Int J Cardiovasc Imaging (2014) 30:925–934 DOI 10.1007/s10554-014-0402-3
ORIGINAL PAPER
Myocardial scar identification based on analysis of Look–Locker and 3D late gadolinium enhanced MRI Qian Tao • Hildo J. Lamb • Katja Zeppenfeld Rob J. van der Geest
•
Received: 3 December 2013 / Accepted: 12 March 2014 / Published online: 19 March 2014 Ó Springer Science+Business Media Dordrecht 2014
Abstract The aim of this study is to introduce and evaluate an approach for objective and reproducible scar identification from late gadolinium enhanced (LGE) MR by analysis of LGE data with post-contrast T1 mapping from a routinely acquired T1 scout Look–Locker (LL) sequence. In 90 post-infarction patients, a LL sequence was acquired prior to a three-dimensional LGE sequence covering the entire left ventricle. In 60/90 patients (training set), the T1 relaxation rates of remote myocardium and dense myocardial scar were linearly regressed to that of blood. The learned linear relationship was applied to 30/90 patients (validation set) to identify the remote myocardium and dense scar, and to normalize the LGE signal intensity to a range from 0 to 100 %. A 50 % threshold was applied to identify myocardial scar. In the validation set, two observers independently performed manual scar identification, annotated reference regions for the full-width-halfmaxima (FWHM) and standard deviation (SD) method, and analyzed the LL sequence for the proposed method. Compared with the manual, FWHM, and SD methods, the proposed method demonstrated the highest inter-class correlation coefficient (0.997) and Dice overlap index (98.7 ± 1.3 %) between the two observers. The proposed Q. Tao (&) R. J. van der Geest Division of Image Processing (LKEB), Department of Radiology, Leiden University Medical Center, PO Box 9600, 2300 RC Leiden, The Netherlands e-mail: [email protected] H. J. Lamb Department of Radiology, Leiden University Medical Center, Leiden, The Netherlands K. Zeppenfeld Department of Cardiology, Leiden University Medical Center, Leiden, The Netherlands
method also showed excellent agreement with the goldstandard manual scar identification, with a Dice index of 89.8 ± 7.5 and 90.2 ± 6.6 % for the two observers, respectively. Combined analysis of LL and LGE sequences leads to objective and reproducible myocardial scar identification in post-infarction patients. Keywords Late gadolinium enhanced MR Look– Locker sequence T1 map Myocardial scar identification
Introduction Accurate assessment of myocardial scar has important diagnostic and prognostic implications for post-infarction patients [1–4]. Over the last decades, late gadolinium enhanced (LGE) MR has become the gold standard for visualization of myocardial scar [5]. However, the lack of absolute measure in LGE makes it difficult to set an objective threshold for scar identification. In literature, patient-specific threshold values were used to identify the scar. Established methods include the mean ? N standard deviation (SD), with N ranging from 1 to 8 [1, 6, 7], full width at half maximum (FWHM) [8] and percentage maximal signal intensity [2, 3]. These semi-automatic methods require manual annotation of remote myocardium or dense scar regions. To reduce observer-independence, computer methods were also developed to automatically determine the threshold, based on probabilistic signal intensity models and spatial connectivity [9, 10]. However, the methods still relied on LGE alone, which has variable sequence parameters, contrast dose, timing of imaging, and acquisition noise. Myocardial T1 mapping is a MR modality which can provide quantification of the absolute T1 relaxation rate in
123
926
myocardial tissue [11]. T1 mapping can be realized by the Look–Locker (LL) sequence [12], and in recent years the modified LL inversion-recovery (MOLLI) sequence and shortened modified LL inversion recovery (ShMOLLI) sequence were developed to facilitate T1 mapping in the myocardium [13, 14]. However, in contrast to the 3D LGE sequence which covers the full heart, the T1 sequences are typically acquired at separate breathhold, with limited heart coverage. In most clinical MR protocols, the LGE scan is preceded by a LL sequence also called T1 scout, for determination of the optimal inversion time (TI) of the LGE acquisition. From this T1 scout sequence, a post-contrast T1 map can be derived. Nacif et al. has demonstrated that post-contrast T1 values derived from LL are in close agreement with T1 values derived from MOLLI sequences [15]. The purpose of this work is to utilize the routinely acquired single-slice LL sequence and three-dimensional LGE sequence, one with quantitative measurement of T1 and the other with excellent contrast and full heart coverage, for objective and reproducible myocardial scar identification in post-infarction patients.
Methods Subjects The total study population consisted of 90 consecutive post-infarction patients (age 64 ± 10 years, 73 male) with chronic myocardial infarction who underwent cardiac MRI between 2007 and 2012 at our center, and who had presence of dense myocardial scar at the location of the LL scan. The 90 patients were randomly divided into two groups, referred to as the training group and the validation group: the training group consisted of 60 patients (age 64 ± 9 years, 49 male), whose data were used to train the post-contrast blood-tissue T1 model; the validation group consisted of 30 patients (age 63 ± 11 years, 24 male), whose data were used for validation of the scar identification method. The high number of subjects in the training group is intended to learn a reliable post-contrast bloodtissue T1 relationship. The Dutch Central Committee on Human-related Research (CCMO) allows use of anonymous data without prior approval of an institutional review board provided that the data is acquired for patient care and that the data contains no identifiers that could be traced back to the individual patient. All data used for this study was acquired for clinical treatment, and was completely stripped of any identifying information.
