between the site observers and the core lab was at 7.7 % while the interobserver variation of bi-atrial measurements was slightly larger at 10.6 %.In the analysis per site, the largest interobserver variation for ventricular strain parameters was found for site 1 and the smallest for site 4 (13.1 % vs. 5.4 %).This observation was confirmed in bi-atrial strain measurements (18.9 % vs. 6.0 %).In the analysis per morphological disease phenotype, healthy controls had the lowest interobserver variation between pooled sites and the core lab for all strain parameters (6.0 % for ventricular and 6.9 % for bi-atrial strain), whereas the indistinct disease phenotype showed the largest interobserver variation for ventricular strain (20.7 %) and the dilated disease phenotype showed the largest interobserver variation for bi-atrial (23.3 %).Detailed results for all analyses of interobserver variation are given in Table 4.
The regression analysis was performed on healthy controls to minimize bias.For ventricular strain parameters, a 5-step model did not show a significant influence of scanner type, field strength, phases per cardiac cycle, heart rate, or temporal resolution on the interobserver variation.
For atrial strain parameters however, a 4-step model revealed that a higher field strength and lower temporal resolution had a significant predictive value for higher interobserver variation (ß = 0.38, p = 0.02 for field strength and ß = 0.34, p = 0.02 for temporal resolution).Of note, temporal resolution was eliminated from the prediction model when performing the regression analysis of this study only on the sites with a mean temporal resolution > 30 ms (sites 2-4).Detailed results of both regression analyses are summarized in Table 5.

Discussion
This multicenter study aimed to assess agreement of simplified junctional left ventricular and bi-atrial strain parameters between different observers, scanner types, field strengths and sites.This was   The main results are: (I) Rapid LV and bi-atrial strain parameters showed strong-to-excellent interobserver agreement (II) LV strain parameters showed low interobserver variability while bi-atrial parameters showed intermediate interobserver variability (III) higher field strength as well as lower temporal resolution were associated with higher interobserver variation in bi-atrial strain measurements.
Despite most commercially available software solutions relying on feature-tracking methods, significant inter-vendor differences have been demonstrated for ventricular [10] as well as atrial strain measurements [12].In the absence of free access to vendor-specific algorithms, differences between them often get lost in the 'black box' [7] and simplified strain algorithms have been proposed [15,16].These require far less sophisticated algorithms and can therefore potentially be more easily implemented by different vendors.In addition, they are also less demanding than feature-tracking algorithms in terms of manual corrections and quality assurance needed by the radiologist.

Ventricular parameters
After the adaptation of these research tools into commercial software, previous single-center studies on ventricular strain parameters found no significant bias for different field strengths [21] and demonstrated non-inferiority compared to feature-tracking strain [16].These results were confirmed in the present study, despite this multicenter study showing a larger degree of interobserver variation in the setting compared to the single center setting (7.6 % vs. 3.4 % [16]).Of note, the interobserver variation for all ventricular functional and volumetric parameters such as LV EF and LV end-systolic volume index (ESVi) as well as the feature-tracking strain were similarly low (<10 %).
In the analysis per participating center, site 2 showed lower reproducibility of LV EF due to differences in LV ESVi.Ventricular strain and bi-atrial volumetric and strain parameters however were unaffected and showed excellent reproducibility.The most likely explanations for this could be the greater prevalence of LV hypertrophy in the patient cohort of this center compared to the others, resulting in more manual corrections needed for the endocardial end-systolic contours.In a similar 2.0 18.5 0.97  Continuous parameters are expressed as medians with interquartile range (Q1, Q3).

Table 5
Stepwise Backward Regression Analysis for Interobserver Variation.fashion, the hypertrophy could have also led to differences in contouring of trabeculations in these contours.The fact that the strain values were less affected by this can be explained by the algorithm, which prioritizes end-diastolic over end-systolic contours in order to compute strain.The regression analysis did not show a significant influence of the type of core lab analysis (on-site visit/remote access), scanner vendor, field strength, phases per cardiac cycle, heart rate, or temporal resolution.Thus, the results of this multi-center study corroborate previous singlecenter studies in finding that simplified ventricular rapid strain parameters can be reliably performed in a real-world setting [15,16,21].

