Time-efficient three-dimensional transmural scar assessment provides relevant substrate characterization for ventricular tachycardia features and long-term recurrences in ischemic cardiomyopathy

Delayed gadolinium-enhanced cardiac magnetic resonance (LGE-CMR) imaging requires novel and time-efficient approaches to characterize the myocardial substrate associated with ventricular arrhythmia in patients with ischemic cardiomyopathy. Using a translational approach in pigs and patients with established myocardial infarction, we tested and validated a novel 3D methodology to assess ventricular scar using custom transmural criteria and a semiautomatic approach to obtain transmural scar maps in ventricular models reconstructed from both 3D-acquired and 3D-upsampled-2D-acquired LGE-CMR images. The results showed that 3D-upsampled models from 2D LGE-CMR images provided a time-efficient alternative to 3D-acquired sequences to assess the myocardial substrate associated with ischemic cardiomyopathy. Scar assessment from 2D-LGE-CMR sequences using 3D-upsampled models was superior to conventional 2D assessment to identify scar sizes associated with the cycle length of spontaneous ventricular tachycardia episodes and long-term ventricular tachycardia recurrences after catheter ablation. This novel methodology may represent an efficient approach in clinical practice after manual or automatic segmentation of myocardial borders in a small number of conventional 2D LGE-CMR slices and automatic scar detection.

Delayed gadolinium-enhanced cardiac magnetic resonance (LGE-CMR) imaging represents a well-established method to identify infarct-related scar tissue based on signal intensity criteria after gadolinium administration 1 .
Several studies have shown that scar tissue quantification using LGE-CMR may increase the predictive performance of ventricular arrhythmic events in patients with established myocardial infarction 2,3 . Nonetheless, scar characterization remains controversial and has not been incorporated as a standard stratification tool for ventricular arrhythmic events or ventricular tachycardia (VT) characterization in patients with ischemic cardiomyopathy (ICM) 4 . Two-dimensional LGE-CMR sequences with large slice thickness (normally 8 mm) are the most common images in the clinic 5 . Conversely, 3D sequences with higher and isotropic resolution are substantially less common in daily clinical practice, although frequently performed in research studies 6 . State-of-art 3D LGE-CMR sequences require larger acquisition time, appropriate patient collaboration, and time-consuming manual or semiautomatic processing for accurate myocardial segmentation 7 . Altogether, these drawbacks make 3D LGE-CMR sequences potentially inefficient in highly demanding agendas during regular clinical practice. The common alternative using 2D sequences provides sufficient information for assessing ventricular function. However, 2D-derived scar reconstructions may have limitations to properly assess the myocardial substrate compared to isotropic high-resolution 3D data. Moreover, these differences might have implications on accurate characterization of the ventricular substrate associated with VT episodes in patients with ICM.
Recently, we have developed a 3D methodology to assess ventricular scar using custom transmural criteria and a semiautomatic approach to obtain transmural scar maps 8 . This approach can be also implemented in upsampled 3D reconstructions 9 obtained from conventional 2D LGE-CMR sequences. This methodology might provide an intermediate and time-efficient approach to improve performance of 2D LGE-CMR sequences to characterize the myocardial substrate particularly associated with clinical presentation of ICM-related VT episodes and potential recurrences after ablation. This may be especially relevant in patients without primary prevention indication for implantable cardioverter defibrillator (ICD).
We hypothesize that transmural scar assessment using 3D-upsampled data from 2D-LGE-CMR sequences will be sufficient and superior to conventional analysis from 2D-LGE images to identify myocardial regions relevant for infarct-related substrate characterization in patients with ICM and VT episodes. Patient-specific scar reconstructions from isotropic 3D-acquired sequences were used as benchmark comparisons. LGE-CMR sequences to test and optimize the performance of a 3D transmural-based methodology that takes into account individual-specific myocardial geometry 8 . The validation and clinical impact of the method were tested in the patients' cohort undergoing 3D and 2D LGE-CMR studies within the same protocol. Figure 1 shows the study flowchart. The data that support the findings of this study are available from the corresponding authors upon reasonable request.

Methods
Pig model of myocardial infarction. Pigs (male, large-white strain, ~ 35 kg) underwent percutaneous catheterization of the left anterior descending coronary artery using femoral access and fluoroscopic guidance under general anesthesia. An angioplasty balloon was inflated in the coronary artery for 60 min to generate myocardial infarction. Then, the balloon was deflated and a coronary angiogram was recorded to confirm patency of the coronary artery and reperfusion as described elsewhere (see also Supplementary Information) 7 .
