Magnetic Resonance in Medicine 64:586–594 (2010)
Automated Segmentation of Myocardial Scar in Late Enhancement MRI Using Combined Intensity and Spatial Information Qian Tao,1 Julien Milles,1 Katja Zeppenfeld,2 Hildo J. Lamb,3 Jeroen J. Bax,2 Johan H.C. Reiber,1 Rob J. van der Geest1 Accurate assessment of the size and distribution of a myocardial infarction (MI) from late gadolinium enhancement (LGE) MRI is of significant prognostic value for postinfarction patients. In this paper, an automatic MI identification method combining both intensity and spatial information is presented in a clear framework of (i) initialization, (ii) false acceptance removal, and (iii) false rejection removal. The method was validated on LGE MR images of 20 chronic postinfarction patients, using manually traced MI contours from two independent observers as reference. Good agreement was observed between automatic and manual MI identification. Validation results showed that the average Dice indices, which describe the percentage of overlap between two regions, were 0.83 ± 0.07 and 0.79 ± 0.08 between the automatic identification and the manual tracing from observer 1 and observer 2, and the errors in estimated infarct percentage were 0.0 ± 1.9% and 3.8 ± 4.7% compared with observer 1 and observer 2. The difference between the automatic method and manual tracing is in the order of interobserver variation. In conclusion, the developed automatic method is accurate and robust in MI delineation, providing an objective tool for quantitative assessment of MI in LGE MR imaging. Magn Reson Med 64:586–594, 2010. © 2010 Wiley-Liss, Inc. Key words: LGE MR; myocardium infarction; identification; false acceptance; false rejection
Late gadolinium enhancement (LGE) MRI is a valuable technique for assessment of myocardial infarction (MI) in postinfarction patients. LGE has been shown to be capable of identifying fibrosis, a mixture of fibrosis and viable myocyte, and the extent and size of infarction (1–4), all of which provide important prognostic indications for clinical studies and patient treatment. Therefore, it is of great interest to accurately quantify MI in LGE MR images. The MI sizing problem has been studied in several investigations (5–9). In earlier work, the identification of
1 LKEB – Division of Image Processing, Department of Radiology, Leiden University Medical Center, Leiden, The Netherlands 2 Department of Cardiology, Leiden University Medical Center, Leiden, The Netherlands 3 Department of Radiology, Leiden University Medical Center, Leiden, The Netherlands Grant sponsor: Netherlands Heart Foundation; Grant number: 2008B074; Grant sponsor: Netherlands Organization for Scientific Research (NWO Casimir Project); Grant number: 018.002.016. *Correspondence to: Qian Tao, Ph.D., LKEB – Division of Image Processing, Department of Radiology, Leiden University Medical Center, Leiden, The Netherlands. Email:
[email protected] Received 6 October 2009; revised 24 December 2009; accepted 8 February 2010. DOI 10.1002/mrm.22422 Published online in Wiley InterScience (www.interscience.wiley.com).
© 2010 Wiley-Liss, Inc.
the MI was mainly based on thresholding the signal intensities in LGE MR images, with the manually annotated remote healthy myocardium or necrosis region as reference. For example, the mean and N standard deviation (SD) (N varies from 1 to 6) of the remote normal myocardium (5,6), and full-width at half-maximum of the necrosis region (7), were used as thresholding criterion. More recent studies, however, went beyond empirical thresholds and tried to automate the identification process with pattern recognition and computer vision techniques. Hsu et al. (8) and Hennemuth et al. (9), for example, used implicit or explicit signal intensity models for the MI identification in a more comprehensive manner. In general, the difficulty of computer-aided MI identification method lies in the varied internal (e.g., size, distribution, heterogeneity of the infarction) and external situations (e.g., resolution, measurement noise, surface coil intensity variation, inversion time), which cannot be accounted for by a simple model. This study aims to address such problem in order to achieve automatic, accurate, and robust MI identification in LGE MR images. The method was validated on in vivo LGE MR images of postinfarction patients. MATERIALS AND METHODS Data Acquisition Twenty patients with known chronic MI (all male, mean age 64 ± 8 years, range 45−82 years) referred for viability assessment using LGE MR imaging were included. A 1.5-T 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. Patients were placed in the supine position. After obtaining the scout