Super-resolution reconstruction of late gadolinium-enhanced MRI for improved myocardial scar assessment
2014; Wiley; Volume: 42; Issue: 1 Linguagem: Inglês
10.1002/jmri.24759
ISSN1522-2586
AutoresOleh Dzyubachyk, Qian Tao, Dirk H. J. Poot, Hildo J. Lamb, Katja Zeppenfeld, Boudewijn P. F. Lelieveldt, Rob J. van der Geest,
Tópico(s)Advanced X-ray and CT Imaging
ResumoJournal of Magnetic Resonance ImagingVolume 42, Issue 1 p. 160-167 Original ResearchFree Access Super-resolution reconstruction of late gadolinium-enhanced MRI for improved myocardial scar assessment Oleh Dzyubachyk PhD, Corresponding Author Oleh Dzyubachyk PhD Department of Radiology, Leiden University Medical Center, Leiden, The Netherlands Supported by grant 10894 of the Dutch Technology Foundation STW (Stichting Technische Wetenschappen), grant MEDIATE2-09039 of the EU ITEA, and the IMDI Heart in 4D project of ZonMw.Address reprint requests to: O.D., Division of Image Processing, Department of Radiology, Leiden University Medical Center, Postbus 9600, 2300 RC Leiden, The Netherlands. E-mail: [email protected]Search for more papers by this authorQian Tao PhD, Qian Tao PhD Department of Radiology, Leiden University Medical Center, Leiden, The Netherlands Supported by grant 10894 of the Dutch Technology Foundation STW (Stichting Technische Wetenschappen), grant MEDIATE2-09039 of the EU ITEA, and the IMDI Heart in 4D project of ZonMw.Search for more papers by this authorDirk H.J. Poot PhD, Dirk H.J. Poot PhD Departments of Radiology and Medical Informatics, Erasmus MC — University Medical Center Rotterdam, Rotterdam, The NetherlandsSearch for more papers by this authorHildo J. Lamb MD, PhD, Hildo J. Lamb MD, PhD Department of Radiology, Leiden University Medical Center, Leiden, The NetherlandsSearch for more papers by this authorKatja Zeppenfeld MD, PhD, Katja Zeppenfeld MD, PhD Department of Cardiology, Leiden University Medical Center, Leiden, The NetherlandsSearch for more papers by this authorBoudewijn P.F. Lelieveldt PhD, Boudewijn P.F. Lelieveldt PhD Department of Radiology, Leiden University Medical Center, Leiden, The Netherlands Intelligent Systems Department, Delft University of Technology, Delft, The NetherlandsSearch for more papers by this authorRob J. van der Geest PhD, Rob J. van der Geest PhD Department of Radiology, Leiden University Medical Center, Leiden, The NetherlandsSearch for more papers by this author Oleh Dzyubachyk PhD, Corresponding Author Oleh Dzyubachyk PhD Department of Radiology, Leiden University Medical Center, Leiden, The Netherlands Supported by grant 10894 of the Dutch Technology Foundation STW (Stichting Technische Wetenschappen), grant MEDIATE2-09039 of the EU ITEA, and the IMDI Heart in 4D project of ZonMw.Address reprint requests to: O.D., Division of Image Processing, Department of Radiology, Leiden University Medical Center, Postbus 9600, 2300 RC Leiden, The Netherlands. E-mail: [email protected]Search for more papers by this authorQian Tao PhD, Qian Tao PhD Department of Radiology, Leiden University Medical Center, Leiden, The Netherlands Supported by grant 10894 of the Dutch Technology Foundation STW (Stichting Technische Wetenschappen), grant MEDIATE2-09039 of the EU ITEA, and the IMDI Heart in 4D project of ZonMw.Search for more papers by this authorDirk H.J. Poot PhD, Dirk H.J. Poot PhD Departments of Radiology and Medical Informatics, Erasmus MC — University Medical Center Rotterdam, Rotterdam, The NetherlandsSearch for more papers by this authorHildo J. Lamb MD, PhD, Hildo J. Lamb MD, PhD Department of Radiology, Leiden University Medical Center, Leiden, The NetherlandsSearch for more papers by this authorKatja Zeppenfeld MD, PhD, Katja Zeppenfeld MD, PhD Department of Cardiology, Leiden University Medical Center, Leiden, The NetherlandsSearch for more papers by this authorBoudewijn P.F. Lelieveldt PhD, Boudewijn P.F. Lelieveldt PhD Department of Radiology, Leiden University Medical Center, Leiden, The Netherlands Intelligent Systems Department, Delft University of Technology, Delft, The NetherlandsSearch for more papers by this authorRob J. van der Geest PhD, Rob J. van der Geest PhD Department of Radiology, Leiden University Medical Center, Leiden, The NetherlandsSearch for more papers by this author First published: 19 September 2014 https://doi.org/10.1002/jmri.24759Citations: 13 The first two authors contributed equally to this work. AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract Purpose To develop and validate a method for improving image resolution of late gadolinium-enhanced (LGE) magnetic resonance imaging (MRI) for accurate assessment of myocardial scar. Materials and Methods In a cohort of 37 postinfarction patients, LGE was performed prior to ventricular tachycardia catheter ablation therapy at 1.5T. A super-resolution reconstruction (SRR) technique was applied to the three anisotropic views: short-axis (SA), two-chamber, and four-chamber, to reconstruct a single isotropic volume. For compensation of the interscan heart motion, a joint localized gradient-correlation-based scheme was developed. Scar was identified as either core or gray zone in both the SRR and original SA volumes, and evaluated based on the clinically established bipolar voltage range of the in vivo electroanatomical voltage mapping (EAVM). Results Compared to the SA volume, the SRR method resulted in significantly (P < 0.05) reduced myocardial scar gray zone sizes (10.5 ± 8.8 g vs. 9.2 ± 8.1 g) and improved agreement of the bipolar voltage range of scar gray zone (0.99 ± 0.65 mV vs. 1.46 ± 1.15 mV). Conclusion We propose an SRR method to automatically reconstruct a high-quality isotropic LGE volume from three orthogonal views. Analysis of the in vivo EAVM demonstrated improved myocardial scar assessment from the SRR volume compared with the SA LGE alone. J. Magn. Reson. Imaging 2015;42:160–167. © 2014 Wiley Periodicals, Inc. MYOCARDIAL INFARCTION is a major cause of cardiac arrhythmia and sudden cardiac deaths. Accurate characterization of the postinfarction myocardial scar has important clinical implications for patient treatment and management, such as risk stratification and catheter ablation 1, 2. In recent years, late gadolinium-enhanced (LGE) magnetic resonance imaging (MRI) has become the gold standard technique for myocardial scar imaging in clinical practice 3. In particular, the size of partially infarcted myocardial tissue, named the gray zone, has been shown to be an important predictor of adverse cardiac events, outperforming traditional function parameters such as left ventricular ejection fraction 1, 4. The accuracy of myocardial scar characterization is closely dependent on image resolution. In a recent study, Schelbert et al 5 showed with animal data that the myocardial scar gray zone is especially susceptible to the partial volume effect (PVE), which causes significant overestimation of gray zone size. In a previous human study, it was shown that the MR-visualized myocardial scar size often exceeds the electrophysiologically defined size 6. These findings underscore the importance of LGE image resolution for accurate myocardial scar characterization in clinical research. Unfortunately, current clinical LGE acquisition typically has poor through-plane resolution (between 5 to 10 mm) compared to in-plane resolution (around 1.5 mm), resulting in significant anisotropy and PVE. To better visualize scar in different orientations, MR protocols often include three LGE acquisitions in orthogonal views (by "orthogonal" we refer to both orthogonal as well as nearly orthogonal cases): short-axis (SA), two-chamber (2CH), and four-chamber (4CH). Nevertheless, to evaluate each view separately remains suboptimal with PVE in different orientations. In recent years, the feasibility of navigator-gated free-breathing 3D LGE MRI techniques for isotropic myocardial scar imaging has been demonstrated 7, 8. However, today the most commonly applied approach remains successive acquisition of multiple views using separate breath-hold acquisitions. In this regard, to combine the three existing MR acquisitions provides a theoretically and practically sound solution, which can capture the myocardial scar in all three dimensions in fine resolution, without altering the protocol or introducing new sequences. Super-resolution reconstruction (SRR) is a general term that defines a group of methods that recover a high-resolution image from multiple low-resolution views of the same scene. In application to MR data, this technique is typically used when high-resolution acquisition is either not technically possible or infeasible due to, eg, too long scanning time 9. The original SRR was developed for low-resolution images with subvoxel shift in the slice-select direction 10. Later, Shilling et al 11 showed that the quality of the reconstruction can be improved by applying it to rotated rather than shifted views. Recently, the SRR methodology has been shown to be able to successfully reconstruct an isotropic volume from several anisotropic orthogonal views 12-16. Since then, orthogonal SRR was successfully applied to various types of MR data: neurological DWI 12, 13, 17, otolaryngological 14, and cardiac 15, 18. In the latter, the authors suggested reconstructing isotropic heart volumes from anisotropic orthogonal views typically acquired in clinical practice. However, their reconstruction was only successful for two orthogonal views 15, whereas usage of all three orthogonal views resulted in quality decrease. Recently, Shi et al 16 presented a novel super-resolution algorithm for orthogonal myocardial scar reconstruction. Their method is based on learning image patches from a database, and thus requires extensive prior training. Good alignment of the low-resolution