Magnetic resonance studies. Ten-to-twelve weeks after myocardial infarction pigs underwent substrate characterization using a Philips Achieva 3T-Tx whole-body scanner equipped with a 32-element and phased-array cardiac coil (Philips Healthcare, Best, The Netherlands). Seven minutes after intravenous contrast injection (0.2 mM/kg. Dotarem, Guerbet) 3D LGE-CMR sequences were acquired using an inversion-recovery spoiled turbo field echo (IR-T1TFE) with isotropic resolution of 1.5 × 1.5 × 1.5 mm (reconstruction resolution 0.57 × 0.57 × 0.75 mm). Segmented and ECG-gated cine steady-state free precession was also performed to acquire 11-13 contiguous short-axis slices covering the ventricles from the base to the apex to evaluate left ventricular ejection fraction (LVEF) (

Three-dimensional scar volume reconstruction.
LGE-CMR sequences, in both pigs and patients, were carefully segmented using a semi-automatic approach with custom-made software in Matlab (Mathworks Inc, Natick, US). First, we obtained a region of interest, which roughly contained the heart and minimized surrounding tissue from other intrathoracic structures. Second, we automatically delineated the endocardial and epicardial contours using an automatic algorithm with an active contour method. Third, this initial myocardial segmentation was subsequently complemented with fine manual segmentation by one single operator and further reviewed by a second expert operator ( Fig. 2A,B). This approach, although time-consuming on 3D-acquired sequences (~ 8-12 h) for further implementation in clinical practice, was used as a benchmark reference for highly accurate patient-specific scar volume reconstructions. Finally, scar segmentation was performed using a full-width-half-maximum method 1 . In patients, maximum signal intensity in 3D isotropic acquisitions was calculated using a slice thickness of ≈ 8-mm (i.e. average of 12 neighbouring voxels in the through-plane direction) to prevent any potential signal-to-noise effect from smaller pixel sizes compared to calculations in lower resolution 2D sequences ( Supplementary Fig. 1). Scar threshold criterion was set at 0.45 of the maximum signal intensity based on our previous data using high resolution in vivo and ex vivo comparisons 7 . Representative Three-dimensional transmural-based scar assessment. In pigs, 3D LGE-CMR images in the coronal plane were first reoriented in the short axis using 3D Slicer (Slicer 4.10.2) and then reconstructed with different slice thicknesses. For each slice thickness reconstruction, scar segmentations were created by averaging the highresolution 3D scar segmentations to the new reconstructed slice thickness (Fig. 1). Downsampled images in pigs or 2D LGE-CMR images in patients, together with their associated myocardial segmentations, were upsampled to the 3D isotropic resolution using a topologic interpolation algorithm (Fig. 3A,B) as described elsewhere 9 . Three-dimensional transmurality maps of the left ventricle (i.e., the ratio of myocardial wall thickness covered by scar) were represented on the epicardial surface and computed as reported elsewhere 8 . This method also provides point-wise correspondences between the endocardium and the epicardium (as other authors have done in the past) 10 Mapping and ablation procedure. The invasive electrophysiology study was performed using percutaneous venous and arterial femoral access to reach the right and left ventricles, respectively. Additional epicardial access, if necessary, was achieved using a percutaneous subxiphoid approach. Ventricular geometries were generated using the Carto3 (Biosense Webster, Diamond Bar, CA) or the Ensite NavX (Abbott, St. Paul, MN) electroanatomic mapping system. VT induction was attempted using programmed ventricular stimulation from www.nature.com/scientificreports/ the right ventricular apex. All stimuli were synchronized with the intrinsic QRS complex. The protocol consisted of two sequential basic drive cycle lengths (CLs) of 10 beats each (S1) at 600 or 400 ms. Each basic drive CL was followed by up to 3 extrastimuli (S2, S3, S4) decremented in CL until reaching refractoriness or a minimum coupling interval of 200 ms. If the arising VT was hemodynamically stable, activation maps and entrainment maneuvers were performed with the ablation catheter or a 20-pole steerable mapping catheter (Pentaray, Biosense Webster) to localize the protected isthmus of the reentrant circuit. Ablation was performed with an open-irrigated-tip catheter (Navistar Thermocool, Biosense Webster or FlexAbility, Abbott). Direct current defibrillation was delivered to restore sinus rhythm in case of ventricular fibrillation or hemodynamic collapse. In the case of hemodynamic instability during mapping of VT morphologies documented on admission (i.e. spontaneous clinical VT), deductive reconstruction of the reentrant circuit and ablation was performed using pacemapping as described by de Chillou et al. 14   www.nature.com/scientificreports/ for correlation analysis between scar volumes or transmural-based scar areas and VT CLs. A p < 0.05 was considered statistically significant for differences in group comparisons and for a non-null Pearson correlation coefficient. The intraclass correlation coefficient was also used as a measure of concordance (ICC 2.1 version). All data were analyzed in Graphpad Prism6 (California, US) and custom-made software in Matlab.