123
Int J Cardiovasc Imaging (2014) 30:925–934
MR imaging protocol 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 12–15 imaging levels 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 9 320 mm2; matrix of 256 9 206 pixels; slice thickness 10 mm, no slice gap; flip angle 50°; echo time 1.67 ms; and pulse repetition time 3.3 ms. A LL sequence was acquired approximately 15 min after bolus injection of gadolinium DTPA (Magnevist; Schering, Berlin, Germany; 0.15 mmol/kg), at one midventricular short-axis level. Typical parameters were as follows: field of view 370 9 370 mm2; matrix of 128 9 77 pixels; slice thickness 10 mm; flip angle 8 °; echo time 1.8 ms; in total 33 phases (typical TI of the first phase 159 ms with phase interval of 15–18 ms); and pulse repetition time 3.7 ms. Immediately following the LL acquisition, T1-weighted LGE images were acquired with an inversion-recovery three-dimensional turbo-field echo sequence with parallel imaging (SENSE, acceleration factor 2). The TI was determined from the LL sequence to null the normal myocardial signal. The heart was imaged with 20–26 imaging levels in the short axis view in one or two threedimensional stacks, each stack acquired within one breathhold. Signal outside the field of view was suppressed using two saturation slabs to avoid fold-over artifacts. Typical parameters were as follows: field of view 400 9 400 mm2; matrix of 256 9 206 pixels; slice thickness 10 mm with 5 mm overlap; flip angle 15°; echo time 1.1 ms; and pulse repetition time 3.7 ms. One inversion pulse per heartbeat was used. Image processing For all 90 subjects, the endocardial and epicardial contours of the left ventricle (LV) in the short-axis LGE sequence were semi-automatically traced using the MASS Research software (version V013-EXP; Leiden University Medical Center, The Netherlands). The endocardial and epicardial contours of the LL sequences were semi-automatically defined using MASS with the following protocol: firstly, the observer manually traced the contours in three separate phases, namely, the end-diastolic, end-systolic, and earlydiastolic phases. Then the contours in the other phases were automatically generated by the software using linear
Int J Cardiovasc Imaging (2014) 30:925–934
927
A
229 ms
299 ms
369 ms
LGE slice
B
800
C
438 ms
400
500 400 300 200
200
648 ms
100 0
718 ms
T1 curves
400
T1 map (ms)
600
578 ms
508 ms
healthy myocardium dense scar
D signal intensity
159 ms
300
200 100
0
0
200
400
600
800
time (ms) Fig. 1 a The LL sequence. b The reference LGE slice with annotated remote myocardium (green) and dense scar (magenta) regions. c The computed T1 map from the LL sequence. d The averaged T1 curve for remote myocardium and dense scar
interpolation, resulting in a series of smoothly changing contours throughout the entire cardiac cycle. Subsequently, the T1 map of the myocardial tissue within endocardial and epicardial contours was computed in a pixel-wise manner: firstly the location of each pixel is defined by its radial (r) and circumferential (h) position relative to the contours, and then the pixels at different frames but with the same r and h were connected to construct a local T1 relaxation curve. The T1 value at this local pixel was computed by parametric fitting of this curve: B f ðTI Þ ¼ A B exp TI=T1 ; T1 ¼ ð1Þ T A 1 where f(TI) is the longitudinal magnetization recovery curve, parameterized by A, B and T*1, as shown in Fig. 1d. The T1 is derived from the three parameters, and the relaxation rate R1 is computed as 1/T1. The LGE slice at the same short-axis level as the LL sequence was computed using the spatial information from the DICOM header. This reconstructed LGE slice was referred to as the reference LGE slice. The image processing, data fitting, and statistical analysis were performed using the scientific computing tool MATLAB (version 2012a; MathWorks, Natick, MA). Learning the linear relationship between post-contrast tissue and blood relaxation rates The change of relaxation rate in myocardial tissue is correlated to the change of relaxation rate in blood, during the equilibrium phase (5–45 min post bolus injection) when the gadolinium clearance is compensated in tissue and blood [16]:
0
DR1tissue R1tissue R1tissue k¼ ¼ 0 DR1blood R1blood R1blood
ð2Þ
where k is the ratio between change, also called partition coefficient, R1tissue and R1blood denote the post-contrast T1 relaxation rates of myocardial tissue and blood, respectively, and R01tissue and R01blood denote their pre-contrast relaxation rates. Without pre-contrast measurements available, we utilized the following linear relationship: R1tissue ¼ ktissue R1blood þ xtissue
ð3Þ
where xtissue ¼ R01tissue ktissue R01blood . We can perform linear fitting on the R1 for remote myocardium and dense scar: R1myo ¼ kmyo R1blood þ xmyo ; R1scar ¼ kscar R1blood þ xscar
ð4Þ
where kmyo and kscar are the slopes, and xmyo and xscar are the intercepts of the linear fitting. Therefore, with the slope and intercept parameters learned, one can estimate the relaxation rates of remote myocardium and dense scar from the relaxation rates of the blood. Consequently, the remote myocardium and dense scar region can be identified from the T1 map. For each subject of the training group, two types of myocardial tissue were manually annotated on the LGE reference slice, as extremely dark or bright regions. An example is given in Fig. 1. The post-contrast relaxation rates of the remote myocardium and dense scar were computed as the averaged R1 in the annotated regions. The post-contrast relaxation rate of the blood was computed from the blood pool region excluding papillary muscles. From the scatter plot, the slope and intercept parameters were estimated by least square linear regression.
123
928
T1-based myocardial scar identification From the LL-derived T1 map, the remote myocardium and dense scar regions were identified: Firstly, the average relaxation rate of the blood was taken from the blood pool region. Then the relaxation rate of remote myocardium and dense scar was computed accordingly based on the learned relationship as expressed in Eq. 3. The remote myocardium regions with relaxation rates R1 R1myo and dense scar regions with R1 R1scar was automatically identified, respectively. Finally the two regions on the T1 map were mapped onto the reference LGE slice, providing the T1identified reference regions for remote myocardium and dense scar. Since the LGE volume is acquired by a 3D sequence with breathhold, there was no variation of intensity distribution expected across the 3D LGE volume. Therefore the reference regions from a single T1 map can be used to normalize the signal intensity of the entire 3D LGE volume. Based on the T1-identified reference regions, the LGE image was normalized as follows: (1) the regions with signal intensity equal to or lower than the average signal intensity of the T1-identificed remote myocardium region were assigned value 0 %; (2) The regions with signal intensity equal to or higher than the average signal intensity of the T1-identified dense scar region were assigned value 100 %; (3) The transitional regions were linearly mapped to the intermediate value from 0 to 100 %. The same normalization was applied to the entire three-dimensional LGE volume covering the full LV. A 50 % threshold was applied to identify the myocardial scar region. In addition, automated post-processing was applied to remove falsely accepted regions, e.g. small isolated noisy patches, by screening the size and location of the connected components [9]. Validation on reproducibility and accuracy In the validation group of 30 subjects, two observers manually segmented the myocardial scar region in the LGE images using the MASS software. Observer 1 has over 20 years’ experience and Observer 2 has over 5 years’ experience in analyzing cardiac MR images. The protocol was as follows: Firstly, both observers reviewed short-axis, long-axis, four-chamber LGE, and cine images. Secondly, both observers segmented the scar region by setting signal intensity thresholds per slice based on visual interpretation of enhancement and wall motion abnormality. Finally, the observers manually excluded those regions with an intensity value above the threshold but not considered as scar. To evaluate the reproducibility of the proposed T1-based method, both observers also manually analyzed the LL
123
Int J Cardiovasc Imaging (2014) 30:925–934
sequences using the protocol previously describe for all 30 subjects in the validation group. We further compared our method with two established scar identification methods: (1) The FWHM method, which uses 50 % of the maximum signal intensity in the annotated dense scar region as the threshold. (2) The SD method, which uses N SD above the mean value in the annotated remote myocardium region as the threshold. For our dataset, N = 6 was taken as an empirical value which agreed best with visual scar assessment. For comparison with FWHM and SD methods, the two observers also manually annotated the dense scar and remote myocardium regions. The scar size was quantified in grams, estimated by the density of 1.05 g/m. The agreement between two identification results were quantified by the Dice index, computed as 2 |V1\V2|/(|V1| ? |V2|), describing the percentage of overlap between two scar volumes V1 and V2 in three dimension. Statistical analysis Continuous variables were expressed as mean ± SD. The paired t test was performed to determine whether there exist significant differences between variables. A p value \0.05 was considered significant. Bland–Altman analysis was used to evaluate inter-observer variability. The interclass correlation coefficient (ICC) was computed to evaluate the reproducibility of the results from two observers.