images is a prerequisite for successful reconstruction, as it is based on the fundamental SRR assumption that all the images depict the same scene. Such alignment is achieved by performing image registration, for which a large variety of methods exist 19. In particular, correlation-based registration methods 20 are popular for recovering unknown translation, rotation (only in 2D), and scaling as they are exhaustive, and thus do not require initialization. Recently, Tzimiropoulos et al 21 developed a gradient-correlation-based method, which is shown to be more robust than other correlation-based methods, in particular, for MR data 22. The purpose of this study was to apply SRR to improve image resolution of LGE MRI for more accurate assessment of myocardial scar. Preliminary results of this research were presented at MICCAI 2013 23. This article considerably extends the conference paper with respect to methodology, number of test datasets, and quality measures used for validation. MATERIALS AND METHODS MRI Acquisition Thirty-seven postinfarction patients (31 male and 6 female, age 64 ± 12 years) who were referred for MR prior to catheter ablation of ventricular tachycardia were involved in this study. For retrospective anonymized studies from routine patient care in Dutch University Medical Centers, Institutional Review Board approval is not required. The patient cohort had on average a myocardial infarction percentage (relative to LV) of 15 ± 10%, and an ejection fraction (LVEF) of 38 ± 10%. A 1.5T Gyroscan ACS-NT MRI scanner (Philips Medical Systems, Best, The Netherlands) equipped with Power Track 6000 gradients and five-element cardiac synergy coil was used. After acquiring the scout and cine sequences, a Look-Locker (LL) sequence was acquired ∼15 minutes after bolus injection of gadolinium DTPA (Magnevist; Schering, Berlin, Germany; 0.15 mmol/kg) at one SA level. After the LL acquisition, T1-weighted LGE images were acquired with an inversion-recovery 3D turbofield echo sequence with parallel imaging (SENSE, acceleration factor 2) at late diastolic phase. The inversion time was determined based on the LL sequence to null the normal myocardial signal. All views were acquired at the same late-diastolic phase. Due to the high number of slices required for full left ventricle coverage in the SA view, it was typically acquired as two stacks in two separate breath-holds. For the other two views, images were acquired in one stack within a single breath-hold. For the SA acquisition, the slice thickness was 10 mm with 5 mm overlap; for the other two acquisitions, the slice thickness was 12 mm with 6 mm overlap. Signal outside the field of view was suppressed using two saturation slabs to avoid fold-over artifacts. All images were acquired using a scan matrix size 256 × 206 and reconstructed to 256 × 256 pixels with in-plane resolution of 1.56 × 1.56 mm2. Typical parameters were as follows: field of view 400 × 400 mm2; flip angle 15°; echo time 1.06 msec; and pulse repetition time 3.7 msec. One inversion pulse per heartbeat was used. Heart Registration Proper alignment of the low-resolution volumes is a crucial step for success of the subsequent SRR. LGE MR is typically performed during a breath-hold to reduce motion artifacts. This may cause significant variation in position of the heart in the cardiac views acquired during different breath-holds. To compensate for the interscan heart displacement, we perform registration of two views by using the gradient-based correlation algorithm 21. Such registration allows estimation of the interscan heart motion with voxel precision without requiring initialization. More precise estimate of the registration parameters (subvoxel displacement, rotation, and scaling) is performed during the SRR, as explained in the "Orthogonal Super-Resolution Reconstruction" section. In general, our registration framework, which is illustrated in Fig. 1, follows the one described previously 22. Initially, all views are transformed to the same coordinate space. In this case, the coordinate system associated with the SA volume was used, which is the customary way to analyze these kind of data in clinical practice. Next, the volume of interest containing the heart and the area around it is calculated as the largest cuboid H that entirely fits in the overlap region R of all three views (see Fig. 2). Finally, the gradient-based correlation map between each volume pair is calculated 21. The most probable displacement vector is calculated from the position of the highest peak of the corresponding correlation map where x defines location on the correlation map in Cartesian coordinates. Figure 1Open in figure viewerPowerPoint Flowchart of the algorithm for geometrically consistent joint registration of three orthogonal heart views. Solid arrows represent registration and the corresponding correlation map. Dashed arrows represent splitting of the SA volume into two subvolumes in case the SA volume was acquired with two separate breath-holds. Figure 2Open in figure viewerPowerPoint Registration example. 