Scar volume quantification showed direct correlation with the cycle length of spontaneous VT episodes. Scar volume quantification from the original 2D
LGE-CMR sequences showed larger scar volumes than those from 3D-acquired LGE reconstructions (Table 1, Fig. 5A). Conversely, left ventricular wall volume (including myocardium and scar) was not significantly different between 2D-and 3D-acquired ventricular reconstructions (Fig. 5B). Data from one representative patient who underwent sequential 3D-2D-3D CMR- LGE acquisitions further supported larger scar volume from 2D images compared to scar volumes obtained from any of the 3D acquisitions ( Supplementary Fig. 6). Both 2D-and 3D-acquired scar volume reconstructions (see representative examples in Fig. 5C) showed statistically significant correlation with the CL of spontaneous VT episodes documented at the time of hospital admission (r = 0.57; p = 0.04 and r = 0.61; p = 0.03, respectively. Figure 5D). The larger the scar volume, the slower was the CL of the spontaneous VT episode. More specific analysis using a double cutoff criterion at 0.45 and 0.67 of the maximum signal intensity (for heterogeneous and dense scar characterization, respectively) 7 , showed higher correlation coefficients using 3D-acquired heterogeneous scar volumes (r = 0.72; p = 0.01. Supplementary Fig. 7A). Conversely, dense scar volumes did not show any statistically significant correlation with spontaneous VT CLs (Supplementary Fig. 7B). The median CL of inducible VT episodes per patient did not show statistically significant correlation with 2D-or 3D-acquired total scar volumes (Fig. 5E).
Transmural-based scar assessment in 3D-upsampled models is sufficient to identify scar regions associated with the cycle length of spontaneous VT episodes. Scar area quantification in myocardial regions with 3D transmurality < 0.2 for 3D-upsampled and < 0.1 for 3D-acquired models showed significant correlation with the CL of spontaneous VT episodes (r = 0.69; p < 0.01 and r = 0.79; p < 0.01, respectively. Figure 6A,B). The correlation value using transmural scar areas from 3D-upsampled models was higher than the one obtained from scar volume reconstructions from 3D-acquired images (Figs. 6B, 5D, respectively). The transmurality threshold in 3D-upsampled models was selected using the highest correlation coefficient in the original 2D LGE-CMR slices (Fig. 6C,D). In fact, scar areas from 3D-upsampled models showed higher cor- LGE images (r = 0.48; p = 0.10) using the same transmurality threshold (< 0.2, Fig. 6C). A representative sample case of 3D visualization of scar areas using < 0.2 (for 3D-upsampled models) and < 0.1 (for 3D-acquired models) transmurality criteria is shown in Fig. 6E. Similar to correlation analysis using scar volume reconstructions (Fig. 5E), the median CL of inducible VT episodes per patient did not show statistically significant correlation with 3D transmural-based scar assessment (Supplementary Fig. 8).
