ORIGINAL RESEARCH
Assessment of Left Atrial Function by MRI Myocardial Feature Tracking Morgane Evin, PhD,1,2,3,4* Philippe Cluzel, MD, PhD,1,2,3,4,5 Jer^ ome Lamy, MS,1,2,3,4 David Rosenbaum, MD,1,2,3,4 Slawek Kusmia, PhD,1,6 Carine Defrance, MD,2,3,4 Gilles Soulat, MD,6 Elie Mousseaux, MD, PhD,7 Charles Roux, MD,1,2,3,4 Karine Clement, MD, PhD,1 St ephane N. Hatem, MD, PhD,1,8 Alban Redheuil, MD, PhD,1,2,3,4,5 and Nadjia Kachenoura, PhD,1,2,3,4 Background: Left atrium (LA) volumes and function are predictors of cardiovascular events. Because LA function cannot be assessed from cardiovascular magnetic resonance imaging (MRI) using the well-established left ventricular tagging techniques, we hypothesized that adequate feature tracking (FT) applied to conventional cine MRI data could characterize LA function accurately. Methods: We studied 10 young (28 6 7 years) and 10 elderly (64 6 6 years) healthy subjects, as well as 20 patients with moderate to severe aortic valve stenosis (AVS; 73 6 15 years, effective aortic valve area: 0.67 6 0.36 cm2). MRI cine two-, three-, and four-chamber views were analyzed using a newly proposed FT method based on spatial correlation and endocardial detection resulting in: regional and global longitudinal strain and strain rate, radial motion fraction and relative velocity for the three LA motion phases including reservoir, conduit, and LA contraction. Results: FT reliability was indicated by a good overlap between tracking results and manual LA endocardial borders, the low error for comparison against theoretical strains introduced in a synthetic phantom and the good inter-observer reproducibility (coefficient of variation < 15%). While all LA functional parameters were significantly impaired in AVS patients (p < 0.04), subclinical age-related variations induced a decreasing trend on all LA parameters but were significant only for radial conduit function parameters (p < 0.03). Finally, LA functional parameters characterized LA alteration in AVS with higher sensitivity than conventional LA volumetric parameters. Conclusions: Left atrial FT is feasible on MRI cine images and its addition to conventional analysis tools might enhance the diagnosis value of MRI in many heart diseases. J. MAGN. RESON. IMAGING 2015;00:000–000.
A function has been associated with atrial fibrillation1 as well as cardiovascular morbidity and mortality.2 Left atrium (LA) function can be quantified by both volumetric and contractile function parameters and its alteration is strongly associated with left ventricular (LV) performance because of the functional interplay between both chambers in healthy aging as well as in various pathological conditions. While LV diastolic and systolic function, whether assessed using global or regional indices, are widely
L
described in the literature, functional characterization of the LA is mostly performed through global morphological indices such as diameters, areas and volumes. Such indices might be insufficient in describing the complex LA function, which is divided into reservoir, conduit and contraction phase, and which is influenced by LV contraction, relaxation and filling pressures, pulmonary vein orientation as well as LA electrical activity. Accordingly, LA function indices such as strain and strain rate have been proposed using non-invasive imaging modalities such as
View this article online at wileyonlinelibrary.com. DOI: 10.1002/jmri.24851 Received Nov 12, 2014, Accepted for publication Dec 26, 2014. *Address reprint requests to: M.E., 91 bvd de l’H^ opital, 75013 Paris. E-mail: [email protected] es, UPMC Univ Paris 06, UMR 7371, UMRS 1146, Laboratoire From the 1Institute of Cardiometabolism and Nutrition, Paris, France; 2Sorbonne Universit edicale, Paris, France; 4CNRS, UMR 7371, Laboratoire d’Imagerie d’Imagerie Biom edicale, Paris, France; 3INSERM, UMRS 1146, Laboratoire d’Imagerie Biom opital Piti e-Salp^ etrie`re, Paris, France; 6ICAN Imaging Core Biom edicale, Paris, France; 5Department of Cardiovascular Radiology, Institut of Cardiologie, H^ Lab, Paris, France; 7Cardiology Departement, European Hospital Georges Pompidou, Paris, France; and 8INSERM, UMRS 1166, Laboratoire d’Imagerie Biom edicale, Paris, France Additional Supporting Information may be found in the online version of this article.