2CH (a) and 4CH (b) views, shown in blue, are transformed into the coordinate system associated with the SA view, shown in orange. The blue and the orange add up to gray when the images are properly aligned. Even though strong breathing motion is observed, locality of our registration guarantees that the area around the heart is well-registered, as indicated by the predominantly gray color in that region. (c) Gradient map of the SA view together with the overlap region of all three views (R) and the region used for the actual registration (H). One slice from a 3D image stack is shown. High anisotropy of the data and small typical sizes of the volume of interest result in relatively broad correlation peaks, which hampers reliable estimation of the most probable displacement between the two volumes. To increase robustness of the interscan motion correction, we propose using the aggregated information from all three correlation maps. More precisely, we search the best triple of the displacement vectors as the one that maximizes the sum of the corresponding gradient-correlation maps: (1)with an additional constraint that the vector addition rule: (2)has to be satisfied (see Fig. 1). Usage of such constraint ensures geometrical consistency of the registration. All the gradient-correlation maps are scaled to the interval [0;1] beforehand. The set is calculated by performing a global search, where for each correlation map GC(x) only the voxels with an intensity higher than the 99.9 percentile were considered to decrease the computational complexity. A typical registration result is shown in Fig. 2. Slice Shift Correction in SA Volume Some of the SA volumes are acquired as two separate subvolumes, SA(1) and SA(2), each within its own breath-hold. This may result in misalignment not only with respect to the other two views, but also between the two SA subvolumes, referred to as a slice shift. For correction of this slice shift, we employ the 2CH and the 4CH volumes in a similar manner to the one described in the previous section. More precisely, we independently register both SA subvolumes to the other two views. In this case, the set of volume pairs contains five elements: and the corresponding set of optimal displacement vectors is determined by solving the problem in Eq. 1 with the following equivalent of the geometrical constraints given by Eq. 2: (3)The slice shift between SA(1) and SA(2) is calculated from the estimated displacements: Subsequently, both SA(1) and SA(2) are passed to the reconstruction routine as separate volumes. A nonzero component of the estimated displacement in the slice direction z results either in an undesired gap in the middle of the image stack, or the need to average the SA subvolumes in the overlap area. To prevent such cases, we restrict the optimization [1] to in-plane slice shifts by adding two additional constraints: Orthogonal Super-Resolution Reconstruction The high-resolution images were reconstructed from the three orthogonal views with a method extended from 24. The method iteratively improved the alignment and high-resolution reconstruction and corrected for global intensity differences between the N acquired images. The reconstruction was obtained with: (4)Here I is the high-resolution image being reconstructed, μ contains the parameters that specify the alignment of the images, the vector b contains the intensity correction factors, the i-th acquired image, w the 3D sinc point spread function of the acquisition, T the affine transformation model, λ a regularization scale factor, and Δ is the Laplacian operator. The images I and Vi are accessed (only) at the discrete sample locations y and x, respectively. The optimization is performed by alternately optimizing with respect to I and (μ,b) with iterative methods, and employing 24 for efficient application of the transformation and point spread function. In the optimization of (μ,b), the sum over x is restricted to a region in which all source images overlap. To improve the convergence, an initial smooth reconstruction is obtained by choosing a high value of λ = 5. Subsequently, this initial estimate was iteratively refined with decreasing values of λ until the final reconstruction was obtained. The final reconstruction used a value of λ = 0.13/2 that was visually perceived to be optimal. This value was the same for all test datasets. Scar Identification Scar identification included manual tracing of the LV endocardial and epicardial contours on the short-axis MRI slices and automatic identification of the hyperenhanced myocardial scar. The MASS software package (Research Version 2014; LKEB; Leiden University Medical Center, The Netherlands) was used