A novel multiparametric imaging approach to acute myocarditis using T2-mapping and CMR feature tracking

After obtaining approval by the local institutional review board, data sets of 86 patients who had been consecutively referred to our department for CMR imaging after clinical diagnosis of acute myocarditis (mean symptom duration before referral: 4.8 ± 4.4 days; all patients demonstrated symptoms for less than 14 days) were retrospectively analyzed. The clinical diagnosis of myocarditis was based upon the clinical suspicion of acute myocarditis and on the clinical criteria in the current recommendations [22] (Table 1). Patients with CMR findings characteristic for other diseases than myocarditis were excluded from further analyses. Of 86 available data sets, 4 were excluded due to severe respiratory motion artifacts in T2 mapping data resulting in non-diagnostic image quality. 2 patients had to be excluded because of lacking T2 black blood (T2 BB), EGE and LGE imaging and 4 datasets were excluded from further analyses due to severe motion artifacts in cine imaging not suitable for subsequent FT analyses. Finally, datasets of 76 patients with suspected myocarditis were analyzed in a retrospective fashion. Of these 76 datasets, only patients with preserved LVEF (defined as ≥ 50%) were included in further statistical analyses (n = 67).

CMR data from 17 age and gender matched healthy volunteers served as control. Healthy subjects were selected as previously described [23], inclusion criteria being: i) uneventful medical history, ii) no symptoms of inflammation, iii) absence of any symptoms indicating cardiovascular dysfunction, iv) normal cardiac dimensions and function on cine CMR. For each volunteer written informed consent was obtained prior to the study after approval by the local institutional review board.

Characteristics of patients and controls are shown in Table 2.

CMR was performed on a 1.5 T MR system (Achieva, Philips Healthcare, Best, The Netherlands) using a standard five-element cardiac phased array coil and a 4-lead vectorcardiogram. A balanced steady-state free precession (b-SSFP) sequence in breath-hold technique and with retrospective ECG-triggering was acquired in three horizontal long axes and a stack of short axes (SAX) covering the left ventricle (LV) for functional analysis and subsequent Feature Tracking analyses. Imaging parameters were chosen as previously described [23]. Volumetry was performed on a standard post-processing platform (IntelliSpace Portal, Version 6, Philips Healthcare).

In patients and controls, edema-sensitive black blood T2-weighted images with fat saturation were acquired in SAX orientation covering the entire LV [24]. Myocardial early gadolinium enhancement was assessed in all patients using fast spin- echo T1-weighted images in axial orientation during the first minutes after 0.1 mmol/kg Gd-DOTA (Dotarem; Guerbet, Villepinte, France) contrast administration as previously described [25]. LGE imaging was performed in all patients 15 min after 0.2 mmol/kg Gd-DOTA (Dotarem; Guerbet, Villepinte, France) contrast administration using an inversion-recovery gradient-echo sequence in the horizontal long axes and SAX as previously described [26].

T2-mapping data were acquired in 3 SAX slices evenly distributed across the LV (resulting in one apical, midventricular and basal slice) using an ECG-triggered, breathhold Gradient Spin Echo technique (GraSE) [27]. Image parameters were chosen as previously described [17]. A pixel-wise myocardial T2-map was generated using a mono-exponential fit [9] on the magnitude data using a maximum likelihood estimator (MLE), where a Rician noise distribution was assumed.

Statistical analysis was performed in R 3.1.2 [31], using the packages ggplot2 [32] for graphical visualization, pastecs [33] for descriptive statistics, rpart [34] for fitting single classification trees [35], and ROCR [36] for using receiver operating curve (ROC) analyses.

All continuous data are given as mean ± standard deviation. A p-value of < .05 was regarded as statistically significant. Testing for significant differences between patients and controls was performed using Wilcoxon sum rank test or Welch independent T-test, depending on normality distribution of the data. Classification models were built using multiple logistic regression analyses. Models were compared according to the Akaike Information Criterion (AIC; where lower AIC represents a better model fit). Optimal cut-off values were defined using classification trees as previously described [17]. The diagnostic accuracy as well as diagnostic sensitivity and specificity of optimal predictive parameters were evaluated using ROC analyses. In models with two (three) parameters, sensitivity was computed as the proportion of myocarditis patients with 2 (3) out of 2 (3) parameters exceeding the cut-off, relative to all myocarditis patients.

Significant differences between patients with suspected myocarditis and controls were observed for LV global longitudinal strain (GLS_LV) and SR, and LV global circumferential strain (GCS_LV; Fig. 1; Table 3). No significant differences were observed for any of the RV parameters (Fig. 1; Table 3). All strain parameters revealed a wide range of values in controls and patients as well as a considerable overlap between the values of the two groups (Fig. 1).

