Discussion
This study describes the first use of a novel method for 3D cine CMR feature-tracking (FT) strain analysis using sequentially stacked 2-dimensional cine CMR images. This study also demonstrates the utility of this method in patients with DMD CMP. We showed that localized strain metrics derived from this 4D image analysis are both sensitive and specific for characterizing DMD CMP disease severity. To our knowledge, this is the first report that localized surface area strain and strain rate metrics derived from 3D cine CMR were used to characterize CMP in a DMD cohort. We observed that basal Ea, Ecc, and Err peak strain and strain rate were the most sensitive and specific metrics for differentiating DMD CMP from healthy control subjects. In general, we found that for every strain metric assessed, slice-level localized strain rate was better able to differentiate between DMD and control subjects compared to corresponding peak strain alone. Basal Ea systolic strain rate had the best differentiation with an AUC of 0.96. Basal Err, basal Err systolic strain rate and basal Ecc systolic strain rate magnitude values were significantly decreased in mild cardiomyopathy (LGE-, LVEF > 55%) compared to a healthy control group. Localized Ea and Ecc peak strain and strain rate metrics also had the strongest correlation with LVEF, T1, and ECV values while Err and Ell were less strongly correlated with LVEF, T1, and ECV. These data suggest that localized and kinematic analysis of 3D cine CMR images in DMD patients may provide a more robust analysis for assessing CMP than global or peak strain values alone.
Patients with DMD universally develop cardiomyopathy [
45]. However, the age of onset, the time course, and the severity of CMP are highly variable necessitating more refined measures for assessing CMP severity [
14,
15]. Although LGE is present as early as 7 years of age it is only apparent when a significant amount of damage has occurred. Traditional functional assessment by LVEF is limited as it is only abnormal in later stages of disease when the process may no longer be reversible. As such, developing early markers of disease can shift the treatment paradigm from rescue to prevention. These early abnormalities provide novel biomarkers and surrogate outcome measures of disease progression.
In this study, we observed the most significant differences, largest AUC values, and strongest correlation to LVEF in basal circumferential and basal E
a when comparing DMD patients to healthy controls. The results demonstrate the value of E
cc as an imaging biomarker in DMD and are consistent with other studies [
46,
47]. In a recent study, Siddiqui et al. also showed 3D CMR-derived E
cc was better able to predict the onset of DMD CMP than conventional 2D CMR-derived strain values [
48]. Similar to our technique, Siddiqui et al. accomplished 3D FT in DMD cardiomyopathy using 3D interpolation of the endocardial and epicardial boundaries from 2D slices. This 3D FT strain technique has been shown to have superior reproducibility compared to 2D FT in CMR and has been well described by Liu et al. [
49].
To our knowledge, no previous studies have explored the role of strain rate derived from 3D cine CMR in a population of patients with DMD-associated CMP. Strain rate has been a predominantly echocardiographic-based measure likely due to its superior temporal resolution (< 5 ms) compared to CMR (~ 20–50 ms) [
50]. Additionally, strain rate measurements that use tagging may be less reproducible due to tag fading [
51]. In the CMR FT technique we present here, we do not have issues with tag fading which provides an advantage for strain rate estimation. In addition, our method of interpolation between frames allows us to estimate strain rate despite a relatively lower temporal resolution. However, as with all strain and strain rate estimation methods, it should be noted that variability in heartrate will impact the temporal resolution differently for each cardiac cycle analyzed. This variability may affect strain rate estimations. Our strain rate calculation method, like others used for CMR [
52] is based on calculating the slope of the strain curve and as such values may be under- or overestimated depending on heartrate. We do note a difference in heartrate between our DMD and healthy control subjects (Table
1), however there is no significant difference between average heartrates between the DMD subgroups we analyzed in this study. One major advantage to our slope calculation and normalization method is that this standardization allows for some added consistency when comparing strain curve slope and shape between patients and groups- even when heartrate is variable within a single patient scan or between multiple scans and patients. These benefits and limitations, however, should be considered when examining the significance of strain rate findings.