C 2015 Wiley Periodicals, Inc. V 1
Journal of Magnetic Resonance Imaging
echocardiography speckle tracking3–5 and color tissue Doppler. Although speckle tracking is presently the only available reference for LA strain estimation, ultrasound beam direction as well as heart motion relative to the probe may influence measurements6 and inter-vendor variability, essentially described in the setting of LV function, need to be further investigated.7 Cardiac MRI has been reported as the gold standard for LV mass and volumes as well as systolic function evaluation. However, only few studies have been dedicated to LA geometry and volumes,8 highlighting the higher accuracy of MRI measurements.9 This latter study advised LA characterization in all recommended routine MRI exams. Furthermore, despite the improvements in terms of spatial resolution required to capture the thin LA wall, MRI is the only non-invasive modality able to characterize myocardial tissue in both LV and LA.10–12 Tissue magnetization strategies such as tagging or strain encoding have been developed for myocardial strain evaluation from MRI data and have been widely used for LV regional function characterization. However, such techniques cannot be used for LA strain evaluation because of the small LA myocardial wall thickness which does not allow obtaining sufficient myocardial signal. In this context, our hypothesis is that the addition of innovative image processing tools for LA strain estimation from routine cine MRI data would enhance the clinical usefulness of MRI for LA characterization and allow retrospective use on large existing datasets. Ultimately, their combination with recent advances in LA electrophysiological mapping might help in planning resynchronization therapy. Accordingly, the objectives of our study were to develop a feature tracking algorithm and to assess its ability to evaluate LA function from conventional cine MRI imaging, which is systematically acquired in standard clinical MRI exams. Such development should account for LA complex geometry and wall motion. The reliability of the tracking technique was studied in terms of: (i) LA borders detection, (ii) inter-operator variability, and (iii) its ability to characterize LA function alterations in healthy aging and in moderate to severe aortic valve stenosis (AVS).
Materials and Methods Synthetic Phantom Data To test the algorithm developed for feature tracking on cine MRI data in controlled conditions, we created a dynamic dataset of synthetic images simulating LA myocardial motion as in the study proposed on the LV by Hor et al.13. Briefly, LA shape is simulated by an ellipse fitted on the maximum contraction time of a real LA image truncated between extremities of the mitral valve. An additional truncated ellipse is added at the opposite side of LA to simulate the LV and to test mitral valve annulus tracking. Ellipse motion is generated using physiological longitudinal strain values 2
(39.5% max strain) and the respective time variations extracted from a previous study14. The ellipse has a 2 pixels grey border and was filled with white to mimic blood pool signal in MRI images. This phantom comprised 60 time phases and its pixel size was 0.73 mm. Moreover, Rician noise has been added to simulate realistic MRI images. Finally, effect of spatial and temporal resolution on functional parameters was studied while: (i) varying pixel size between 0.73 and 3.14 mm and (ii) decreasing the temporal resolution by a factor of 2, 3, 4, 5, and 6 (by averaging successive time frames to mimic the blurring effect of low temporal resolution). The resulted functional parameters of radial and longitudinal strains were compared against the reference used to generate the initial phantom.
Study Population and Data Acquisition Forty subjects including 20 healthy volunteers free from overt cardiovascular disease (10 young controls: 27.7 6 6.8 years and 10 elderly controls: 63.7 6 5.7years) and 20 patients with AVS (73 6 14.5 years) were studied. In this latter group, severity of AVS was characterized by echocardiography (mean pressure gradient: 45.8 6 13.9 mmHg and effective orifice area: 0.67 6 0.36 cm2). Subjects with atrial fibrillation at the time of MRI were not included in the study. MRI exam, including steady state free precession (SSFP) acquisition on a 1.5 Tesla (T) magnet in three long axis views and short axis views, during breathholding, to cover the whole left heart, was performed using the following scan parameters: acquisition matrix 5 260 3 192, TR 5 3.7 ms, TE 5 1.5 ms, flip angle 5 50 , pixel size 5 0.74 mm 3 0.74 mm, slice thickness 5 8 mm, views per segment 5 12, temporal resolution ranged between 20 and 30 ms. The study protocol was approved by the institutional review board and informed consent was obtained from all participants.
LA Volumes and Ejection Fraction End-systolic and end-diastolic LA endocardial borders were manually traced on both two- and four-chamber views using the Qmass software (v7.6, Leiden, Netherlands) to derive LA volumes and ejection fraction (EF). The same software was also used for the estimation of LV mass and volumes as well as LV EF after manual tracing of LV endocardial and epicardial borders on a stack of short axis slices at end-systole and end-diastole.