Among several multiple logistic regression models including different strain parameters, GCS_LV performed best as an independent predictor of suspected myocarditis according to the AIC (81.69). When using different strain parameters in combination, all models revealed collinearity of the selected strain parameters, as either one was only significant in the model when the other one was not included. ROC-analyses finally revealed an only moderate potential in differentiating healthy controls from myocarditis patients for the three LV strain parameters (Fig. 2a; Table 4).

All parameters derived from T2 mapping showed significantly higher values in patients with suspected myocarditis when compared to controls (Fig. 3, Table 3). Similar to the strain parameters, however, meanT2 and meanSD both demonstrated a large variation of values in myocarditis patients as well as a considerable overlap between both groups (Fig. 3).

Among several multiple logistic regression models including different T2 parameters alone or in combination, the model containing madSD as an independent predictor of suspected myocarditis performed best according to the AIC (72.04). The model combining madSD and maxT2 as previously recommended [20] resulted in a minimally higher AIC (73.08) and revealed some collinearity between the two parameters, as either one was only significant in the model when the other one was not included.

In ROC analyses, madSD resulted in the highest area under the curve (AUC) when compared to the other three T2 parameters (Fig. 2b, Table 4) and all single T2 parameters were superior to the single strain parameters.

All single LLC demonstrated only a moderate potential for discriminating between patients with suspected myocarditis and healthy controls (Fig. 2c; Table 4). A visual analysis of T2 BB edema imaging was superior to T2-ratio in our cohort.

Finally, different multiparametric imaging models were tested in a sequential fashion in order to evaluate, which combination exhibits the highest diagnostic potential in suspected acute myocarditis.

Multiple logistic regression analyses showed that models containing madSD and either GLS_LV or GCS_LV performed better than the model containing madSD alone (AIC 66.55 for madSD + GLS_LV and 68.12 for madSD + GCS_LV).

In ROC analyses, the multiparametric model containing madSD and GCS_LV resulted in the highest diagnostic performance (Fig. 4a; Table 4). Including GLS_LV instead of GCS_LV in the model resulted in an equal sensitivity but considerably lower specificity (Fig. 4a; Table 4). Both multiparametric models exhibited a slightly better diagnostic performance when compared to the model containing madSD alone and all three models were superior to the LLC (Fig. 4a; Table 4).

Adding LGE to both multiparametric imaging models resulted in further increased diagnostic performances (Fig. 4b, Table 4) and the model containing madSD, GCS_LV and LGE yielded the highest diagnostic sensitivity.

Using classification trees in order to define optimal cut-off values for the best multiparametric imaging model (madSD + GCS_LV + LGE) resulted in a cut-off of ≥1.8 ms for madSD, of ≥ −25% for GCS_LV and LGE positive. Using these cut-offs in combination led to an excellent diagnostic performance (Table 4). Using the two cut-offs for madSD and GCS_LV without gadolinium (LGE) led to a still excellent discriminatory performance (Table 4).

In the present study, GLS_LV and GCS_LV showed significant differences of mean values, but also revealed a considerable overlap between healthy subjects and patients with suspected myocarditis, what led to an only moderate diagnostic potential of these two strain parameters when used as single diagnostic parameters in myocarditis patients with preserved LVEF. These observations are in concordance with a previous study of our group focusing on biventricular strain parameters in a different small cohort of myocarditis patients [8] as well as with a recent study by André et al. In these studies, both LV strain parameters were significantly correlated with LVEF, elucidating, that larger differences of mean values between patients and controls might most likely be introduced by patient subgroups with impaired LVEF. Hence, as strain parameters are sought to be more sensitive diagnostic markers than LVEF and thereby might reflect different aspects of LV systolic function [37], we now focused on myocarditis patients with preserved LVEF only. This also might explain the lower diagnostic performance of GLS_LV in our study when compared to a recent study by Luetkens et al. [11]. They demonstrated a considerably better diagnostic performance of GLS_LV alone (AUC 0.83) in their small patient cohort, which might be due to the fact that they did not differentiate patients on the basis of LVEF and thereby introduced LVEF as a biasing factor.

In the present study, GCS_LV exhibited a slightly higher diagnostic potential than GLS_LV, either when used alone or in combination with madSD ± LGE. Nevertheless, the differences between the diagnostic potential of the two LV strain parameters was only small and there might be a role for both parameters in the diagnostic algorithm of suspected acute myocarditis, as already proposed by other groups [11, 16]. However, the better reproducibility of GCS_LV as compared to GLS_LV [38, 39] as well as the better agreement with tagging derived parameters representing the gold standard for myocardial strain analysis [39] speak in favor of using GCS_LV in future diagnostic approaches.