Results Training In the training group of 60 subjects, the T1 value was 300.6 ± 39.7 ms for the blood pool, 344.9 ± 43.8 ms for the annotated remote myocardium, and 282.8 ± 43.7 ms for the annotated dense scar. The linear regression between R1myo and blood R1blood is shown in Fig. 2a, and the linear regression between R1scar and R1blood is shown in Fig. 2b. The Pearson’s correlation coefficients of R1myo and R1scar with R1blood were 0.70 and 0.84, respectively, with p \ 10-10 for both. The estimated kmyo was 0.56 and kscar was 1.03. Validation The validation included two parts: (1) the inter-observer variability and (2) the agreement with gold-standard. In absence of the histological data, the manual results from Observer 1, who has over 20 years’ experience in
Int J Cardiovasc Imaging (2014) 30:925–934
929
Fig. 2 a The linear regression between relaxation rates of remote myocardium and blood. b The linear regression between the relaxation rates of dense scar and blood
analyzing cardiac MR images, was considered as the gold standard in scar identification. Figure 3 shows the Bland–Altman plots of the results from the two observers for all four methods. Table 1 reported the inter-observer variability, including the identified scar size, Dice index, p values from the paired t test, and ICC coefficient. It is shown that the ICC is high for all four methods ([0.9). However, for the manual and FWHM methods, a bias existed between the two observers, with Observer 2 estimating the scar size slightly higher than Observer 1 (p \ 0.05). The SD methods yielded no statistical bias between the observers, but with relatively large variation in their differences. The T1-based method showed the highest ICC (0.997) and Dice index (98.7 ± 1.3 %), and with lowest inter-observer difference (0.0 ± 1.6 ml). Figure 4 shows the results from Observer 2 in comparison to the gold standard manual scar identification. It is can be seen from the Bland–Altman plot that the T1-based method has the least variation across the scar size. Table 2 compares the results of the three methods from both Observer 1 and 2 with the gold standard. No significant differences in scar sizing were observed for all three methods (p [ 0.05). For both observers, the T1-based method produced the highest Dice index with the gold standard. Figure 5 displays three examples of the scar identification results from all four methods: T1-based, manual, FWHM, and SD. In Fig. 5b, the intermediate results, i.e., the T1-normalized LGE image, is also shown.
instead of quantitative measurement. In comparison, the T1 mapping is able to provide quantitative measurement of tissue relaxation rates, but has limited LV coverage. This work aims to combine the advantage of both sequences for objective and reproducible myocardial scar identification. Compared to manually annotating the remote myocardium and dense scar regions in LGE, the proposed method is able to automatically identified the two reference regions based on combined analysis of T1 map and LGE, avoiding visual interpretation. In this work, we utilized the linear relationship between R1myo and R1scar to R1blood , which can be learned from a training set. In Fig. 2, it shows that the absolute relaxation rates of remote myocardium and dense scar do vary substantially among subjects, but a linear relationship to the relaxation rates of blood holds in both cases (p \ 10-10). In [17], a similar principle was used to correct T1 of tissue with T1 of blood for more stable measurement of fibrosis in the myocardium. Compared with what has been reported in literature [18], the partition coefficient of the myocardium kmyo obtained from our training data was identical (0.56), but that of the dense scar kscar was higher (1.03 vs. 0.78). The increase in kscar may be explained by the fact that our T1 map was at pixel level, while in [18] the T1 map was computed per radial sector of 60°, encompassing both dense scar and viable tissue.
Comparison with FWHM and SD method Discussion This study demonstrated that analysis of T1 and LGE sequences can lead to reproducible and accurate myocardial scar identification in post-infarction patients. Combining information from T1 and LGE LGE is able to delineate myocardial scar in fine resolution covering the whole heart, but it is based on relative contrast
We have compared our method with the two established methods: FWHM and SD, and evaluated the reproducibility between two observers, and the agreement with gold standard. Our study demonstrated that the T1-based method has the highest ICC and Dice index between two observers. The high reproducibility is explained by the fact that the two reference regions were objectively identified from the T1 values, instead of manually. The reproducibility of the SD method was the lowest since the SD value within the annotated remote myocardium largely depends on its
123
930
A 50
difference scar size (g)
100
manual Observer 2 (g)
Fig. 3 Comparison of the myocardial scar identification results between 2 observers: a manual, b FWHM, c SD, D. T1-based
Int J Cardiovasc Imaging (2014) 30:925–934
80 60 40 20 0
0
50
0
-50
100
0
manual Observer 1 (g)
B
difference scar size (g)
60 40 20
0
50
difference: 2.6±4.2
0
-50
100
0
FWHM Observer 1 (g)
80 60 40 20
0
0
-50
100
0
50
100
average scar size (g)
100
50
80 60 40 20
1
T -based Observer 2 (g)
50
difference: 0.7±11.1
STD Observer 1 (g)
D
0 0
50
100
T1-based Observer 1 (g)
location as well as size. For different dataset, N needs to be carefully tuned to the MRI data quality as SD describes the noise level. In our dataset, N = 6 leads to the best agreement with manual scar identification, as also suggested in a previous study [6].
123
100
50
difference scar size (g)
100
0
50
average scar size (g)
difference scar size (g)
STD Observer 1 (g)
C
100
50
80
0
50
average scar size (g)
100
FWHM Observer 2 (g)
difference: 3.2±3.9
difference: 0.0±1.6
0
-50
0
50
100
average scar size (g)
Comparison with the gold-standard manual scar identification showed that the T1-based method has the least variation across scar size, compared with the FWHM and SD methods. This can be explained by the fact that in the FWHM and SD methods, the variability of manual
Int J Cardiovasc Imaging (2014) 30:925–934 Table 1 Inter-observer comparison of four methods: manual, FWHM, SD, and T1based
931
Method Manual
Scar size (g) Observer 2
Dice index (%)
p value
ICC
94.0 ± 4.6
\0.001
0.977
94.6 ± 5.2
0.002
0.979
88.9 ± 10.1
0.73
0.924
98.7 ± 1.3
0.99
0.997
44.9 ± 23.4
48.1 ± 23.5
(5.2–117.8)
(7.2–116.5)
FWHM
45.2 ± 23.4
47.8 ± 23.9
(6.5–112.4)
(5.6–111.8)
SD
45.4 ± 27.8
46.1 ± 28.3
(6.6–147.5)
(8.3–144.5)
44.6 ± 21.9
44.6 ± 22.1
(5.3–124.5)
(5.5–124.5)
T1-based
100
50
difference scar size (g)
A FWHM Observer 2 (g)
Fig. 4 Comparison of the myocardial scar identification results to the gold standard: a FWHM, b SD, c T1-based
Scar size (g) Observer 1
80 60 40 20 0
0
50
difference: 0.3±11.2
0
-50
100
0
manual Observer 1 (g) 100 80 60 40 20 0
0
50
difference: 0.5±13.0
0
-50
100
0
manual Observer 1 (g) 100
100
50
80 60 40 20 0
50
average scar size (g)
difference scar size (g)
1
T -based Observer 2 (g)
C
100
50
difference scar size (g)
STD Observer 2 (g)
B
50
average scar size (g)
50
0
100
manual Observer 1 (g)
annotation may increase with the scar size, while in the T1based method the determination of the reference regions is not dependent on visual interpretation. In our method, a similar rationale as in the FWHM method, i.e. applying a 50 % threshold, was used to
difference: -0.2±7.6
0
-50
0
50
100
average scar size (g)
identify the scar. The difference however lies in the way to identify the reference regions: the FWHM method uses a single maximum signal intensity in the dense scar region as reference, while our method uses two T1-identified reference regions including both dense scar and remote
123
932
Int J Cardiovasc Imaging (2014) 30:925–934
Table 2 Comparison of the FWHM, SD, and T1-based methods to the gold-standard manual scar identification Method
Observer 1 Dice index (%)
Observer 2 p value
Dice index (%)
p value
FWHM
87.3 ± 9.4
0.87
87.4 ± 10.3
0.19
SD
85.2 ± 8.3
0.94
85.3 ± 8.5
0.63
T1-based
89.8 ± 7.5
0.86
90.2 ± 6.6
0.86
In FWHM and SD methods, one or two reference regions needs to be manually identified. In comparison, the proposed method automatically identifies such reference regions from the T1 map based on the blood-tissue T1 relationship, thereby avoiding any observer bias due to visual interpretation of the signal enhancement in LGE images. Clinical applicability
myocardium. Our method therefore accommodates imperfect myocardium nulling (i.e. when the signal intensity of remote myocardium is also elevated).