views, cine images of the heart were obtained from apex to base, with 12 to 15 imaging levels (dependent on heart size, one slice per breath hold) in the short-axis view, using a balanced turbo-field echo sequence with parallel imaging (SENSE: sensitivity encoding, acceleration factor 2). Typical parameters were as follows: field of view 400 × 320 mm2 ; matrix of 256 × 206 pixels; slice thickness, 10mm; no slice gap; flip angle (α) 50◦ ; echo time 1.67 ms; and pulse repetition time 3.3 ms. After cine MRI acquisition, T1 -weighted contrastenhanced images were acquired approximately 15 min after bolus injection of gadolinium diethylenetriamine pentaacetic acid (Magnevist; Schering, Berlin, Germany; 0.15 mmol/kg) with an inversion-recovery three-dimensional turbo-field echo sequence with parallel imaging
586
Automated Segmentation of Myocardial Scar
(SENSE, acceleration factor 2). Inversion time was determined with real-time plan scan to null normal myocardial signal. The heart was imaged in one breath hold, with 20 to 24 imaging levels (dependent on heart size) in the shortaxis view. Signal outside the field of view was suppressed (using two saturation slabs) to avoid foldover artifacts. Typical parameters were as follows: field of view, 400 × 400 mm2 ; matrix, 256 × 206 pixels; slice thickness, 10 mm with 5 mm overlap; α, 15◦ ; echo time 1.06 ms; and pulse repetition time 3.7 ms. One inversion pulse per heartbeat was used. Images were reconstructed to a matrix of 256 × 256 after zero-filling (scan matrix 256 × 160). Two independent observers evaluated the extent of MI in the LGE MR images. Observer 1 had full access to both the cine and LGE MR images. First, the cine images of the same cardiac phase as the LGE images were used by observer 1 to trace the epicardial and endocardial borders, which were then projected and fitted to the left ventricle in the LGE images. Subsequently, the regions of MI within the myocardium in LGE images were traced by the same observer, taking into consideration the wall motion abnormalities in the cine sequence as a reference of MI presence. Observer 2, however, was provided only with the LGE images, including the epicardial and endocardial borders from observer 1, and traced the MI regions based on this information. For the analysis, both observers used the MASS software package (research version; LKEB; Leiden University Medical Center, the Netherlands) (10). Algorithm Theoretical Density Model Prior to presenting our methods, a brief introduction of the theoretical normal myocardial tissue and MI intensity models is given. In theory, the intensity of homogeneous tissue in MR images in the presence of noise is shown to be governed by the Rician distribution (11), and with enough signal-to-noise ratio, the distribution reduces to gaussian distribution (11). For this reason, in the LGE MR images, the normal myocardium tissue is often modeled as Rician as the signal intensity of the normal myocardium tissue is low due to the rapid washing out of contrast agent, while the MI is often modeled as gaussian as the signal intensity of the MI is high due to hyperenhancement; hence, fairly
587
FIG. 1. Theoretical model of the T1 -intensity distributions of the normal myocardium tissue and the MI: Rician and gaussian. The vertical line illustrates one possible threshold. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]
high signal-to-noise ratio. The density model of the two tissues is illustrated in Fig. 1. Figure 1 shows the ideal intensity model of the normal myocardium and MI distribution. In realistic situations, the distributions may change according to particular patient situations, inversion time, acquisition protocols, and device specifications. For example, the distribution of the MI intensities will be dependent on the homogeneity and size of the MI, which varies from patient to patient. Robust Initialization The LGE MR images exhibit good contrast between the MI and the normal myocardium but often poor contrast between the MI and the blood pool (12). Figure 2a and b are two examples that demonstrate low contrasts between the blood pool and MI. In our work, this property is exploited by including the blood pool into the initial analysis and merging it into the same class as the MI, in contrast to the normal myocardium class. We argue that the blood pool provides useful reference both from a visual and computational point of view:
FIG. 2. Three example short-axis LGE MR images: (a) slice with lateral MI, (b) slice with full MI, (c) slice with no MI. Manually traced endocardial and epicardial contours are also shown. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]
588
Tao et al.