Feature Tracking Algorithm First, two-, three-, and four-chamber SSFP images were filtered using a non-local mean filter15 to enhance the contrast between LA blood pool and myocardium. The feature tracking algorithm developed using Matlab (Matworks, 8.0 R2012b, Natick, MA) is based on spatial correlation, as commonly done for speckle tracking in echocardiography and constrained with endocardial contours detection which can be easily performed on the highly contrasted MRI images. It comprised the following two main steps (Fig. 1): (i) user manual initialization of the LA endocardial border on a single phase corresponding to maximal LA dilatation by positioning N markers which were interpolated into N*3 markers and corrected manually if required to optimize the superimposition of the resulting contour on the sharp transition in contrast between Volume 00, No. 00
Evin et al.: LA Function by MRI Myocardial Feature Tracking
FIGURE 1: Description of the feature tracking algorithm on MRI data and an example of the synthetic phantom along with the normal longitudinal strain curve used for its generation. a: Spatial correlation between 2 ROI weighted by the distance map to LA edge and/or active contour resulting in a final map. b: Ellipsoidal phantom with realistic size fitted on real LA data, and the maximum dilatation and minimum contraction were obtained from typical normal longitudinal strain curve (c) derived from literature (max 39.5%). The phantom had a pixel size of 0.73mm and 60 frames.
LA blood and myocardium. Indeed, conversely to speckle tracking where an entire myocardial feature was considered, features on MRI data consists in an interface between blood and myocardium, in other words the LA endocardial wall. Markers were carefully positioned to exclude LA appendage and junctions with pulmonary veins. (ii) Tracking of the initial markers was performed on neighboring images toward the beginning and the end of the cardiac cycle. Two-dimensional correlation was used to track similar anatomic features on a delimited neighborhood of each marker. Indeed, rectangular (6 3 6 pixels) region of interest (ROI) was defined around the current marker, perpendicular to the LA endocardial border to take into account physiological knowledge about LA wall motion. Spatial correlation between the defined ROI and the rectangular regions centered on each of its pixels and defined on the neighboring image were calculated. This resulted in a rectangular correlation map in which only pixels comprised in a conic shape defined from the LA center of mass were considered to avoid erroneous tracking. The highest correlation measured from this map corresponded to the new position of the marker on the neighboring image. In addition to correlation, another weighting which takes into account the distance of the marker to: (a) the approximated endocardial edge stored in a binary image calculated from maximal contrast variations (edge method), (b) the approximated endocardial border detected on the native image using active contours, based on Chan and Vese16. The aforementioned processing steps were combined into three different tracking options: tracking on gradient images calculated from native images (Track 1), tracking on filtered native images with the final weighting performed using either edge images (Track 2) or both active contours and edge images (Track 3). For all tracking options, radial position of each marker relative to the LA center of mass was temporally filtered by a 3 median filter to smooth the LA endocardial border and correct for erroneous trackMonth 2015
ing. Finally, the user defined three segments by choosing their first and last markers on the initial trace.
LA Functional Parameters Contours resulting from the tracking were used to calculate regional longitudinal strain and regional radial motion fraction indices on standardized LA segments (anterior/inferior, antero-septal/infero-lateral, infero-septal/antero-lateral, see Fig. 2a and b). LA regional function was characterized by longitudinal and radial indices as suggested in Geyer et al4 and adapted from the common denomination for the LV measurements: longitudinal strain and radial motion fraction (Fig. 2a; Supplementary File S1, which is available online). Longitudinal strain (Fig. 2a and c) is defined as the temporal variation of the length of the studied segment and is calculated as: t SlðtÞ5 L0L2L , with L0 the initial length of the segment and Lt its 0 length at time t. Radial motion fraction (Fig. 2a and e) corresponded to the radial relative displacement of the considered segment: t MrðtÞ5 R0R2R , with R0 the initial radius, Rt the radius at time t. 0 For each endocardial marker, radius was calculated relative to the LA center of mass and then averaged for all the markers of the studied segment. Of note, LA center of mass was calculated for each phase of the cardiac cycle using the tracked endocardial markers. Time derivatives of longitudinal and radial indices were calculated resulting in longitudinal strain rate and radial relative velocity curves. For longitudinal measurements, the following functional indices were derived from strain and strain rate curves (Fig. 2c and d): reservoir strain (SlR: first strain peak), left atrial contraction strain (SlA: second strain peak) and the conduit strain (SlE: difference between R and A peaks), systolic strain rate (SRlS’: positive strain rate peak), early LV filling strain rate (SRlE’: first negative strain rate 3
Journal of Magnetic Resonance Imaging
FIGURE 2: Global and regional LA functional parameters. The two-chamber view restricted to the LA (a), schematic views of LA orientation and segments (b), longitudinal strain curves (c), longitudinal strain rate curves (d), radial motion fraction (e), radial relative velocities for a control subject (f). Long, longitudinal; LAA, left atrium appendage; Ant, Anterior segment; Roof, LA roof segment; Inf, Inferior segment.