In previous studies, strain rate used in assessing myocardial infarction showed good reproducibility using 2D CMR FT [
53] and in a healthy control group for both 2D and 3D CMR FT [
49]. In this study, we found that for every strain metric assessed, slice-level localized strain rate was better able to differentiate between DMD and healthy control subjects compared to corresponding peak strain alone. These results suggest that valuable information might be missed when only peak strain values are considered. Strain rate differences, particularly those between DMD groups with increasing CMP severity may be an early indication of mechanistic changes in the heart. In myocardial infarction, strain rate has been shown to correlate with regional ischemia and akinetic regions [
54,
55]. One explanation for strain rate differences in DMD CMP therefore could be an early manifestation of regional heterogeneity of systolic function that is spatially correlated with regional fibrofatty replacement of healthy myocardium. Further work using animal models or larger clinical datasets may help elucidate a mechanistic explanation for these findings.
Within the group of DMD patients we analyzed (n = 43), we were able to identify three distinct groups- LGE-/LVEF > 55%, LGE + /LVEF > 55%, and LGE + /LVEF < 55%. In our particular cohort of subjects, no patient was observed to be both LGE- and have an LVEF < 55%. Interestingly, we observed that a few of our tested metrics—basal Err, basal Err systolic strain rate and basal Ecc systolic strain rate magnitude values—were significantly decreased in each DMD group compared to a healthy control group. Other metrics including peak basal Ecc, peak basal Ea, and basal Ea systolic strain rate showed significant differences between LGE-/LVEF > 55% and LGE + /LVEF < 55% groups. These results suggest that regional strain metrics derived from 4D CMR may be able to detect early dysfunction even prior to LGE or overt LV dysfunction and differentiate between more mild and severe disease. A comprehensive longitudinal study describing these changes over time in the same patients would be a valuable extension of this work.
While strain rate measurements improved differentiation for every strain quantity we calculated over peak strain alone, we did not observe strong correlations between radial strain and LVEF, T1, T2, or ECV or longitudinal strain and LVEF, T1, T2, or ECV. This may be due in part to the characteristic pattern of DMD associated CMP which primarily affects myofibers in the subepicardial free wall of the LV, though as the condition progresses, transmural fibrosis becomes increasingly more prevalent [
44]. Importantly, we also note that we estimate longitudinal strain using the stack of short-axis images with limited resolution in the longitudinal acquisition plane. This low spatial and contrast resolution may contribute to less reliable feature-tracking and E
ll estimations. Another reason we may not be observing these correlations is the wider variation of E
ll and E
rr strain compared to E
cc, making correlative measures less reliable. A 3D FT CMR analysis in DMD patients done by Siddiqui et al. [
48] similar to ours showed insignificant differences in global E
ll between DMD and healthy control subjects but did observe differences in global E
rr and global E
cc. This study also reported similar ranges to those we found for E
ll, E
cc, and E
rr, derived from 3D FT CMR in DMD CMP. A different meta-analysis examining global longitudinal strain measured by 2D speckle tracking echocardiography in eight studies showed that global longitudinal strain and circumferential strain but not radial strain were significantly decreased in DMD vs healthy subjects, though the study did show heterogeneity in results [
56]. As with all strain estimation methods, differences in acquisition modality, method, and analysis should all be considered when interpreting results.
In our analysis we observed significant differences in E
rr between healthy and DMD patients, though relatively lower AUC values compared to E
cc and E
a. This may be due in part to the wide range of E
rr values in our method. This variability could be due to dyskinesia resulting in a shift of time of contraction or pathological issues related to radial thickening seen in the DMD patients. It could also be a tracking issue due to through-plane motion from circumferential and longitudinal deformation exacerbated by thinner myocardial walls in more advanced cardiomyopathy. E
rr estimation has historically been a more difficult metric to measure consistently. For example, Cao et al. showed moderate differences in E
cc and E
ll between vendors using CMR FT, but very large differences in E
rr [
57]. Despite these considerations, we observed that E
rr may still be a valuable metric and warrants further validation and comparison in DMD populations.