FIG. 3. Distributions of voxel intensities: (a) within the myocardium excluding the blood pool, per slice; (b) within the epicardial contour including the blood pool, per slice; (c) within the myocardium excluding the blood pool, per volume; (d) within the epicardial contour including the blood pool, per volume. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]
• First, the blood pool introduces a more stable reference than merely comparing the normal myocardium and MI, especially in cases when the volume of MI is small or nonexistent, illustrated by Fig. 2c. In such cases, it is difficult to determine upon the MI without external reference. Analysis that relies only on the myocardium might still falsely detect MI due to the intensity variation or noise in the normal myocardium. • Second, the blood pool, which occupies a large volume inside the left ventricle, contributes a large number of intensity samples, making it statistically more stable and robust for intensity-based analysis. To further increase the robustness of density analysis, the voxel intensities are taken from the entire myocardial volume, i.e., all the slices. This is particularly of interest in cases like Fig. 2b, where there is insufficient contrast between the two classes no matter whether the blood pool is included or not on a single slice level. To illustrate the effects of the proposed strategies, the distribution of the intensity values in each of the 20 slices from from apical to basal is shown in Fig. 3a and b for excluding and including the blood pool, respectively. For comparison, the distribution of the intensity values within the entire left ventricular volume is shown in Fig. 3c and d for excluding and including the blood pool, respectively. Obviously, the bimodality of the distribution is more pronounced when the blood pool is included either on slice or
volume level. Furthermore, the density functions become much more reliable and smooth when voxel intensities are taken from the entire volumetric region. We note that the three-dimensional initialization is an extension of the twodimensional situations. The method remains applicable in cases where fewer slices show hyperenhancement: individual histograms in Fig. 3b still show a more distinct bimodal distribution. However, detection can be hampered in situations where contrast is insufficient within the epicardial contour, as depicted in Fig. 2b. With such bimodal distribution, it is a natural thought to separate the two classes with a well-chosen threshold; for example, one in the valley of two distributions. For this purpose, the Otsu thresholding method (13) is used. Without assuming any specific underlying density models (i.e., no parametric form like Rician or gaussian), the Otsu method uses a general optimization criterion, similar to Fisher’s discriminant (14), to maximize the ratio of between-class 2 variance σb2 and within-class variance σw , both of which are functions of the threshold T Totsu = arg max T
σb2 (T ) 2 (T ) σw
[1]
This criterion can be further simplified (see (13)) and lead to fast and reliable estimation of T and therefore robust initialization of the MI identification.
Automated Segmentation of Myocardial Scar
589
FIG. 4. Illustration of connectivity filtering and region growing: (a) the original image, (b) the initially identified MI with the Otsu threshold, (c) the resulting MI after connectivity filtering and region growing. Arrows indicate the direction of growing. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]
Notwithstanding the accuracy and robustness of threshold determination, a method that relies on a threshold is subject to two types of classification errors: false acceptances (or false positive) and false rejections (or false negative). Errors originate from the overlapping of two density distributions, as illustrated in Fig. 1. Theoretically, they cannot be avoided by thresholding alone. In the following, we will address the two types of error respectively, using more information from the LGE MR images. For the initialization, we have used the intensity information, and for the refinement, spatial information will be exploited. False-Acceptance Removal by Connectivity Filtering Besides the false acceptance caused by scattering of normal tissue density distribution, the false acceptances in the LGE MR images are more often caused by small parts or thin layers of epicardial fat or blood pool, which also exhibit high signal intensities. Such false acceptance arises from small tracing errors of the endocardial and epicardial contours, as well as the partial-volume effect on anatomical borders. In addition to the signal intensity information, the connectivity of the initially identified voxels provides useful clues with regard to their likelihood of being MI. We used a two-pass algorithm (15) to study the connectivity of the initial MI. The idea of the two-pass algorithm is briefly described as follows: the first pass records connectivity and assigns temporary labels to the voxels, while the second pass iteratively replaces the temporary label with that of the connected voxels. The two-pass strategy leads to fast and efficient searching, even in large regions. For further details, see (15,16). Once a number of isolated regions are identified, analysis is done per region. The connectivity filtering criterion is designed based on physical constraints and applied in the following order: (i) the connected regions with too small volume (e.g., ≤ three voxels) are excluded, considered as noise; (ii) the connected regions attached to the epicardial or endocardial contours, thin layer in shape (criterion can be formulated as a limit on the ratio between the voxels connected to the contours and the entire volume, e.g., 0.05), but lower than those of observer 2, with a difference of 3.8 ± 4.7% (P < 0.05). We would like to stress that the interobserver variation is actually underestimated in this case as observer 2 used the same epicardial and endocardial contours traced by observer 1. The comparison, therefore, is more on the MI identification given pretraced endocardial and epicardial borders since both manual and automatic MI identification used the same input. DISCUSSION The purpose of this study was to design an automatic and robust MI identification method for LGE MR images of the postinfarction patients. Taking into account both internal and external variability of the MI in LGE images, the method aims to achieve segmentation in an objective manner, with minimal human intervention, and reduce the false acceptance and false rejection by maximally exploiting intensity and spatial information. Previous Methods In LGE MR images, the infarcted myocardial tissue is hyperenhanced due to accumulation of the contrast agent and can be differentiated from the healthy tissue (21). In the literature, therefore, thresholding the signal intensity in myocardium has been the most popular way for MI identification. The emphasis of those methods, however, has been different. In the paper by Yan et al. (3), the signal intensity of the remote healthy tissue was used as reference and the thresholding criterion was based on the mean
592
Tao et al.