peak), and LA contraction strain rate (SRlA’: second negative strain rate peak). Similarly, the following indices were extracted from radial motion fraction and relative velocity curves (Fig. 2e and f ): reservoir motion fraction (MrR: first motion fraction peak), left atrial contraction motion fraction (MrA: second motion fraction peak) and the conduit motion fraction (MrE: difference between R and A peaks), systolic relative velocity (VrS’: positive relative velocity peak), early LV filling relative velocity (VrE’: first negative relative velocity peak) and LA contraction relative velocity (VrA’: second negative relative velocity peak). Finally, global indices were calculated similarly when considering the three LA segments of each view.
Feature Tracking Algorithm Evaluation On phantom data, quality of tracking and performance of the three aforementioned tracking options were assessed by the comparison of the estimated functional indices against those initially introduced in the model. For MRI data, evaluation in terms of contours overlap between tracking and manual Qmass contours was performed, using the standard measurements commonly used for segmentation algorithms evaluation such as Dice, Jaccard, mean and maximal distances17. Such analysis was performed on 10 subjects (3 views, all phases of the cardiac cycle). These two steps enabled the selection of the best tracking option. Finally, because initialization of LA markers is the only manual intervention, inter-operator variability regarding such initialization was investigated on the two-, three-, and four-chamber 4
views of 10 subjects in terms of contour overlap and differences in LA functional indices.
Statistical Analysis All continuous variables are given as mean 6 standard deviation. Bland and Altman analysis was performed for comparisons between repeated measurements and mean bias and limits of agreement (mean6 1.96*standard deviation) were provided. Boxplot graphs were used to illustrate LA functional parameters variation between young and elderly controls as well as AVS patients. Differences between young controls and elderly control as well as elderly controls and AVS patients were tested using the nonparametric Wilcoxon test and a pvalue < 0.05 indicated statistical significance. Associations between continuous variables were studied using linear regression and Pearson correlation coefficients were provided. For inter-observer variability, a coefficient of variation was calculated as the standard deviation of the differences between two series of measurements divided by the mean of the measurements. The ability of the newly proposed LA functional parameters as well as conventional volumetric parameters to detect LA functional alteration underlying AVS, in terms of sensitivity, specificity, negative and positive predictive values (NPV and PPV) as well as accuracy, was evaluated using a receiver operating characteristic (ROC) analysis and optimal thresholds were defined. Statistical analysis was performed using the R software.
Results Feature Tracking Algorithm Evaluation Table 1 summarizes absolute differences and percentages of difference between results of the three tracking options and Volume 00, No. 00
Evin et al.: LA Function by MRI Myocardial Feature Tracking
TABLE 1. Feature Tracking Algorithms Evaluationa
Absolute long. Strain diff. (Percentage) Methods
Absolute radial motion fraction diff. (Percentage)
SlR
SlE
SlA
MrR
MrE
MrA
Track 1
0.1 (0.2)
1.8 (7.6)
1.9 (11.9)
1.8 (4.2)
2.0 (7.9)
0.2 (1.3)
Track 2
1.5 (3.7)
0.4 (1.8)
1.9 (11.9)
2.3 (5.7)
2.0 (8.2)
0.3 (1.8)
Track 3
1.3 (3.1)
0 (0.1)
1.3 (8.0)
1.5 (3.5)
0.9 (3.6)
0.6 (3.5)
a
Absolute and percentage differences between functional parameters estimated using the three feature tracking algorithm options and reference strain initially used to generate the synthetic phantom. Track 1: tracking on gradient images; Track2: tracking on filtered native images using edge images; Track 3: tracking on filtered native images using both active contours and edge images. Functional parameters were: SlR: reservoir longitudinal strain, SlE: conduit longitudinal strain, SlA: LA contraction longitudinal strain, MrR: reservoir radial motion fraction, MrE: conduit radial motion fraction, MrA: LA contraction radial motion fraction.