Transmural-based scar assessment identifies myocardial substrates associated with long-term VT recurrences after ablation. Entrainment maneuvers and pace mapping were performed in 6 and 8 patients, respectively, to target and ablate the myocardial substrate associated with clinical VT morphologies. This and further substrate ablation to eliminate other inducible VT morphologies showed that most of radiofrequency lesions (75 [71, 81] %) were delivered in LAVA regions, although complete LAVA elimination was not pursued in any of the cases. The total ablation time during the procedure was 15.8 (10.1, 17.8) min. Noninducibility of clinical VT morphologies was achieved in 13 out of 14 cases. In 3 patients, other non-clinical and non-mappable morphologies remained inducible on programmed ventricular stimulation after ablation. Epicardial mapping was only attempted in one patient, although severe pericardial adherences did not enable further mapping of the region of interest and radiofrequency energy was only delivered from the endocardium. At 5 years of follow-up VT recurrences of any morphology were documented in 5 out of 15 patients. Four patients were under class III antiarrhythmic drugs at the end of the follow-up. Patients with VT recurrences showed lower scar areas on 3D-acquired and 3D-upsampled models using < 0.1 and < 0.2 transmurality criteria,  . (B,C) Correlation analysis between the cycle length (CL) of spontaneous VT episodes and scar areas for wall transmurality < 0.2 (for 3D-upsampled models), < 0.1 (for 3D-acquired models) and < 0.2 [for original 2D LGE-CMR slices, in (C)]. (D) Pearson correlation coefficients between scar assessment at sequential TsM thresholds and the CL of spontaneous VT episodes. (E) 3D visualization of scar areas (in blue) using < 0.2 (for 3D-upsampled models) and < 0.1 (for 3D-acquired models) transmurality criteria in the sample case shown in (A).  Fig. 7B). Further analysis establishing boundaries in the myocardial wall depth at 0.33, 0.67 and 0.97 of the myocardial thickness showed that the majority of scar tissue in the regions with 3D transmurality < 0.1 (for 3D-acquired models) and < 0.2 (for 3D-upsampled models) was located in the endocardial layer (97.56% and 80.49%, respectively, Fig. 7C,D and Supplementary Figs. 9, 10).

Discussion
This study shows that semiautomatic transmural scar assessment on 3D models from both 3D-upsampled and 3D-acquired LGE-CMR sequences provide clinically relevant scar characterization in patients with established myocardial infarction. The CL of spontaneous VT episodes significantly correlated with transmural scar values mainly reflecting endocardial scar. Moreover, patients with VT recurrences after a long-term follow-up showed significantly lower scar areas on 3D-upsampled and 3D-acquired models compared with patients without recurrences. Conversely, scar volume quantifications did not reach statistically significant differences between patients with and without VT recurrences. The latter reflects the relevance of endocardial scar assessment rather than the entire scar volume in ICM-related VT episodes. In fact, transmurality criteria at < 0.1 (for 3D-acquired models) and < 0.2 (for 3D-upsampled models), mainly reflected endocardial scar (Fig. 7C,D). Hence, the methodology we propose in this paper may represent an efficient approach in clinical practice after segmentation of myocardial borders and automatic scar detection. Manual accurate segmentation in a small number (≈ 12) of conventional 2D LGE-CMR slices can be performed in less than 15 min by expert operators. However, reliable novel software tools and machine learning algorithms will be a time-efficient alternative for the initial segmentation step 16 .
Several studies have shown that scar tissue quantification using LGE-CMR images may represent an independent predictor of ventricular arrhythmic events after myocardial infarction 3,17 . However, CMR-based scar characterization and quantification to predict ventricular arrhythmic events or VT features remains controversial 18 . Moreover, these associations have not been consistent across studies 19,20 . Imaging resolution, myocardial and scar segmentation, and partial volume averaging represent important factors that may explain the lack of uniform results and variable scar identification criteria among series 21,22 . Partial volume averaging may also explain the differences we documented in myocardial wall thickness using 2D and 3D LGE-CMR images. Larger scar www.nature.com/scientificreports/ extension is expected to be present in lower resolution images 7,21 , which might substantially affect the predictive performance of scar assessment from LGE-CMR sequences. In fact, our results are consistent with larger scar volumes in patient-specific 3D models obtained from 2D LGE-CMR sequences compared to those from 3D LGE-CMR studies (Fig. 5A). This may explain higher correlation coefficients between the CL of spontaneous VT episodes and scar volumes obtained from higher-resolution 3D-acquired sequences. However, 3D sequences require larger acquisition times compared to 2D sequences (Table 1) and time-consuming reliable segmentation on sequences with > 100 slices, which directly affect regular implementation in clinical practice. The proposed methodology based on 3D transmural scar assessment, using a < 0.2 transmurality threshold in 3D-upsampled models from conventional 2D LGE-CMR sequences, may help to overcome these limitations since scar areas provided higher correlation coefficients with the CL of spontaneous