Fig. 5 Examples of myocardial scar identification results: a The original LGE images. b T1normalized LGE images. c T1based scar identification results. b Manual scar identification results. e FWHM scar identification results. f SD scar identification results
123
The proposed method is compatible with standard LGE protocols, in which a single-slice LL and a three-dimensional LGE are acquired sequentially. This implies that no
Int J Cardiovasc Imaging (2014) 30:925–934
extra scanning is needed and the method can be applied to standard LGE dataset. To utilize the available T1 information from the T1 scout sequence, we analyzed the LL sequences in a semi-automated manner. As described in the protocol, the contouring of the LL sequence is relatively easy and fast: only three cardiac phases need to be analyzed, and the rest of contours are generated automatically. For trained observers, the process takes \3 min per LL scan. Potentially, the manual contouring can be replaced by computer methods that exploit the mutual information to automatically link the individual pixels in the LL sequence [19, 20].
933
3.
4.
5.
6.
Limitations 7.
A major limitation in our study is that we do not have histological data as the gold standard. The validation of the results was based on comparison with the manual results from an experienced observer, and in addition the evaluation of reproducibility between two observers. Another limitation is that the method requires that the single-slice T1 map cover both the remote myocardium and dense scar in order to get T1-based reference regions. The condition may not always be satisfied. Alternatively, if the LL is acquired at a location without dense scar, the proposed method may potentially provide a reference region for remote myocardium, which can serve as input to the SD method.
8.
9.
10.
11.
Conclusions 12.
We have proposed a myocardial scar identification method based on analysis of routinely acquired LL and LGE sequences. The proposed method demonstrated high scar identification accuracy compared with the gold standard manual identification, with excellent inter-observer reproducibility.
13.
14.
Acknowledgments The authors would like to acknowledge the financial support from the EU MEDIATE Project (ITEA2-09039). Conflict of interest
None.
15.
16.
References 1. Yan AT, Shayne AJ, Brown KA et al (2006) Characterization of the peri-infarct zone by contrast-enhanced cardiac magnetic resonance imaging is a powerful predictor of post-myocardial infarction mortality. Circulation 114:32–39. doi:10.1161/CIRCU LATIONAHA.106.613414 2. Schmidt A, Azevedo CF, Cheng A et al (2007) Infarct tissue heterogeneity by magnetic resonance imaging identifies enhanced cardiac arrhythmia susceptibility in patients with left ventricular
17.
18.
dysfunction. Circulation 115:2006–2014. doi:10.1161/CIRCU LATIONAHA.106.653568 Roes SD, Borleffs CJW, van der Geest RJ et al (2009) Infarct tissue heterogeneity assessed with contrast-enhanced MRI predicts spontaneous ventricular arrhythmia in patients with ischemic cardiomyopathy and implantable cardioverter-defibrillator. Circ Cardiovasc Imaging 2:183–190. doi:10.1161/CIRCIMA GING.108.826529 Bello D, Fieno DS, Kim RJ et al (2005) Infarct morphology identifies patients with substrate for sustained ventricular tachycardia. J Am Coll Cardiol 45:1104–1108. doi:10.1016/j.jacc.2004.12.057 Kim RJ, Fieno DS, Parrish TB et al (1999) Relationship of MRI delayed contrast enhancement to irreversible injury, infarct age, and contractile function. Circulation 100:1992–2002 Harrigan CJ, Peters DC, Gibson CM et al (2011) Hypertrophic cardiomyopathy: quantification of late gadolinium enhancement with contrast-enhanced cardiovascular MR imaging. Radiology 258:128–133. doi:10.1148/radiol.10090526 Beek AM, Bondarenko O, Afsharzada F, van Rossum AC (2009) Quantification of late gadolinium enhanced CMR in viability assessment in chronic ischemic heart disease: a comparison to functional outcome. J Cardiovasc Magn Reson 11:6. doi:10.1186/ 1532-429X-11-6 Amado LC, Gerber BL, Gupta SN et al (2004) Accurate and objective infarct sizing by contrast-enhanced magnetic resonance imaging in a canine myocardial infarction model. J Am Coll Cardiol 44:2383–2389. doi:10.1016/j.jacc.2004.09.020 Tao Q, Milles J, Zeppenfeld K et al (2010) Automated segmentation of myocardial scar in late enhancement MRI using combined intensity and spatial information. Magn Reson Med 64:586–594. doi:10.1002/mrm.22422 Hsu L-Y, Ingkanisorn WP, Kellman P et al (2006) Quantitative myocardial infarction on delayed enhancement MRI. Part II: clinical application of an automated feature analysis and combined thresholding infarct sizing algorithm. J Magn Reson Imaging 23:309–314. doi:10.1002/jmri.20495 Mewton N, Liu CY, Croisille P et al (2011) Assessment of myocardial fibrosis with cardiovascular magnetic resonance. J Am Coll Cardiol 57:891–903. doi:10.1016/j.jacc.2010.11.013 Look DC, Locker DR (1970) Time saving in measurement of NMR and EPR relaxation times. Rev Sci Instrum 41:250–251. doi:10.1063/1.1684482 Messroghli DR, Radjenovic A, Kozerke S et al (2004) Modified look-locker inversion recovery (MOLLI) for high-resolution T1 mapping of the heart. Magn Reson Med 52:141–146. doi:10. 1002/mrm.20110 Piechnik SK, Ferreira VM, Dall’Armellina E et al (2010) Shortened modified look-locker inversion recovery (ShMOLLI) for clinical myocardial T1-mapping at 1.5 and 3 T within a 9 heartbeat breathhold. J Cardiovasc Magn Reson 12:69. doi:10. 1186/1532-429X-12-69 Nacif MS, Turkbey EB, Gai N et al (2011) Myocardial T1 mapping with MRI: comparison of look-locker and MOLLI sequences. J Magn Reson Imaging 34:1367–1373. doi:10.1002/jmri.22753 Wendland MF, Saeed M, Lauerma K et al (1997) Alterations in T1 of normal and reperfused infarcted myocardium after GdBOPTA versus GD-DTPA on inversion recovery EPI. Magn Reson Med 37:448–456 Choi E-Y, Hwang SH, Yoon YW et al (2013) Correction with blood T1 is essential when measuring post-contrast myocardial T1 value in patients with acute myocardial infarction. J Cardiovasc Magn Reson 15:11. doi:10.1186/1532-429X-15-11 Flacke SJ, Fischer SE, Lorenz CH (2001) Measurement of the gadopentetate dimeglumine partition coefficient in human myocardium in vivo: normal distribution and elevation in acute and chronic infarction. Radiology 218:703–710
123