for manual contour tracking by defining a number of control points on the myocardial border. The contour was traced by an experienced observer blinded to the catheter measurements and ablation results. For myocardial scar identification and further differentiation of scar core and gray zones, we used a validated method 25, which identifies the myocardial scar region using the Otsu threshold to separate the delayed enhanced signal from normal signal. Within this region, the scar core zone and gray zone were further differentiated by the full-width-half-maximum method that has been validated on histological data 26. The scar sizes were quantified in grams, approximated with a density of 1.05 g/ml. Validation With Electroanatomical Mapping All patients underwent catheter ablation of ventricular tachycardia. During the procedure, electroanatomical voltage mapping (EAVM) was performed on the endocardial surface to measure the electrophysiological (EP) characteristics in the myocardial scar region. EAVM was performed using the CARTO XP system (Biosense Webster, Diamond Bar, CA) and an irrigated-tip mapping catheter (NaviStar ThermoCooled, Biosense Webster) during sinus rhythm. The amplitude of the bipolar voltage reflects the local in vivo EP activities. Previous clinical studies have established the empirical voltage thresholds for myocardial tissue characterization 27: normal myocardium v ≥ 1.5 mV, scar gray zone 0.5 mV < v < 1.5 mV, and scar core zone v ≤ 0.5 mV. In the absence of histological data, the EAVM measurement provides the best available gold standard for characterizing the underlying myocardial tissue composition. Therefore, we validated the results by evaluating the agreement between the scar identification results and corresponding EAVM bipolar voltage range. Registration between the EAVM and MRI data was performed by minimizing the point-to-surface distance under the assumption of a fixed heart orientation 28, given that the patients stayed in supine position during all acquisitions. After registration, the acquired bipolar voltages were labeled according to their mapped location on the endocardial surface. If there is >80% normal healthy tissue within the 5 mm-radius region, the mapping location is classified as normal tissue; if there is >80% gray zone within the 1 cm-diameter region, the mapping location is classified as gray zone; if there is >80% core zone within the 1 cm-diameter region, the mapping location is classified as core zone. The 5-mm radius region is the approximate capturing range of the catheter tip 29, while 80% is an empirical threshold for zone evaluation 30. Bipolar voltages of the three types of tissue were obtained by pooling the voltage values on all classified mapping locations. Experiment Settings To validate our approach, we designed a number of quantitative experiments targeting different quality aspects of the reconstructed data. In particular, the reconstructed volume using the proposed method ("SRR") was compared to 1) the corresponding SA volume ("Interpolation"), interpolated by cubic spline to the same resolution as SRR; 2) the average of the three co-registered views ("Averaging"), also interpolated to the same resolution; 3) SRR reconstruction using two views only, "SRR (SA+2CH)" and "SRR (SA+4CH)". These volumes were compared and evaluated in terms of the estimated size of core and gray zones. The agreement between the identified myocardial scar core and gray zone with EAVM was evaluated by comparing the range of measured local voltages with clinically established empirical range. The paired Wilcoxon signed rank test with the significance threshold P = 0.05 was used to test for the difference of the scar size and voltage distributions. For the latter, we used the unpaired test since the number of the classified points at each zone can differ among volumes. Our method was implemented in MATLAB R2012b (MathWorks, Natick, MA) and partially in C/C++, and executed on a 3.60 GHz Intel(R) Xeon(R) computer with 32 GB RAM. RESULTS Scar Size Quantification Results of the size quantification of the core and the gray zones in grams are reported in Table 1. Compared to using the SA volume alone ("Interpolation"), significantly reduced gray zone sizes were identified from the SRR volume: 9.2 ± 8.1 g vs. 10.5 ± 8.8 g, P < 0.05, while the core zone sizes remained similar: 9.0 ± 8.7 g vs. 9.6 ± 8.6 g, P = 0.4 in accordance with previous animal experiment results in Schelbert et al 5. Compared to averaging of the three orthogonal views, both zones identified from SRR were smaller, suggesting less PVE blurring. It is observed that the results from the SRR of three volumes also differed from that of the SRR of two volumes, SRR (SA+2CH) and SRR (SA+4CH), with the generally smaller identified size. Table 1. Myocardial Scar Size Quantification Results (in grams)a Core zone Gray zone SRR 9.0 ± 8.7 