VT episodes than reconstructed scar volumes from 3D-acquired LGE sequences. This highlights that the transmural-based approach was able to identify scar regions relevant for ICM-related spontaneous VT episodes. In other myocardial substrates with less relevance of endocardial scars, 3D transmural-based scar assessment would also permit quantifying scar using different myocardial layer criteria. A relevant limitation for further implementation of CMR-LGE sequences as a standard stratification tool in the clinic is the of lack of consensus on a uniform scar evaluation criterion and signal intensity thresholding for scar tissue 6,23 , which varies among series and makes substrate characterization particularly sensitive to sampling bias. This requires additional efforts among the scientific community to reach a more general agreement on scar identification. Here, we used a single signal intensity threshold value to simplify scar tissue differentiation from healthy myocardium. We chose a scar cutoff at 0.45 of the maximum signal intensity based on previous analyses in animal models reporting that below such threshold, in high-resolution postcontrast CMR images, remote myocardial areas started to show false positive scar detection 7 . Although mild fibrotic remodeling may also be present in remote ventricular regions, outside the infarcted region, using a 0.45 criterion on signal intensity we documented that this remote scar was ≤ 1% of the remote myocardium, which is consistent with histopathological data reported in animal models with myocardial infarction 24 . Our results further support that a single 0.45 criterion on signal intensity may be a good reference cutoff for substrate characterization in patients.
Our data provide new insights into the role of scar characterization in clinical presentation and long-term follow-up of patients with infarct-related VT episodes undergoing catheter-based ablation, especially in patients without primary ICD indication (14 out of 15 patients in this series). Lower scar areas in patients with VT recurrences at 5 years of follow-up might reflect more viable myocytes within scar regions 25 , which settles a heterogeneous substrate with areas of potentially slow conduction leading to reentrant ventricular arrhythmias. Two small series by Woie et al. 26 and Alexandre et al. 17 have previously shown that the number of scar islands and scar size, respectively, were associated with the mean CL of VT episodes recorded by ICDs over a segment of 12-to-13 VT intervals from the first non-sustained or sustained monomorphic VT morphology. These short recording periods for analysis have limitations to extrapolate results to longer duration VT episodes or clinically relevant spontaneous episodes admitted to the emergency department. Our current prospective series including patients admitted to hospital with spontaneous VT, without neither previous documented episodes nor ICD, supports the role of LGE-CMR imaging to identify relevant myocardial regions and potentially predict VT features in patients with established myocardial infarction. The results are also consistent with other series suggesting that scar extension is not associated with the mean CL of inducible VT episodes during an electrophysiological study 27 . Similar results were reported in a previous study using a pig model with established myocardial infarction, in which we did not document any statistically significant correlation between total, heterogeneous or dense scar volumes with the median CL of inducible VT episodes after programmed ventricular stimulation 25 . Altogether these results highlight that other factors, beyond the underlying substrate, may be involved in inducible VT episodes (e.g. functional and unstable re-entrant circuits, stimulation site, etc.).
Limitations. Time-to-imaging acquisition after gadolinium administration was not the same for 3D and 2D LGE-CMR acquisitions. However, in both sequences scar identification was performed using the same spatial resolution and signal intensity criterion. Scar volume results were also consistent with previous data reporting the implications of imaging resolution and partial volume effects on scar assessment 7,21 . Data from one representative patient undergoing sequential 3D-2D-3D LGE-CMR acquisitions also supported the notion that lower resolution images show larger scar regions (Supplementary Fig. 6).
Clinical implications of transmural scar assessment in spontaneous VT features were only studied in patients with established myocardial infarction. Other ventricular events as ventricular fibrillation and polymorphic VT episodes have not been studied in this series, which mainly included patients with mild to moderate decrease in LVEF, without ICD at the time of hospital admission 4 . New data from currently ongoing trials (e.g. CMR-GUIDE trial; NCT01918215) will provide more insight into this specific question. Statistical association of scar areas derived from transmurality analysis with VT recurrences might have been affected by individual specific substrate modification during the ablation procedure. Further series are warranted to confirm these descriptive results in relatively limited ablation strategies without extensive scar homogenization.

Conclusion
Three-dimensional transmural scar assessment in ventricular models reconstructed from 3D-upsampled or 3D-acquired LGE sequences may provide relevant scar characterization for clinical presentation and long-term ablation outcomes in patients with infarct-related spontaneous VT episodes.