934 19. Xue H, Shah S, Greiser A et al (2012) Motion correction for myocardial T1 mapping using image registration with synthetic image estimation. Magn Reson Med 67:1644–1655. doi:10.1002/ mrm.23153
123
Int J Cardiovasc Imaging (2014) 30:925–934 20. Van de Giessen M, Tao Q, van der Geest RJ, Lelieveldt BPF (2013) Model-based alignment of look-locker mri sequences for calibrated myocardical scar tissue quantification. International Symposium on Biomedical Imaging (ISBI): from nano to macro
ORIGINAL PAPER
Myocardial scar identification based on analysis of Look–Locker and 3D late gadolinium enhanced MRI Qian Tao • Hildo J. Lamb • Katja Zeppenfeld Rob J. van der Geest
•
Received: 3 December 2013 / Accepted: 12 March 2014 / Published online: 19 March 2014 Ó Springer Science+Business Media Dordrecht 2014
Abstract The aim of this study is to introduce and evaluate an approach for objective and reproducible scar identification from late gadolinium enhanced (LGE) MR by analysis of LGE data with post-contrast T1 mapping from a routinely acquired T1 scout Look–Locker (LL) sequence. In 90 post-infarction patients, a LL sequence was acquired prior to a three-dimensional LGE sequence covering the entire left ventricle. In 60/90 patients (training set), the T1 relaxation rates of remote myocardium and dense myocardial scar were linearly regressed to that of blood. The learned linear relationship was applied to 30/90 patients (validation set) to identify the remote myocardium and dense scar, and to normalize the LGE signal intensity to a range from 0 to 100 %. A 50 % threshold was applied to identify myocardial scar. In the validation set, two observers independently performed manual scar identification, annotated reference regions for the full-width-halfmaxima (FWHM) and standard deviation (SD) method, and analyzed the LL sequence for the proposed method. Compared with the manual, FWHM, and SD methods, the proposed method demonstrated the highest inter-class correlation coefficient (0.997) and Dice overlap index (98.7 ± 1.3 %) between the two observers. The proposed Q. Tao (&) R. J. van der Geest Division of Image Processing (LKEB), Department of Radiology, Leiden University Medical Center, PO Box 9600, 2300 RC Leiden, The Netherlands e-mail: [email protected] H. J. Lamb Department of Radiology, Leiden University Medical Center, Leiden, The Netherlands K. Zeppenfeld Department of Cardiology, Leiden University Medical Center, Leiden, The Netherlands
method also showed excellent agreement with the goldstandard manual scar identification, with a Dice index of 89.8 ± 7.5 and 90.2 ± 6.6 % for the two observers, respectively. Combined analysis of LL and LGE sequences leads to objective and reproducible myocardial scar identification in post-infarction patients. Keywords Late gadolinium enhanced MR Look– Locker sequence T1 map Myocardial scar identification
Introduction Accurate assessment of myocardial scar has important diagnostic and prognostic implications for post-infarction patients [1–4]. Over the last decades, late gadolinium enhanced (LGE) MR has become the gold standard for visualization of myocardial scar [5]. However, the lack of absolute measure in LGE makes it difficult to set an objective threshold for scar identification. In literature, patient-specific threshold values were used to identify the scar. Established methods include the mean ? N standard deviation (SD), with N ranging from 1 to 8 [1, 6, 7], full width at half maximum (FWHM) [8] and percentage maximal signal intensity [2, 3]. These semi-automatic methods require manual annotation of remote myocardium or dense scar regions. To reduce observer-independence, computer methods were also developed to automatically determine the threshold, based on probabilistic signal intensity models and spatial connectivity [9, 10]. However, the methods still relied on LGE alone, which has variable sequence parameters, contrast dose, timing of imaging, and acquisition noise. Myocardial T1 mapping is a MR modality which can provide quantification of the absolute T1 relaxation rate in
123
926
myocardial tissue [11]. T1 mapping can be realized by the Look–Locker (LL) sequence [12], and in recent years the modified LL inversion-recovery (MOLLI) sequence and shortened modified LL inversion recovery (ShMOLLI) sequence were developed to facilitate T1 mapping in the myocardium [13, 14]. However, in contrast to the 3D LGE sequence which covers the full heart, the T1 sequences are typically acquired at separate breathhold, with limited heart coverage. In most clinical MR protocols, the LGE scan is preceded by a LL sequence also called T1 scout, for determination of the optimal inversion time (TI) of the LGE acquisition. From this T1 scout sequence, a post-contrast T1 map can be derived. Nacif et al. has demonstrated that post-contrast T1 values derived from LL are in close agreement with T1 values derived from MOLLI sequences [15]. The purpose of this work is to utilize the routinely acquired single-slice LL sequence and three-dimensional LGE sequence, one with quantitative measurement of T1 and the other with excellent contrast and full heart coverage, for objective and reproducible myocardial scar identification in post-infarction patients.
Methods Subjects The total study population consisted of 90 consecutive post-infarction patients (age 64 ± 10 years, 73 male) with chronic myocardial infarction who underwent cardiac MRI between 2007 and 2012 at our center, and who had presence of dense myocardial scar at the location of the LL scan. The 90 patients were randomly divided into two groups, referred to as the training group and the validation group: the training group consisted of 60 patients (age 64 ± 9 years, 49 male), whose data were used to train the post-contrast blood-tissue T1 model; the validation group consisted of 30 patients (age 63 ± 11 years, 24 male), whose data were used for validation of the scar identification method. The high number of subjects in the training group is intended to learn a reliable post-contrast bloodtissue T1 relationship. The Dutch Central Committee on Human-related Research (CCMO) allows use of anonymous data without prior approval of an institutional review board provided that the data is acquired for patient care and that the data contains no identifiers that could be traced back to the individual patient. All data used for this study was acquired for clinical treatment, and was completely stripped of any identifying information.