annotation necessary. Second, no specific density model is assumed, and hence no potential overfitting introduced. Third, as shown in Fig. 3, by including the blood pool and populating within the entire myocardial volume, we maximize the bimodality and stabilize the threshold estimation. The third characteristic is especially of interest when the MI size is small, where the bimodality within the myocardium tends to disappear, as will be seen in Fig. 7(a).
Accuracy and Robustness of the Method
FIG. 6. Bland-Altman analysis of the %MI: (a) between observer 1 and observer 2, (b) between automatic identification and observer 1, (c) between automatic identification and observer 2. [Color figure can be viewed in the online issue, which is available at www.interscience. wiley.com.]
and SD of the remote region. Amado et al. (7), in comparison, used the MI region as reference and advocated the full-width at half-maximum criterion. Of importance, for both methods prior manual annotation of the reference region, either the remote area or the MI, is required. Using a different methodology, Hennemuth et al. (9) studied the distribution (histogram) of mixed normal and infarcted tissues. By assuming parametric underlying models of both tissue types, the relevant parameters were estimated using iterative expectation maximization, and the threshold was automatically derived using the watershed method. Likewise, the histogram was also analyzed by Hsu et al. (8), and the mean and SD of the normal tissue were estimated to calculate the threshold, using a criterion similar to that used by Yan et al. (3). Thresholding the signal intensity is used in our method, but only as an initialization step of MI identification. There are three characteristics in the way we derive it. First, the threshold determination is fully automatic, with no manual
From a classification point of view, thresholding two overlapping intensity distributions leads to false acceptances and false rejections. To remove them and refine the identification result, the spatial contextual information in MR images is utilized. The false acceptances, appearing as noise in the normal tissue, tiny patches or thin slices of epicardial fat, or blood pool, can be discriminated by their size, shape, and relative location within the left ventricle and eventually removed. As manual tracing of the endocardial and epicardial contours is not only tedious but also difficult due to the poor contrast among the blood pool, MI, and perimyocardial fat, it is highly desirable that the method be insensitive to those occasional errors caused by small mismatch of contour tracing. Due to the partial-volume effect, the regions with intermingling bundles of fibrotic and viable myocytes will exhibit lower-intensity values compared to those of the fibrotic region and cause false rejections. Referring to the density models in Fig. 1, they correspond to the left-end tail of the MI distribution, which is rejected by a brutal threshold. From a Markov random field point of view, those falsely rejected regions are likely to be connected to the detected MIs with higher conditional probability, while from a physiological point of view, they are also likely to be located at the MI borders, according to the wavefront phenomenon (22), which states that those regions further away from the infarct core contain less fibrotic tissue and hence a lower degree of enhancement. By carefully extending the initially localized MI, therefore, an explicit delineation of entire MI region can be achieved. The growing criterion is taken from the normal myocardium, based on the mean and SD of this region, as described in earlier work (3,18,19). In general, the statistics of the normal myocardium region are expected to be relatively stable compared to that of the MI region, which is subject to higher variability originating from the varying size, age, and heterogeneity of the infarction. The validation results showed that the automatic method produced accurate identification of the MI compared to the manual tracing from two independent observers. The difference between the automatic identification and the manual tracing is in the order of interobserver variation. A better agreement with observer 1, who had access to more patient information (i.e., cine, wall motion abnormalities), implies the benefit of a robust threshold that makes use of the global intensity information. In other words, the threshold is no longer slice specific but is more comprehensive and stable and therefore gives a more consistent prediction of the MI across slices. The robustness also