We also explored the use of E
a and E
a strain rate metrics from 3D + time CMR imaging in DMD patients. This metric, while unique to CMR analysis in DMD patients, has been explored with 3D speckle tracking echocardiography (3D-STE) in DMD CMP. For example, Yu et al
. demonstrated that E
a derived from 3D-STE had an 85.7% sensitivity and a 71.0% specificity for differentiating DMD patients (n = 56) from controls (n = 31) [
58]. E
a is a relatively novel metric, unique to 3D imaging, that takes into account both longitudinal and circumferential shortening [
59]. Also, since the myocardium is relatively incompressible, radial thickening during systole influences E
a setting up an inverse relationship between E
a and E
rr. The integration of these effects into a single strain parameter makes E
a potentially useful in examining subclinical dysfunction. Since this is a relatively novel parameter, more studies are needed to determine its full value, especially in CMR imaging. Conventional echocardiographic strain estimation techniques often have a higher temporal resolution (< 5 ms) compared to CMR (20–50 ms), though CMR offers superior contrast resolution [
50]. Additionally, methodological differences make a direct comparison of strain values between STE and CMR difficult; depending on technique and study population, these values may not be in agreement [
60]. Transthoracic echocardiography (TTE) is used for the initial screening of cardiac function in nearly every patient population, including those at risk for DMD. However, as the disease progresses, limited TTE windows and image artifacts due to scoliosis and fat deposition make cardiac assessment with TTE increasingly difficult [
61]. In many DMD patients, only a small number of measures of LV function can be reliably estimated from TTE [
61]. Thus, while 3D-STE is a promising characterization technique in some patients, CMR remains the gold standard for evaluation in this patient population.
Many other strain imaging techniques are currently being used to analyze CMR data. The technique described in this work is best characterized as a 3D feature-tracking Lagrangian deformation estimation technique that utilizes image features in CMR scans to estimate strain. Ventricular boundaries, brightness, and homogeneity are all tracked throughout the cardiac cycle to produce deformation. Others have used similar 2D feature-tracking techniques in DMD cohorts to detect morphologic changes in the absence of LGE as well as to detect changes between DMD and healthy subjects not seen using 3D-STE [
62]. One major benefit to the feature-tracking approach we use is that it utilizes conventional short-axis cine images and thus does not increase the complexity or length of a typical CMR imaging protocol. Another major advantage of our platform is that it provides additional and more comprehensive 3D imaging strain metrics (i.e. surface area strain, strain rate) compared to conventional metrics typically available through other commercial platforms. The nature of our platform also allows for raw export of image and segmentation data for further, more extensive analysis- one example being the 48-point surface representation of the ventricle which lends itself well to machine and deep learning algorithms. We note as well that this 48-point surface representation does not employ a more traditional basal, mid-ventricular, and apical sections, but rather utilizes a 4-slice length-based analysis (25%, 50%, 75%, 100%). We found that the use of this 4-slice representation allowed for reasonably accurate feature tracking while also permitting a simple and rapid analysis. One drawback to this method is the need to interpolate between slices. A benefit to this approach, however, is that the segmentation framework is adaptable such that more or less slices can easily be added to enhance future analysis. A major drawback to our method is that the post-processing steps for each scan require a moderate amount of training and additional time to complete. Non-expert, non-clinical users of the graphical user interface felt comfortable with its use after 20–30 min of training. After some practice, users were able to complete a full 4D analysis in 30–45 min. While this analysis time is likely not feasible outside of a research setting, efforts are being made to further simplify this approach and incorporate machine learning techniques to automate manual segmentation and analysis [
24].
It is important to note that all CMR strain imaging techniques have benefits and drawbacks. HARmonic Phase (HARP) analysis is perhaps the most utilized strain imaging method incorporated into the largest number of post-processing software packages [
63‐
65]. HARP is a CMR tagging strain analysis technique that isolates one Fourier component of the amplitude modulated data, and tracks pixels with consistent phase [
64]. This allows for a relatively rapid and reproducible strain estimation compared to other techniques that involve additional scan time. HARP analysis in conjunction with CMR tagging has been used to reliably estimate strain in DMD patients allowing for robust patient stratification [
46,
47]. While this technique is useful in many clinical scenarios, additional scan time, lack of standardization, low spatial resolution, and tag fading are all drawbacks to this method [
66,
67].