123
Int J Cardiovasc Imaging (2014) 30:925–934
MR imaging protocol 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 12–15 imaging levels 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 9 320 mm2; matrix of 256 9 206 pixels; slice thickness 10 mm, no slice gap; flip angle 50°; echo time 1.67 ms; and pulse repetition time 3.3 ms. A LL sequence was acquired approximately 15 min after bolus injection of gadolinium DTPA (Magnevist; Schering, Berlin, Germany; 0.15 mmol/kg), at one midventricular short-axis level. Typical parameters were as follows: field of view 370 9 370 mm2; matrix of 128 9 77 pixels; slice thickness 10 mm; flip angle 8 °; echo time 1.8 ms; in total 33 phases (typical TI of the first phase 159 ms with phase interval of 15–18 ms); and pulse repetition time 3.7 ms. Immediately following the LL acquisition, T1-weighted LGE images were acquired with an inversion-recovery three-dimensional turbo-field echo sequence with parallel imaging (SENSE, acceleration factor 2). The TI was determined from the LL sequence to null the normal myocardial signal. The heart was imaged with 20–26 imaging levels in the short axis view in one or two threedimensional stacks, each stack acquired within one breathhold. Signal outside the field of view was suppressed using two saturation slabs to avoid fold-over artifacts. Typical parameters were as follows: field of view 400 9 400 mm2; matrix of 256 9 206 pixels; slice thickness 10 mm with 5 mm overlap; flip angle 15°; echo time 1.1 ms; and pulse repetition time 3.7 ms. One inversion pulse per heartbeat was used. Image processing For all 90 subjects, the endocardial and epicardial contours of the left ventricle (LV) in the short-axis LGE sequence were semi-automatically traced using the MASS Research software (version V013-EXP; Leiden University Medical Center, The Netherlands). The endocardial and epicardial contours of the LL sequences were semi-automatically defined using MASS with the following protocol: firstly, the observer manually traced the contours in three separate phases, namely, the end-diastolic, end-systolic, and earlydiastolic phases. Then the contours in the other phases were automatically generated by the software using linear
Int J Cardiovasc Imaging (2014) 30:925–934
927
A
229 ms
299 ms
369 ms
LGE slice
B
800
C
438 ms
400
500 400 300 200
200
648 ms
100 0
718 ms
T1 curves
400
T1 map (ms)
600
578 ms
508 ms
healthy myocardium dense scar
D signal intensity
159 ms
300
200 100
0
0
200
400
600
800
time (ms) Fig. 1 a The LL sequence. b The reference LGE slice with annotated remote myocardium (green) and dense scar (magenta) regions. c The computed T1 map from the LL sequence. d The averaged T1 curve for remote myocardium and dense scar
interpolation, resulting in a series of smoothly changing contours throughout the entire cardiac cycle. Subsequently, the T1 map of the myocardial tissue within endocardial and epicardial contours was computed in a pixel-wise manner: firstly the location of each pixel is defined by its radial (r) and circumferential (h) position relative to the contours, and then the pixels at different frames but with the same r and h were connected to construct a local T1 relaxation curve. The T1 value at this local pixel was computed by parametric fitting of this curve: B f ðTI Þ ¼ A B exp TI=T1 ; T1 ¼ ð1Þ T A 1 where f(TI) is the longitudinal magnetization recovery curve, parameterized by A, B and T*1, as shown in Fig. 1d. The T1 is derived from the three parameters, and the relaxation rate R1 is computed as 1/T1. The LGE slice at the same short-axis level as the LL sequence was computed using the spatial information from the DICOM header. This reconstructed LGE slice was referred to as the reference LGE slice. The image processing, data fitting, and statistical analysis were performed using the scientific computing tool MATLAB (version 2012a; MathWorks, Natick, MA). Learning the linear relationship between post-contrast tissue and blood relaxation rates The change of relaxation rate in myocardial tissue is correlated to the change of relaxation rate in blood, during the equilibrium phase (5–45 min post bolus injection) when the gadolinium clearance is compensated in tissue and blood [16]:
0
DR1tissue R1tissue R1tissue k¼ ¼ 0 DR1blood R1blood R1blood
ð2Þ
where k is the ratio between change, also called partition coefficient, R1tissue and R1blood denote the post-contrast T1 relaxation rates of myocardial tissue and blood, respectively, and R01tissue and R01blood denote their pre-contrast relaxation rates. Without pre-contrast measurements available, we utilized the following linear relationship: R1tissue ¼ ktissue R1blood þ xtissue
ð3Þ
where xtissue ¼ R01tissue ktissue R01blood . We can perform linear fitting on the R1 for remote myocardium and dense scar: R1myo ¼ kmyo R1blood þ xmyo ; R1scar ¼ kscar R1blood þ xscar
ð4Þ
where kmyo and kscar are the slopes, and xmyo and xscar are the intercepts of the linear fitting. Therefore, with the slope and intercept parameters learned, one can estimate the relaxation rates of remote myocardium and dense scar from the relaxation rates of the blood. Consequently, the remote myocardium and dense scar region can be identified from the T1 map. For each subject of the training group, two types of myocardial tissue were manually annotated on the LGE reference slice, as extremely dark or bright regions. An example is given in Fig. 1. The post-contrast relaxation rates of the remote myocardium and dense scar were computed as the averaged R1 in the annotated regions. The post-contrast relaxation rate of the blood was computed from the blood pool region excluding papillary muscles. From the scatter plot, the slope and intercept parameters were estimated by least square linear regression.
123
928
T1-based myocardial scar identification From the LL-derived T1 map, the remote myocardium and dense scar regions were identified: Firstly, the average relaxation rate of the blood was taken from the blood pool region. Then the relaxation rate of remote myocardium and dense scar was computed accordingly based on the learned relationship as expressed in Eq. 3. The remote myocardium regions with relaxation rates R1 R1myo and dense scar regions with R1 R1scar was automatically identified, respectively. Finally the two regions on the T1 map were mapped onto the reference LGE slice, providing the T1identified reference regions for remote myocardium and dense scar. Since the LGE volume is acquired by a 3D sequence with breathhold, there was no variation of intensity distribution expected across the 3D LGE volume. Therefore the reference regions from a single T1 map can be used to normalize the signal intensity of the entire 3D LGE volume. Based on the T1-identified reference regions, the LGE image was normalized as follows: (1) the regions with signal intensity equal to or lower than the average signal intensity of the T1-identificed remote myocardium region were assigned value 0 %; (2) The regions with signal intensity equal to or higher than the average signal intensity of the T1-identified dense scar region were assigned value 100 %; (3) The transitional regions were linearly mapped to the intermediate value from 0 to 100 %. The same normalization was applied to the entire three-dimensional LGE volume covering the full LV. A 50 % threshold was applied to identify the myocardial scar region. In addition, automated post-processing was applied to remove falsely accepted regions, e.g. small isolated noisy patches, by screening the size and location of the connected components [9]. Validation on reproducibility and accuracy In the validation group of 30 subjects, two observers manually segmented the myocardial scar region in the LGE images using the MASS software. Observer 1 has over 20 years’ experience and Observer 2 has over 5 years’ experience in analyzing cardiac MR images. The protocol was as follows: Firstly, both observers reviewed short-axis, long-axis, four-chamber LGE, and cine images. Secondly, both observers segmented the scar region by setting signal intensity thresholds per slice based on visual interpretation of enhancement and wall motion abnormality. Finally, the observers manually excluded those regions with an intensity value above the threshold but not considered as scar. To evaluate the reproducibility of the proposed T1-based method, both observers also manually analyzed the LL
123
Int J Cardiovasc Imaging (2014) 30:925–934
sequences using the protocol previously describe for all 30 subjects in the validation group. We further compared our method with two established scar identification methods: (1) The FWHM method, which uses 50 % of the maximum signal intensity in the annotated dense scar region as the threshold. (2) The SD method, which uses N SD above the mean value in the annotated remote myocardium region as the threshold. For our dataset, N = 6 was taken as an empirical value which agreed best with visual scar assessment. For comparison with FWHM and SD methods, the two observers also manually annotated the dense scar and remote myocardium regions. The scar size was quantified in grams, estimated by the density of 1.05 g/m. The agreement between two identification results were quantified by the Dice index, computed as 2 |V1\V2|/(|V1| ? |V2|), describing the percentage of overlap between two scar volumes V1 and V2 in three dimension. Statistical analysis Continuous variables were expressed as mean ± SD. The paired t test was performed to determine whether there exist significant differences between variables. A p value \0.05 was considered significant. Bland–Altman analysis was used to evaluate inter-observer variability. The interclass correlation coefficient (ICC) was computed to evaluate the reproducibility of the results from two observers.