Automated Segmentation of Myocardial Scar
593
FIG. 7. Histogram revisited to demonstrate the difference between manually and automatically identified MI. Different regions are compared: manually traced MI regions from observer 1, observer 2, and automatically identified MI. Histograms of the myocardium (including normal myocardium and MI) and myocardium and blood pool are also plotted in the background. The Otsu threshold is shown as the dashed vertical line, which gives an initialization for the automatic identification method. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]
lies in the fact that our method does not rely on a number of algorithmic parameters. For initialization, the Otsu thresholding method is parameter free (13). For connectivity filtering, the criterion to remove false acceptance is purely physical (e.g., volume, shape) and easy to formulate without introducing sensitivity. The only parameter needed is the number of SD (N ) in region growing, which has been discussed in many previous studies (7,8,23). It has been demonstrated that the frequently suggested N = 2 readily produces good matching with manual tracing in our validation set. Furthermore, another merit of the computerized method, as shown in Fig. 5, is that it easily follows the MI boundaries, which can be exhausting and time consuming for observers to delineate. This property becomes increasingly valuable in analyzing a high-resolution threedimensional LGE study, where there are a large number of slices and manual tracing is even more laborious. Density Model Revisited In literature, it has been suggested that a simple intensity threshold is unreliable (3,7). In our work, a single uniform threshold is avoided and we could demonstrate that it is possible to discriminate voxels of the same intensity, either to be normal myocardium or MI, depending on their spatial prior. To illustrate this, we revisit the model presented in Fig. 1 and Fig. 3 by computing the histogram of the automatic identification and comparing it with that of the manual contouring. Histograms of the myocardium excluding and including the blood pool are also plotted in the background for reference. Figure 7 shows two examples from the validation set. The effect of the proposed method on false acceptance (seen as the histogram difference between the initialization and result) and false rejection (seen as the extension the left tail of the result MI beyond the Otsu threshold) can be observed in both examples. In Fig. 7a, a good match can be seen among all three
identification results, while in Fig. 7b, the relatively large MI region of observer 2 can be noticed. Study Limitations The method was validated on three-dimensional LGE MR data, and as has been pointed out, the automatic initialization may encounter problems in the single-slice situation where there is insufficient contrast within the epicardial contour. Furthermore, the method was evaluated on a relatively small number of patients with chronic MI. To test the full clinical potential of the method, in future work, elaborate evaluation of the method will need to be performed in a larger patient cohort, including patients with MI of nonischemic origin (e.g., MI with microvascular obstruction or at a subepicardial location). With such a population, MI distribution, morphology, and location can be different and, as a result, the identification criterion would need to be adapted. Another interesting future project is to study the potential of the method to automatically and objectively discriminate within the identified MI region the core zone and the gray zone, which can provide more in-depth clinical and diagnostic cues. CONCLUSION The paper presents a fully automatic MI identification method in LGE MR images of postinfarction patients. In the proposed framework that combines both intensity and spatial information, the common problems of MI identification, in the form of false acceptances and false rejections, are solved in a systematic and robust manner. Quantitative evaluation was carried out on patient data and demonstrated that the proposed automatic method is able to produce accurate MI delineation, providing an objective tool for quantitative assessment of MI in LGE MRI.