9.2 ± 8.1 Interpolation 9.6 ± 8.6 10.5 ± 8.8* Averaging 9.8 ± 9.1* 10.1 ± 8.3* SRR (SA+2CH) 9.8 ± 8.9* 9.7 ± 8.2 SRR (SA+4CH) 9.2 ± 8.1 9.9 ± 8.0* a An asterisk indicates statistically significant difference with the SRR method with the significance level 0.05. In Fig. 3, the improvement of image quality is observed in the two through-plane views, as indicated by the red arrows. This figure also illustrates how our registration algorithm corrects the slice shift between two SA stacks. Figure 3Open in figure viewerPowerPoint A 3D reconstruction of the myocardial scar and one slice of each orthogonal view. The interpolated SA image (a) and the SRR reconstruction (b) for SA volumes acquired as two separate volumes. The locations of improvement in scar definition are indicated by the red arrows. The artifact due to slice shift between two separately acquired SA volumes is indicated by the blue arrows. Agreement With Electroanatomical Mapping Results of the bipolar voltage quantification are given in Table 2. No significant difference was observed between the bipolar voltages at the normal tissue from all five reconstructed volumes. However, we observed significant differences between the SRR and interpolated volumes in terms of the bipolar voltages at the scar core zone: 0.68 ± 0.98 mV vs. 0.87 ± 0.88 mV, P < 0.05, and gray zone: 0.99 ± 0.65 mV vs. 1.46 ± 1.15 mV, P < 0.05. The results imply that overestimation of scar gray zone indeed occurred with the interpolated SA volume, encompassing part of the healthy myocardium. The same holds for the averaged volume, with the voltages in the identified gray zone reaching a higher upper limit than that of SRR, although not statistically significant. Generally speaking, the difference between SRR and SRR (SA+2CH)/SRR (SA+4CH) was smaller than the difference between SRR and interpolation/averaging. Table 2. Bipolar Voltage Quantification Resultsa Bipolar voltage, mV Normal tissue Gray zone Core zone Empirical range v ≥ 1.5 0.5 < v < 1.5 v ≤ 0.5 SRR 3.67 ± 3.08 0.99 ± 0.65 0.68 ± 0.98 Interpolation 3.72 ± 3.10 1.46 ± 1.15* 0.87 ± 0.88* Averaging 3.66 ± 3.08 1.38 ± 2.16 1.01 ± 1.18* SRR (SA+2CH) 3.69 ± 3.09 1.39 ± 0.77* 0.87 ± 1.14 SRR (SA+4CH) 3.73 ± 3.09 1.08 ± 0.72 0.52 ± 0.46 a An asterisk indicates statistically significant difference with the SRR method with the significance level 0.05. Figure 4 presents an overview of the voltage distributions for comparison. Figure 5 shows two examples of 3D scar reconstruction for both SRR and interpolated MRI volumes, with the corresponding color-coded EAVM points superimposed; healthy myocardium is associated with purple points, gray zone—with blue to green points, and core zone—with red points. From this figure, better agreement of EAVM with SRR volume can be appreciated. Figure 4Open in figure viewerPowerPoint Boxplot of the bipolar voltages in identified normal tissue, scar gray zone and core zone. Dashed red lines indicate the empirical 0.5 mV and 1.5 mV threshold. Figure 5Open in figure viewerPowerPoint The 3D reconstruction of the myocardial scar of two patients, including core (orange) and gray (yellow) zones, superimposed on the EAVM data. The left column (a,c) shows two examples of interpolated SA, and the right column (b,d) shows the corresponding SRR reconstruction. The EAVM colors indicate: healthy myocardium (purple), core scar (red), and gray zone (intermediate colors). The red arrows point to locations of the most significant differences between the interpolated SA and SRR volumes in terms of characterization of the gray zone and the core zone. Execution Performance Registration of three orthogonal views took 0.20 ± 0.16 seconds. Initial smooth reconstruction and one refinement iteration took respectively 5.08 ± 0.87 minutes and 2.71 ± 0.48 minutes. The total execution time per reconstructed volume was 15.98 ± 2.5 minutes. DISCUSSION The balance between MRI image resolution, quality, and acquisition time is an important concern in clinical practice. Balancing all aspects is especially challenging for cardiac MRI, since the heart is a moving object with both local (contraction) and global (respiration) movement. To achieve high image quality within limited scan time, it has become a common practice to acquire the LGE MRI in three separate views, namely: SA, 2CH, and 4CH, each within one or two breath-holds. The details of the myocardial scar in 3D and its composition in terms of core and gray zone have important clinical implications for individualized catheter ablation or risk stratification. This implies that the clinicians who evaluate the LGE would preferably mentally combine the three acquisitions to reconstruct the 3D myocardial scar, to better appreciate the morphological and textural details. In addition, the different degree of PVE needs to be taken into account during this reconstruction, as the through-pla
Referência(s)