Results Training In the training group of 60 subjects, the T1 value was 300.6 ± 39.7 ms for the blood pool, 344.9 ± 43.8 ms for the annotated remote myocardium, and 282.8 ± 43.7 ms for the annotated dense scar. The linear regression between R1myo and blood R1blood is shown in Fig. 2a, and the linear regression between R1scar and R1blood is shown in Fig. 2b. The Pearson’s correlation coefficients of R1myo and R1scar with R1blood were 0.70 and 0.84, respectively, with p \ 10-10 for both. The estimated kmyo was 0.56 and kscar was 1.03. Validation The validation included two parts: (1) the inter-observer variability and (2) the agreement with gold-standard. In absence of the histological data, the manual results from Observer 1, who has over 20 years’ experience in
Int J Cardiovasc Imaging (2014) 30:925–934
929
Fig. 2 a The linear regression between relaxation rates of remote myocardium and blood. b The linear regression between the relaxation rates of dense scar and blood
analyzing cardiac MR images, was considered as the gold standard in scar identification. Figure 3 shows the Bland–Altman plots of the results from the two observers for all four methods. Table 1 reported the inter-observer variability, including the identified scar size, Dice index, p values from the paired t test, and ICC coefficient. It is shown that the ICC is high for all four methods ([0.9). However, for the manual and FWHM methods, a bias existed between the two observers, with Observer 2 estimating the scar size slightly higher than Observer 1 (p \ 0.05). The SD methods yielded no statistical bias between the observers, but with relatively large variation in their differences. The T1-based method showed the highest ICC (0.997) and Dice index (98.7 ± 1.3 %), and with lowest inter-observer difference (0.0 ± 1.6 ml). Figure 4 shows the results from Observer 2 in comparison to the gold standard manual scar identification. It is can be seen from the Bland–Altman plot that the T1-based method has the least variation across the scar size. Table 2 compares the results of the three methods from both Observer 1 and 2 with the gold standard. No significant differences in scar sizing were observed for all three methods (p [ 0.05). For both observers, the T1-based method produced the highest Dice index with the gold standard. Figure 5 displays three examples of the scar identification results from all four methods: T1-based, manual, FWHM, and SD. In Fig. 5b, the intermediate results, i.e., the T1-normalized LGE image, is also shown.
instead of quantitative measurement. In comparison, the T1 mapping is able to provide quantitative measurement of tissue relaxation rates, but has limited LV coverage. This work aims to combine the advantage of both sequences for objective and reproducible myocardial scar identification. Compared to manually annotating the remote myocardium and dense scar regions in LGE, the proposed method is able to automatically identified the two reference regions based on combined analysis of T1 map and LGE, avoiding visual interpretation. In this work, we utilized the linear relationship between R1myo and R1scar to R1blood , which can be learned from a training set. In Fig. 2, it shows that the absolute relaxation rates of remote myocardium and dense scar do vary substantially among subjects, but a linear relationship to the relaxation rates of blood holds in both cases (p \ 10-10). In [17], a similar principle was used to correct T1 of tissue with T1 of blood for more stable measurement of fibrosis in the myocardium. Compared with what has been reported in literature [18], the partition coefficient of the myocardium kmyo obtained from our training data was identical (0.56), but that of the dense scar kscar was higher (1.03 vs. 0.78). The increase in kscar may be explained by the fact that our T1 map was at pixel level, while in [18] the T1 map was computed per radial sector of 60°, encompassing both dense scar and viable tissue.
Comparison with FWHM and SD method Discussion This study demonstrated that analysis of T1 and LGE sequences can lead to reproducible and accurate myocardial scar identification in post-infarction patients. Combining information from T1 and LGE LGE is able to delineate myocardial scar in fine resolution covering the whole heart, but it is based on relative contrast
We have compared our method with the two established methods: FWHM and SD, and evaluated the reproducibility between two observers, and the agreement with gold standard. Our study demonstrated that the T1-based method has the highest ICC and Dice index between two observers. The high reproducibility is explained by the fact that the two reference regions were objectively identified from the T1 values, instead of manually. The reproducibility of the SD method was the lowest since the SD value within the annotated remote myocardium largely depends on its
123
930
A 50
difference scar size (g)
100
manual Observer 2 (g)
Fig. 3 Comparison of the myocardial scar identification results between 2 observers: a manual, b FWHM, c SD, D. T1-based
Int J Cardiovasc Imaging (2014) 30:925–934
80 60 40 20 0
0
50
0
-50
100
0
manual Observer 1 (g)
B
difference scar size (g)
60 40 20
0
50
difference: 2.6±4.2
0
-50
100
0
FWHM Observer 1 (g)
80 60 40 20
0
0
-50
100
0
50
100
average scar size (g)
100
50
80 60 40 20
1
T -based Observer 2 (g)
50
difference: 0.7±11.1
STD Observer 1 (g)
D
0 0
50
100
T1-based Observer 1 (g)
location as well as size. For different dataset, N needs to be carefully tuned to the MRI data quality as SD describes the noise level. In our dataset, N = 6 leads to the best agreement with manual scar identification, as also suggested in a previous study [6].
123
100
50
difference scar size (g)
100
0
50
average scar size (g)
difference scar size (g)
STD Observer 1 (g)
C
100
50
80
0
50
average scar size (g)
100
FWHM Observer 2 (g)
difference: 3.2±3.9
difference: 0.0±1.6
0
-50
0
50
100
average scar size (g)
Comparison with the gold-standard manual scar identification showed that the T1-based method has the least variation across scar size, compared with the FWHM and SD methods. This can be explained by the fact that in the FWHM and SD methods, the variability of manual
Int J Cardiovasc Imaging (2014) 30:925–934 Table 1 Inter-observer comparison of four methods: manual, FWHM, SD, and T1based
931
Method Manual
Scar size (g) Observer 2
Dice index (%)
p value
ICC
94.0 ± 4.6
\0.001
0.977
94.6 ± 5.2
0.002
0.979
88.9 ± 10.1
0.73
0.924
98.7 ± 1.3
0.99
0.997
44.9 ± 23.4
48.1 ± 23.5
(5.2–117.8)
(7.2–116.5)
FWHM
45.2 ± 23.4
47.8 ± 23.9
(6.5–112.4)
(5.6–111.8)
SD
45.4 ± 27.8
46.1 ± 28.3
(6.6–147.5)
(8.3–144.5)
44.6 ± 21.9
44.6 ± 22.1
(5.3–124.5)
(5.5–124.5)
T1-based
100
50
difference scar size (g)
A FWHM Observer 2 (g)
Fig. 4 Comparison of the myocardial scar identification results to the gold standard: a FWHM, b SD, c T1-based
Scar size (g) Observer 1
80 60 40 20 0
0
50
difference: 0.3±11.2
0
-50
100
0
manual Observer 1 (g) 100 80 60 40 20 0
0
50
difference: 0.5±13.0
0
-50
100
0
manual Observer 1 (g) 100
100
50
80 60 40 20 0
50
average scar size (g)
difference scar size (g)
1
T -based Observer 2 (g)
C
100
50
difference scar size (g)
STD Observer 2 (g)
B
50
average scar size (g)
50
0
100
manual Observer 1 (g)
annotation may increase with the scar size, while in the T1based method the determination of the reference regions is not dependent on visual interpretation. In our method, a similar rationale as in the FWHM method, i.e. applying a 50 % threshold, was used to
difference: -0.2±7.6
0
-50
0
50
100
average scar size (g)
identify the scar. The difference however lies in the way to identify the reference regions: the FWHM method uses a single maximum signal intensity in the dense scar region as reference, while our method uses two T1-identified reference regions including both dense scar and remote
123
932
Int J Cardiovasc Imaging (2014) 30:925–934
Table 2 Comparison of the FWHM, SD, and T1-based methods to the gold-standard manual scar identification Method
Observer 1 Dice index (%)
Observer 2 p value
Dice index (%)
p value
FWHM
87.3 ± 9.4
0.87
87.4 ± 10.3
0.19
SD
85.2 ± 8.3
0.94
85.3 ± 8.5
0.63
T1-based
89.8 ± 7.5
0.86
90.2 ± 6.6
0.86
In FWHM and SD methods, one or two reference regions needs to be manually identified. In comparison, the proposed method automatically identifies such reference regions from the T1 map based on the blood-tissue T1 relationship, thereby avoiding any observer bias due to visual interpretation of the signal enhancement in LGE images. Clinical applicability
myocardium. Our method therefore accommodates imperfect myocardium nulling (i.e. when the signal intensity of remote myocardium is also elevated).