reference longitudinal strain as well as radial motion fraction parameters, initially applied to generate the synthetic phantom. Percentage of differences was
Assessment of Left Atrial Function by MRI Myocardial Feature Tracking Morgane Evin, PhD,1,2,3,4* Philippe Cluzel, MD, PhD,1,2,3,4,5 Jer^ ome Lamy, MS,1,2,3,4 David Rosenbaum, MD,1,2,3,4 Slawek Kusmia, PhD,1,6 Carine Defrance, MD,2,3,4 Gilles Soulat, MD,6 Elie Mousseaux, MD, PhD,7 Charles Roux, MD,1,2,3,4 Karine Clement, MD, PhD,1 St ephane N. Hatem, MD, PhD,1,8 Alban Redheuil, MD, PhD,1,2,3,4,5 and Nadjia Kachenoura, PhD,1,2,3,4 Background: Left atrium (LA) volumes and function are predictors of cardiovascular events. Because LA function cannot be assessed from cardiovascular magnetic resonance imaging (MRI) using the well-established left ventricular tagging techniques, we hypothesized that adequate feature tracking (FT) applied to conventional cine MRI data could characterize LA function accurately. Methods: We studied 10 young (28 6 7 years) and 10 elderly (64 6 6 years) healthy subjects, as well as 20 patients with moderate to severe aortic valve stenosis (AVS; 73 6 15 years, effective aortic valve area: 0.67 6 0.36 cm2). MRI cine two-, three-, and four-chamber views were analyzed using a newly proposed FT method based on spatial correlation and endocardial detection resulting in: regional and global longitudinal strain and strain rate, radial motion fraction and relative velocity for the three LA motion phases including reservoir, conduit, and LA contraction. Results: FT reliability was indicated by a good overlap between tracking results and manual LA endocardial borders, the low error for comparison against theoretical strains introduced in a synthetic phantom and the good inter-observer reproducibility (coefficient of variation < 15%). While all LA functional parameters were significantly impaired in AVS patients (p < 0.04), subclinical age-related variations induced a decreasing trend on all LA parameters but were significant only for radial conduit function parameters (p < 0.03). Finally, LA functional parameters characterized LA alteration in AVS with higher sensitivity than conventional LA volumetric parameters. Conclusions: Left atrial FT is feasible on MRI cine images and its addition to conventional analysis tools might enhance the diagnosis value of MRI in many heart diseases. J. MAGN. RESON. IMAGING 2015;00:000–000.
A function has been associated with atrial fibrillation1 as well as cardiovascular morbidity and mortality.2 Left atrium (LA) function can be quantified by both volumetric and contractile function parameters and its alteration is strongly associated with left ventricular (LV) performance because of the functional interplay between both chambers in healthy aging as well as in various pathological conditions. While LV diastolic and systolic function, whether assessed using global or regional indices, are widely
L
described in the literature, functional characterization of the LA is mostly performed through global morphological indices such as diameters, areas and volumes. Such indices might be insufficient in describing the complex LA function, which is divided into reservoir, conduit and contraction phase, and which is influenced by LV contraction, relaxation and filling pressures, pulmonary vein orientation as well as LA electrical activity. Accordingly, LA function indices such as strain and strain rate have been proposed using non-invasive imaging modalities such as
View this article online at wileyonlinelibrary.com. DOI: 10.1002/jmri.24851 Received Nov 12, 2014, Accepted for publication Dec 26, 2014. *Address reprint requests to: M.E., 91 bvd de l’H^ opital, 75013 Paris. E-mail: [email protected] es, UPMC Univ Paris 06, UMR 7371, UMRS 1146, Laboratoire From the 1Institute of Cardiometabolism and Nutrition, Paris, France; 2Sorbonne Universit edicale, Paris, France; 4CNRS, UMR 7371, Laboratoire d’Imagerie d’Imagerie Biom edicale, Paris, France; 3INSERM, UMRS 1146, Laboratoire d’Imagerie Biom opital Piti e-Salp^ etrie`re, Paris, France; 6ICAN Imaging Core Biom edicale, Paris, France; 5Department of Cardiovascular Radiology, Institut of Cardiologie, H^ Lab, Paris, France; 7Cardiology Departement, European Hospital Georges Pompidou, Paris, France; and 8INSERM, UMRS 1166, Laboratoire d’Imagerie Biom edicale, Paris, France Additional Supporting Information may be found in the online version of this article.