Like in the initial proof of concept [17] and the subsequent confirmation study [20], madSD exhibited a high diagnostic potential in the present study cohort and its diagnostic performance was better than for any other tested T2 parameter and even than combined LLC. This and the controversial results of previous studies using averaged myocardial T2 values for diagnosing suspected acute myocarditis [7, 10‐13, 18] underline the potential of the novel approach of “mapping tissue inhomogeneity” in acute myocarditis, as myocarditis lesions exhibit a focal nature and often are confined to small areas with abnormal T2 times compared to large areas of myocardial tissue exhibiting normal T2 times [40]. In addition, measuring global T2 times is of limited diagnostic value when it comes to multi-center, multi-vendor studies, as T2 times are known to vary with the sequence and field strengths used [23]. However, it is still unclear whether madSD might be more independent of such technical factors when compared to global T2 values. Further studies therefore should look at the potential of madSD to overcome the sequence and field strengths dependency [23] of mapping techniques.

Interestingly, combining madSD with maxT2 in a model revealed some collinearity between the two parameters in the present study and unlike the results of our previous studies, combining madSD with maxT2 in the present study cohort was not superior to using madSD alone as a predictor. As maxT2 represents a parameter, which might be more prone to outliers e.g. related to artifacts (as a single artifact in one segment has the inherent potential to cause a false-high segmental T2 value) [20], using madSD alone for detecting myocardial inflammation might be the better choice for future diagnostic algorithms.

Although madSD alone already yielded an excellent diagnostic performance in the present study cohort, the proposed multiparametric models exhibited an inherent potential to further improve the diagnosis of suspected acute myocarditis. Using the best multiparametric model in the present study, i.e. the combination of madSD, GCS_LV and LGE missed only 2 patients with suspected acute myocarditis (compared to 28 missed by LLC). This parallels the results of Luetkens et al. [11] who recently could demonstrate the incremental diagnostic value of a combination of LGE with global myocardial T2 over LLC or single quantitative CMR parameters.

Even without the use of gadolinium, the combination of madSD and GCS_LV exhibited an excellent diagnostic performance in our study cohort, missing only 5 patients with suspected acute myocarditis. This approach might prove especially advantageous in patients with contraindications for the use of gadolinium-based contrast agents. In addition, there is a particular need for a diagnostic tool allowing an identification of LGE-negative subjects with acute myocarditis [11].

We tested the diagnostic performance of different parameters in myocarditis patients compared to a group of healthy subjects. However, the most relevant challenge for the novel quantitative CMR parameters will be their discriminative power in myocarditis patients when compared to patients with other non-inflammatory cardiomyopathies [7] or between patients with acute and chronic forms of myocarditis, as CMR generally appears to provide a better diagnostic performance in the acute setting of myocarditis than in subacute or chronic stages [6, 7, 40, 41]. In addition, we did not include T1 and ECV mapping into our multiparametric models. Yet, two recent studies including T1 and ECV mapping could already demonstrate that only T2, but not T1 or ECV provided an adequate diagnostic performance in patients with acute as well as chronic symptoms and between patients with inflammatory and non-inflammatory dilated cardiomyopathy [7, 12]. Moreover, Luetkens et al. showed an equal diagnostic performance for global myocardial T1 and T2 values in acute myocarditis [11], both parameters being influenced by myocardial water content. Finally, the addition of clinical e.g. serological parameters such as Troponin or markers of inflammation to a multiparametric imaging model might further improve its diagnostic potential. Therefore, future studies will have to address the question whether a combination of T1 and T2 mapping or of other clinical, e.g. serological parameters adds some value to the diagnostic work-up of suspected acute myocarditis. Moreover, further studies should focus on the discriminatory power of madSD between patients with acute and chronic forms of myocarditis or between other types of cardiomyopathy.

The present study has several limitations. First, there are inherent drawbacks of the retrospective study design. Future larger (and multicenter) studies with a prospective selection of patients with suspected acute myocarditis should be performed in order to validate the presented proof of principle experiment of a novel multiparametric imaging approach. In addition, further studies should also evaluate the discriminatory power of the presented multiparametric imaging approach in order to differentiate between patients with acute and chronic forms of myocarditis as well as between other types of cardiomyopathy.

The present study was performed using a clinical reference standard of myocarditis patients and we did not perform EMB. However, EMB is unnecessary or even contraindicated in myocarditis patients with normal systolic function [42]. Hence, we carefully defined the present patient cohort based on clinical criteria similar to other studies using a clinical reference standard [5, 11, 13, 14, 17]. Nevertheless, some patients in whom a differential cause of disease might have been missed despite careful patient selection may have confounded our study cohort.

The results concerning the novel T2 parameter madSD still must be considered as preliminary. As so far only two studies have investigated the diagnostic potential of madT2 in myocardial inflammation, this needs further validation or re-estimation in larger prospective, multi-center studies.