594
REFERENCES 1. Wu K, Zerhouni E, Judd R, Lugo-Olivieri C, Barouch L, Schulman S, Blumenthal R, Lima J. Prognostic significance of microvascular obstruction by magnetic resonance imaging in patients with acute myocardial infarction. Circulation 1998;97:765–772. 2. Kim R, Wu E, Rafael A, Chen E, Parker M, Simonetti O, Klocke F, Bonow R, Judd R. The use of contrast-enhanced magnetic resonance imaging to identify reversible myocardial dysfunction. N Engl J Med 2000;343:1445–1453. 3. Yan A, Shayne A, Brown K, Gupta S, Chan C, Luu T, Di Carli M, Reynolds H, Stevenson W, Kwong R. Characterization of the peri-infarct zone by contrast-enhanced cardiac magnetic resonance imaging is a powerful predictor of post-myocardial infarction mortality. Circulation 2006; 114:32–39. 4. Schmidt A, Azevedo C, Cheng A, Gupta S, Bluemke A, Foo T, Gerstenblith G, Weiss R, Marbán E, Tomaselli G, Lima J, Wu K. Infarct tissue heterogeneity by magnetic resonance imaging identifies enhanced cardiac arrhythmia susceptibility in patients with left ventricular dysfunction. Circulation 2008;115:2006–2014. 5. Lund G, Stork A, Bansmann M, Gerken J, Mester J, Higgins C, Adam G, Meinertz T. Acute myocardial infarction: evaluation with first-pass enhancement and delayed enhancement MR imaging compared with 201tl SPECT imaging. Radiology 2004;232:49–57. 6. Setser R, Bexell D, O’Donnell T, Stillman A, Lieber M, Schoenhagen P, White R. Quantitative assessment of myocardial scar in delayed enhancement magnetic resonance imaging. J Magn Reson Imaging 2003;18:434–441. 7. Amado L, Gerber B, Gupta S, Rettmann D, Szarf G, Schock R, Nasir K, Kraitchman D, Lima J. Accurate and objective infarct sizing by contrast-enhanced magnetic resonance imaging in a canine myocardial infarction model. J Am Coll Cardiol 2004;44:2383–2389. 8. Hsu L, Natanzon A, Kellman P, Hirsch G, Aletras AH, Arai AE. Quantitative myocardial infarction on delayed enhancement MRI, part I: animal validation of an automated feature analysis and combined thresholding infarct sizing algorithm. J Magn Reson Imaging 2006;23:298–308. 9. Hennemuth A, Seeger A, Friman O, Miller S, Klumpp B, Oeltze S. A comprehensive approach to the analysis of contrast enhanced cardiac MR images. IEEE Trans Med Imaging 2008;27:1592–1610.
Tao et al. 10. van der Geest R, Buller V, Jansen E, Lamb H, Baur L, van der Wall E, de Roos A, Reiber J. Comparison between manual and semiautomated analysis of left ventricular volume parameters from short-axis MR images. J Comput Assist Tomogr 1997;21:756–765. 11. Gudbjartsson H, Patz S. The Rician distribution of noisy MRI data. Magn Reson Med 1995;34:910–914. 12. Kellman P, Arai A. Multicontrast delayed enhancement provides improved contrast between myocardial infarction and blood pool. J Magn Reson Imaging 2005;22:605–613. 13. Otsu N. A threshold selection method from gray-level histograms. IEEE Trans Syst, Man Cybern 1979;9:62–66. 14. Duda R, Hart P, Stork D. Pattern Classification (2nd ed.). John Wiley and Sons, New York, 2001. 15. Shapiro L, Stockman G. Computer vision. Prentice Hall, Upper Saddle River, New Jersey, 2001. 16. Suzuki K, Horiba I, Sugie N. Linear-time connected-component labeling based on sequential local operations. Comput Vis Image Underst 2003;89:1–23. 17. Gonzalez R, Woods R. Digital image processing. Prentice Hall; 2002. 18. Klocke F, Wu E, Lee D. “Shades of gray” in cardiac magnetic resonance images of infarcted myocardium: can they tell us what we’d like them to? Circulation 2006;114:8–10. 19. Schuleri K, Centola M, George R, Amado L, Evers K, Kitagawa K, Vavere A, Evers R, Hare J, Cox C, McVeigh E, Lima J, Lardo A. Characterization of peri-infarct zone heterogeneity by contrast-enhanced multidetector computed tomography: a comparison with magnetic resonance imaging. J Am Coll Cardiol 2009;53:1699–1707. 20. Bland J, Altman D. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986;1:307–310. 21. Lima J, Judd R, Bazille A, Schulman S, Atalar E, Zerhouni E. Regional heterogeneity of human myocardial infarcts demonstrated by contrastenhanced MRI. Circulation 1995;92:1117–1125. 22. Reimer K, Lowe J, Rasmussen M, Jennings R. The wavefront phenomenon of ischemic cell death: myocardial infarct size vs. duration of coronary occlusion in dogs. Circulation 1977;56:786–794. 23. Kolipaka A, Chatzimavroudis G, White R, O’Donnell T, Setser R. Segmentation of non-viable myocardium in delayed enhancement magnetic resonance images. Int J Cardiovasc Imaging 2005;21:303–311.