C 2015 Wiley Periodicals, Inc. V 1
Journal of Magnetic Resonance Imaging
echocardiography speckle tracking3–5 and color tissue Doppler. Although speckle tracking is presently the only available reference for LA strain estimation, ultrasound beam direction as well as heart motion relative to the probe may influence measurements6 and inter-vendor variability, essentially described in the setting of LV function, need to be further investigated.7 Cardiac MRI has been reported as the gold standard for LV mass and volumes as well as systolic function evaluation. However, only few studies have been dedicated to LA geometry and volumes,8 highlighting the higher accuracy of MRI measurements.9 This latter study advised LA characterization in all recommended routine MRI exams. Furthermore, despite the improvements in terms of spatial resolution required to capture the thin LA wall, MRI is the only non-invasive modality able to characterize myocardial tissue in both LV and LA.10–12 Tissue magnetization strategies such as tagging or strain encoding have been developed for myocardial strain evaluation from MRI data and have been widely used for LV regional function characterization. However, such techniques cannot be used for LA strain evaluation because of the small LA myocardial wall thickness which does not allow obtaining sufficient myocardial signal. In this context, our hypothesis is that the addition of innovative image processing tools for LA strain estimation from routine cine MRI data would enhance the clinical usefulness of MRI for LA characterization and allow retrospective use on large existing datasets. Ultimately, their combination with recent advances in LA electrophysiological mapping might help in planning resynchronization therapy. Accordingly, the objectives of our study were to develop a feature tracking algorithm and to assess its ability to evaluate LA function from conventional cine MRI imaging, which is systematically acquired in standard clinical MRI exams. Such development should account for LA complex geometry and wall motion. The reliability of the tracking technique was studied in terms of: (i) LA borders detection, (ii) inter-operator variability, and (iii) its ability to characterize LA function alterations in healthy aging and in moderate to severe aortic valve stenosis (AVS).
Materials and Methods Synthetic Phantom Data To test the algorithm developed for feature tracking on cine MRI data in controlled conditions, we created a dynamic dataset of synthetic images simulating LA myocardial motion as in the study proposed on the LV by Hor et al.13. Briefly, LA shape is simulated by an ellipse fitted on the maximum contraction time of a real LA image truncated between extremities of the mitral valve. An additional truncated ellipse is added at the opposite side of LA to simulate the LV and to test mitral valve annulus tracking. Ellipse motion is generated using physiological longitudinal strain values 2
(39.5% max strain) and the respective time variations extracted from a previous study14. The ellipse has a 2 pixels grey border and was filled with white to mimic blood pool signal in MRI images. This phantom comprised 60 time phases and its pixel size was 0.73 mm. Moreover, Rician noise has been added to simulate realistic MRI images. Finally, effect of spatial and temporal resolution on functional parameters was studied while: (i) varying pixel size between 0.73 and 3.14 mm and (ii) decreasing the temporal resolution by a factor of 2, 3, 4, 5, and 6 (by averaging successive time frames to mimic the blurring effect of low temporal resolution). The resulted functional parameters of radial and longitudinal strains were compared against the reference used to generate the initial phantom.
Study Population and Data Acquisition Forty subjects including 20 healthy volunteers free from overt cardiovascular disease (10 young controls: 27.7 6 6.8 years and 10 elderly controls: 63.7 6 5.7years) and 20 patients with AVS (73 6 14.5 years) were studied. In this latter group, severity of AVS was characterized by echocardiography (mean pressure gradient: 45.8 6 13.9 mmHg and effective orifice area: 0.67 6 0.36 cm2). Subjects with atrial fibrillation at the time of MRI were not included in the study. MRI exam, including steady state free precession (SSFP) acquisition on a 1.5 Tesla (T) magnet in three long axis views and short axis views, during breathholding, to cover the whole left heart, was performed using the following scan parameters: acquisition matrix 5 260 3 192, TR 5 3.7 ms, TE 5 1.5 ms, flip angle 5 50 , pixel size 5 0.74 mm 3 0.74 mm, slice thickness 5 8 mm, views per segment 5 12, temporal resolution ranged between 20 and 30 ms. The study protocol was approved by the institutional review board and informed consent was obtained from all participants.
LA Volumes and Ejection Fraction End-systolic and end-diastolic LA endocardial borders were manually traced on both two- and four-chamber views using the Qmass software (v7.6, Leiden, Netherlands) to derive LA volumes and ejection fraction (EF). The same software was also used for the estimation of LV mass and volumes as well as LV EF after manual tracing of LV endocardial and epicardial borders on a stack of short axis slices at end-systole and end-diastole.