Fig. 5 Examples of myocardial scar identification results: a The original LGE images. b T1normalized LGE images. c T1based scar identification results. b Manual scar identification results. e FWHM scar identification results. f SD scar identification results
123
The proposed method is compatible with standard LGE protocols, in which a single-slice LL and a three-dimensional LGE are acquired sequentially. This implies that no
Int J Cardiovasc Imaging (2014) 30:925–934
extra scanning is needed and the method can be applied to standard LGE dataset. To utilize the available T1 information from the T1 scout sequence, we analyzed the LL sequences in a semi-automated manner. As described in the protocol, the contouring of the LL sequence is relatively easy and fast: only three cardiac phases need to be analyzed, and the rest of contours are generated automatically. For trained observers, the process takes \3 min per LL scan. Potentially, the manual contouring can be replaced by computer methods that exploit the mutual information to automatically link the individual pixels in the LL sequence [19, 20].
933
3.
4.
5.
6.
Limitations 7.
A major limitation in our study is that we do not have histological data as the gold standard. The validation of the results was based on comparison with the manual results from an experienced observer, and in addition the evaluation of reproducibility between two observers. Another limitation is that the method requires that the single-slice T1 map cover both the remote myocardium and dense scar in order to get T1-based reference regions. The condition may not always be satisfied. Alternatively, if the LL is acquired at a location without dense scar, the proposed method may potentially provide a reference region for remote myocardium, which can serve as input to the SD method.
8.
9.
10.
11.
Conclusions 12.
We have proposed a myocardial scar identification method based on analysis of routinely acquired LL and LGE sequences. The proposed method demonstrated high scar identification accuracy compared with the gold standard manual identification, with excellent inter-observer reproducibility.
13.
14.
Acknowledgments The authors would like to acknowledge the financial support from the EU MEDIATE Project (ITEA2-09039). Conflict of interest
None.
15.
16.
References 1. Yan AT, Shayne AJ, Brown KA et al (2006) Characterization of the peri-infarct zone by contrast-enhanced cardiac magnetic resonance imaging is a powerful predictor of post-myocardial infarction mortality. Circulation 114:32–39. doi:10.1161/CIRCU LATIONAHA.106.613414 2. Schmidt A, Azevedo CF, Cheng A et al (2007) Infarct tissue heterogeneity by magnetic resonance imaging identifies enhanced cardiac arrhythmia susceptibility in patients with left ventricular
17.
18.
dysfunction. Circulation 115:2006–2014. doi:10.1161/CIRCU LATIONAHA.106.653568 Roes SD, Borleffs CJW, van der Geest RJ et al (2009) Infarct tissue heterogeneity assessed with contrast-enhanced MRI predicts spontaneous ventricular arrhythmia in patients with ischemic cardiomyopathy and implantable cardioverter-defibrillator. Circ Cardiovasc Imaging 2:183–190. doi:10.1161/CIRCIMA GING.108.826529 Bello D, Fieno DS, Kim RJ et al (2005) Infarct morphology identifies patients with substrate for sustained ventricular tachycardia. J Am Coll Cardiol 45:1104–1108. doi:10.1016/j.jacc.2004.12.057 Kim RJ, Fieno DS, Parrish TB et al (1999) Relationship of MRI delayed contrast enhancement to irreversible injury, infarct age, and contractile function. Circulation 100:1992–2002 Harrigan CJ, Peters DC, Gibson CM et al (2011) Hypertrophic cardiomyopathy: quantification of late gadolinium enhancement with contrast-enhanced cardiovascular MR imaging. Radiology 258:128–133. doi:10.1148/radiol.10090526 Beek AM, Bondarenko O, Afsharzada F, van Rossum AC (2009) Quantification of late gadolinium enhanced CMR in viability assessment in chronic ischemic heart disease: a comparison to functional outcome. J Cardiovasc Magn Reson 11:6. doi:10.1186/ 1532-429X-11-6 Amado LC, Gerber BL, Gupta SN et al (2004) Accurate and objective infarct sizing by contrast-enhanced magnetic resonance imaging in a canine myocardial infarction model. J Am Coll Cardiol 44:2383–2389. doi:10.1016/j.jacc.2004.09.020 Tao Q, Milles J, Zeppenfeld K et al (2010) Automated segmentation of myocardial scar in late enhancement MRI using combined intensity and spatial information. Magn Reson Med 64:586–594. doi:10.1002/mrm.22422 Hsu L-Y, Ingkanisorn WP, Kellman P et al (2006) Quantitative myocardial infarction on delayed enhancement MRI. Part II: clinical application of an automated feature analysis and combined thresholding infarct sizing algorithm. J Magn Reson Imaging 23:309–314. doi:10.1002/jmri.20495 Mewton N, Liu CY, Croisille P et al (2011) Assessment of myocardial fibrosis with cardiovascular magnetic resonance. J Am Coll Cardiol 57:891–903. doi:10.1016/j.jacc.2010.11.013 Look DC, Locker DR (1970) Time saving in measurement of NMR and EPR relaxation times. Rev Sci Instrum 41:250–251. doi:10.1063/1.1684482 Messroghli DR, Radjenovic A, Kozerke S et al (2004) Modified look-locker inversion recovery (MOLLI) for high-resolution T1 mapping of the heart. Magn Reson Med 52:141–146. doi:10. 1002/mrm.20110 Piechnik SK, Ferreira VM, Dall’Armellina E et al (2010) Shortened modified look-locker inversion recovery (ShMOLLI) for clinical myocardial T1-mapping at 1.5 and 3 T within a 9 heartbeat breathhold. J Cardiovasc Magn Reson 12:69. doi:10. 1186/1532-429X-12-69 Nacif MS, Turkbey EB, Gai N et al (2011) Myocardial T1 mapping with MRI: comparison of look-locker and MOLLI sequences. J Magn Reson Imaging 34:1367–1373. doi:10.1002/jmri.22753 Wendland MF, Saeed M, Lauerma K et al (1997) Alterations in T1 of normal and reperfused infarcted myocardium after GdBOPTA versus GD-DTPA on inversion recovery EPI. Magn Reson Med 37:448–456 Choi E-Y, Hwang SH, Yoon YW et al (2013) Correction with blood T1 is essential when measuring post-contrast myocardial T1 value in patients with acute myocardial infarction. J Cardiovasc Magn Reson 15:11. doi:10.1186/1532-429X-15-11 Flacke SJ, Fischer SE, Lorenz CH (2001) Measurement of the gadopentetate dimeglumine partition coefficient in human myocardium in vivo: normal distribution and elevation in acute and chronic infarction. Radiology 218:703–710
123
934 19. Xue H, Shah S, Greiser A et al (2012) Motion correction for myocardial T1 mapping using image registration with synthetic image estimation. Magn Reson Med 67:1644–1655. doi:10.1002/ mrm.23153
123
Int J Cardiovasc Imaging (2014) 30:925–934 20. Van de Giessen M, Tao Q, van der Geest RJ, Lelieveldt BPF (2013) Model-based alignment of look-locker mri sequences for calibrated myocardical scar tissue quantification. International Symposium on Biomedical Imaging (ISBI): from nano to macro