Displacement encoding with stimulated echoes (DENSE), strain encoded CMR (SENC), and tissue phase mapping (TPM) are additional methods for strain estimation using CMR, but these are less studied in DMD populations [
46,
47]. DENSE is generally accepted as an accurate and reproducible method for strain imaging as it relies on the phase information of a stimulated echo and is directly proportional to tissue displacement. By analyzing directional-encoded phase images, the Lagrangian displacement fields can be produced. SENC is similar to HARP in that it utilizes parallel tagged lines with out-of-plane phase encoding gradients to estimate strain. Finally, TPM relies solely on the pixel phase from which it encodes velocity from each image allowing for spatial integration and estimation of deformation and strain. Each of these techniques relies on specialized image sequences and takes additional scan time, but they produce relatively high spatial and temporal resolution needed for strain estimates. In addition, each of these methods is not widely used clinically, though they are being used in research studies [
68‐
70]. Each of the additional CMR strain imaging techniques mentioned (HARP, DENSE, SENC, and TPM) are most often used on 2D images, though they could be used in 3D reconstruction techniques for strain estimation. Fully 3D CMR acquisition is being explored but the acquisition time is longer than conventional 2D scanning. Additionally, techniques must be robust enough to overcome motion artifacts, and excitation of 3D volume may diminish image contrast between blood and myocardium [
71].
The primary purpose and scope of this work was to determine the feasibility and utility of our 3D cine strain analysis technique in DMD CMP. While we see similar differences between DMD and healthy control subjects, the ranges for strain values may differ slightly from other techniques. This is not a unique issue in our proposed method as studies have shown differences in 2D CMR strain where FT techniques and in different cardiac pathologies. For example, Chew et al. demonstrated that CMR FT may overestimate strain when compared to SENC in adult and pediatric congenital heart disease cohorts [
72]. Others have shown good agreement between CMR FT and HARP measuring E
cc in a DMD population [
73], and CMR FT and DENSE in adults with myocardial infarction [
74]. Even when using CMR FT on the same cohort of patients with CMP, inter-vendor differences were found in E
cc and E
ll and E
rr [
57]. Small differences have also been shown between 2 and 3D derived CMR FT strain values in DMD [
48] and healthy adult populations [
49]. As with most sequential cine acquisitions it is also possible that we are very slightly underestimating peak strain because our raw image temporal resolution (20–25 images per cardiac cycle) may not perfectly capture the moment of peak myocardial contraction especially with higher heartrate. Thus, as with any other technique, one should interpret the control and DMD strain value ranges considering these differences between techniques, vendors, and disease characteristics. Future prospective studies will be needed to fully compare and validate the approach described here with 2D and 3D CMR strain imaging data acquired from the same patients. Analysis of a larger longitudinal cohort will also allow us to determine if we can predict outcomes from the novel metrics we derive from our 4D CMR strain method.
Limitations
Our study has several limitations. Although we noted significant differences in many strain metrics between DMD and healthy control subjects, this is a cross-sectional study and not a longitudinal study. As such, longitudinal analysis of patients would add stronger evidence as to whether these measures are sensitive and specific and can determine longitudinal changes that may correlate with disease progression, and even mortality. For this study we considered 43 DMD CMP patients and 25 healthy control subjects. While we were able to perform analysis on an age- and sex-matched cohort, this relatively smaller sample size did not allow for robust modeling to account for all confounding factors such as height, weight, or blood pressures. A larger study may be needed to fully validate these findings in light of many potential confounders. Additionally, for many of our strain metrics (E
ll, E
rr, and E
a) we used the small strain linear approximation which may not be as accurate as a finite strain approximation for larger strain values. That said, the relative differences between DMD and healthy subjects should still be evident. Another limitation in data analysis is the labor-intensive process to contour and segment the LV while considering CMR image artifacts caused by gross movement of the patient or diaphragm necessitating manual correction. An automated or even semi-automated approach using machine learning techniques to correct displacement and aid in segmentation may help this process [
24]. Further, the short axis cine images used to create the 3D + time datasets are acquired sequentially and not simultaneously. In addition, the use of a simplified 48-point representation to create a 3D dynamic mesh relies on both spatial and temporal interpolation. While this representation may lend itself well to a simplified feature-tracking method, further work will be needed to validate its use.