Feature Tracking Algorithm First, two-, three-, and four-chamber SSFP images were filtered using a non-local mean filter15 to enhance the contrast between LA blood pool and myocardium. The feature tracking algorithm developed using Matlab (Matworks, 8.0 R2012b, Natick, MA) is based on spatial correlation, as commonly done for speckle tracking in echocardiography and constrained with endocardial contours detection which can be easily performed on the highly contrasted MRI images. It comprised the following two main steps (Fig. 1): (i) user manual initialization of the LA endocardial border on a single phase corresponding to maximal LA dilatation by positioning N markers which were interpolated into N*3 markers and corrected manually if required to optimize the superimposition of the resulting contour on the sharp transition in contrast between Volume 00, No. 00
Evin et al.: LA Function by MRI Myocardial Feature Tracking
FIGURE 1: Description of the feature tracking algorithm on MRI data and an example of the synthetic phantom along with the normal longitudinal strain curve used for its generation. a: Spatial correlation between 2 ROI weighted by the distance map to LA edge and/or active contour resulting in a final map. b: Ellipsoidal phantom with realistic size fitted on real LA data, and the maximum dilatation and minimum contraction were obtained from typical normal longitudinal strain curve (c) derived from literature (max 39.5%). The phantom had a pixel size of 0.73mm and 60 frames.
LA blood and myocardium. Indeed, conversely to speckle tracking where an entire myocardial feature was considered, features on MRI data consists in an interface between blood and myocardium, in other words the LA endocardial wall. Markers were carefully positioned to exclude LA appendage and junctions with pulmonary veins. (ii) Tracking of the initial markers was performed on neighboring images toward the beginning and the end of the cardiac cycle. Two-dimensional correlation was used to track similar anatomic features on a delimited neighborhood of each marker. Indeed, rectangular (6 3 6 pixels) region of interest (ROI) was defined around the current marker, perpendicular to the LA endocardial border to take into account physiological knowledge about LA wall motion. Spatial correlation between the defined ROI and the rectangular regions centered on each of its pixels and defined on the neighboring image were calculated. This resulted in a rectangular correlation map in which only pixels comprised in a conic shape defined from the LA center of mass were considered to avoid erroneous tracking. The highest correlation measured from this map corresponded to the new position of the marker on the neighboring image. In addition to correlation, another weighting which takes into account the distance of the marker to: (a) the approximated endocardial edge stored in a binary image calculated from maximal contrast variations (edge method), (b) the approximated endocardial border detected on the native image using active contours, based on Chan and Vese16. The aforementioned processing steps were combined into three different tracking options: tracking on gradient images calculated from native images (Track 1), tracking on filtered native images with the final weighting performed using either edge images (Track 2) or both active contours and edge images (Track 3). For all tracking options, radial position of each marker relative to the LA center of mass was temporally filtered by a 3 median filter to smooth the LA endocardial border and correct for erroneous trackMonth 2015
ing. Finally, the user defined three segments by choosing their first and last markers on the initial trace.
LA Functional Parameters Contours resulting from the tracking were used to calculate regional longitudinal strain and regional radial motion fraction indices on standardized LA segments (anterior/inferior, antero-septal/infero-lateral, infero-septal/antero-lateral, see Fig. 2a and b). LA regional function was characterized by longitudinal and radial indices as suggested in Geyer et al4 and adapted from the common denomination for the LV measurements: longitudinal strain and radial motion fraction (Fig. 2a; Supplementary File S1, which is available online). Longitudinal strain (Fig. 2a and c) is defined as the temporal variation of the length of the studied segment and is calculated as: t SlðtÞ5 L0L2L , with L0 the initial length of the segment and Lt its 0 length at time t. Radial motion fraction (Fig. 2a and e) corresponded to the radial relative displacement of the considered segment: t MrðtÞ5 R0R2R , with R0 the initial radius, Rt the radius at time t. 0 For each endocardial marker, radius was calculated relative to the LA center of mass and then averaged for all the markers of the studied segment. Of note, LA center of mass was calculated for each phase of the cardiac cycle using the tracked endocardial markers. Time derivatives of longitudinal and radial indices were calculated resulting in longitudinal strain rate and radial relative velocity curves. For longitudinal measurements, the following functional indices were derived from strain and strain rate curves (Fig. 2c and d): reservoir strain (SlR: first strain peak), left atrial contraction strain (SlA: second strain peak) and the conduit strain (SlE: difference between R and A peaks), systolic strain rate (SRlS’: positive strain rate peak), early LV filling strain rate (SRlE’: first negative strain rate 3
Journal of Magnetic Resonance Imaging
FIGURE 2: Global and regional LA functional parameters. The two-chamber view restricted to the LA (a), schematic views of LA orientation and segments (b), longitudinal strain curves (c), longitudinal strain rate curves (d), radial motion fraction (e), radial relative velocities for a control subject (f). Long, longitudinal; LAA, left atrium appendage; Ant, Anterior segment; Roof, LA roof segment; Inf, Inferior segment.