Atrial parameters
Despite broad agreement on the potential that quantification of atrial function and deformation offers in terms of diastolic assessment and prognostic implications in various diseases [22][23][24][25], it currently remains a research tool.There are several reasons for this, such as the thin atrial walls complicating the contouring of the atrium and anatomical difficulties like the orifices of the pulmonary veins and atrial appendages impairing feature-tracking [26,27].
Similarly, to previous studies on simplified LA strain parameters, this study demonstrated no significant inter-observer bias [26].However, the LoA were much broader than previously published.The most likely explanation for this is the more diverse cohort of patients compared to a rather homogenous cohort of heart failure patients in prior trials [26].Interestingly, there were no relevant differences between volumetric and strain parameters.The interobserver variations were low (<10 %) for all LA volumetric parameters as well as all RA strain and volumetric parameters but intermediate for LA rapid strain.This superior reproducibility of the RA rapid strains compared to LA rapid strains is especially noteworthy because all RA parameters were only derived from the 4-chamber long-axis view, while LA parameters were averaged from 4chamber and 2-chamber long-axis views.The most likely explanation for this is that the averaging not only adds another datapoint for more accurate anatomical representation of the LA but also adds an additional layer of potential measurement variations.
Previous studies have found superior interobserver agreement for right ventricular strain at lower field strengths [13].Despite the higher spatial resolution at 3 T which would suggest superior tracking, it is also known that image artifacts such as pulsatile flow artifacts are typically more pronounced at higher field strengths.This potential cause of maltracking is further aggravated because cardiac MRI acquisition are commonly centered around ventricular assessment and sophisticated post-processing techniques are not routinely applied to the atria.Therefore, in this real-world scenario, flow artifacts which did not inhibit visual assessment of the atria might have been tolerated at the time of acquisition but hampered strain assessment nevertheless.This hypothesis is further supported by the fact, that ventricular strain assessment was not significantly confounded by higher field strength.
In terms of temporal resolution as a confounding factor of interobserver reproducibility, however, the available literature is inconsistent.While Fischer et al. [13] did not report temporal resolution as a relevant confounding factor, Backhaus et al. [14] found that higher temporal resolution lead to better reproducibility for right ventricular strain parameters.While the transferability of these results to atrial strain measurements remains unclear, the possible reasons for the influence of the temporal resolution can be transferred to strain measurements in general.These include improved tracking quality due to smaller magnitudes of movement within the smaller time frames between phases and less need for extrapolation between the data points [14].Accordingly, when performing the regression analysis of this study only on the sites with a mean temporal resolution < 30 ms (sites 2-4), the temporal resolution was eliminated from the backward regression model at the first step.Based on these results, it is recommended to acquire the cine sequences with a minimum temporal resolution of 30 ms.

Limitations
First, only one post-processing software was used for strain measurements.In addition, due to its more complex contraction pattern, the right ventricular strain could not be semi-automatically assessed by the simplified strain algorithm and was therefore beyond the scope of this study.
Second, only adult subjects were included, so transferability to children remains to be determined.Lastly, because all sites routinely used intravascular contrast agents, this effect on strain quantification could not be studied.

Conclusion
Simplified left ventricular and bi-atrial strain parameters show strong-to-excellent agreement in a real-world multi-center setting, despite different scanner vendors, field strengths, and imaging protocols.Higher field strength and lower temporal resolution were independently associated with higher interobserver variation in bi-atrial strain measurements.When assessing atrial deformation, it is recommended to acquire the underlying cine sequences with a minimum temporal resolution of 30 ms.

Fig. 1 .
Fig.1.Clinical example of a four-chamber view bSSFP cine sequence from a 71 year old female healthy volunteer.The top row shows images from the end-diastolic phase while the bottom row shows images from the end-systolic phase.Feature-tracking GLS can be seen on the left with myocardial tracking points (yellow) and rapid strain contours are shown on the right (white line marks longitudinal extend between apex and atrioventricular valve plane).

Fig. 2 .
Fig. 2. Boxplots comparing the site observers' results (blue boxes) and the core lab results (orange boxes) for volumetric LV parameters (left) and LV strain parameters (right).

Fig. 3 .
Fig. 3. Boxplots comparing the site observers' results (blue boxes) and the core lab results (orange boxes) for volumetric bi-atrial parameters (left) and bi-atrial strain parameters (right).

Fig. 4 .
Fig.4.Bland-Altman plots showing the agreement between the site observers' results and the core lab for LV rapid strain (left), LA rapid strain (middle), and RA rapid strain (right).Note that the scales for ventricular and atrial strains are different due to differences in limits of agreement.

Table 1
Baseline characteristics for each site.

Table 3
Interobserver agreement over all parameters and cohorts.

Table 4
Variation of Strain Parameters Over All Subjects per Site and per Cohort.