peak), and LA contraction strain rate (SRlA’: second negative strain rate peak). Similarly, the following indices were extracted from radial motion fraction and relative velocity curves (Fig. 2e and f ): reservoir motion fraction (MrR: first motion fraction peak), left atrial contraction motion fraction (MrA: second motion fraction peak) and the conduit motion fraction (MrE: difference between R and A peaks), systolic relative velocity (VrS’: positive relative velocity peak), early LV filling relative velocity (VrE’: first negative relative velocity peak) and LA contraction relative velocity (VrA’: second negative relative velocity peak). Finally, global indices were calculated similarly when considering the three LA segments of each view.
Feature Tracking Algorithm Evaluation On phantom data, quality of tracking and performance of the three aforementioned tracking options were assessed by the comparison of the estimated functional indices against those initially introduced in the model. For MRI data, evaluation in terms of contours overlap between tracking and manual Qmass contours was performed, using the standard measurements commonly used for segmentation algorithms evaluation such as Dice, Jaccard, mean and maximal distances17. Such analysis was performed on 10 subjects (3 views, all phases of the cardiac cycle). These two steps enabled the selection of the best tracking option. Finally, because initialization of LA markers is the only manual intervention, inter-operator variability regarding such initialization was investigated on the two-, three-, and four-chamber 4
views of 10 subjects in terms of contour overlap and differences in LA functional indices.
Statistical Analysis All continuous variables are given as mean 6 standard deviation. Bland and Altman analysis was performed for comparisons between repeated measurements and mean bias and limits of agreement (mean6 1.96*standard deviation) were provided. Boxplot graphs were used to illustrate LA functional parameters variation between young and elderly controls as well as AVS patients. Differences between young controls and elderly control as well as elderly controls and AVS patients were tested using the nonparametric Wilcoxon test and a pvalue < 0.05 indicated statistical significance. Associations between continuous variables were studied using linear regression and Pearson correlation coefficients were provided. For inter-observer variability, a coefficient of variation was calculated as the standard deviation of the differences between two series of measurements divided by the mean of the measurements. The ability of the newly proposed LA functional parameters as well as conventional volumetric parameters to detect LA functional alteration underlying AVS, in terms of sensitivity, specificity, negative and positive predictive values (NPV and PPV) as well as accuracy, was evaluated using a receiver operating characteristic (ROC) analysis and optimal thresholds were defined. Statistical analysis was performed using the R software.
Results Feature Tracking Algorithm Evaluation Table 1 summarizes absolute differences and percentages of difference between results of the three tracking options and Volume 00, No. 00
Evin et al.: LA Function by MRI Myocardial Feature Tracking
TABLE 1. Feature Tracking Algorithms Evaluationa
Absolute long. Strain diff. (Percentage) Methods
Absolute radial motion fraction diff. (Percentage)
SlR
SlE
SlA
MrR
MrE
MrA
Track 1
0.1 (0.2)
1.8 (7.6)
1.9 (11.9)
1.8 (4.2)
2.0 (7.9)
0.2 (1.3)
Track 2
1.5 (3.7)
0.4 (1.8)
1.9 (11.9)
2.3 (5.7)
2.0 (8.2)
0.3 (1.8)
Track 3
1.3 (3.1)
0 (0.1)
1.3 (8.0)
1.5 (3.5)
0.9 (3.6)
0.6 (3.5)
a
Absolute and percentage differences between functional parameters estimated using the three feature tracking algorithm options and reference strain initially used to generate the synthetic phantom. Track 1: tracking on gradient images; Track2: tracking on filtered native images using edge images; Track 3: tracking on filtered native images using both active contours and edge images. Functional parameters were: SlR: reservoir longitudinal strain, SlE: conduit longitudinal strain, SlA: LA contraction longitudinal strain, MrR: reservoir radial motion fraction, MrE: conduit radial motion fraction, MrA: LA contraction radial motion fraction.
reference longitudinal strain as well as radial motion fraction parameters, initially applied to generate the synthetic phantom. Percentage of differences was
