The aim of this study was to compare two-dimensional (2D) and three-dimensional (3D) methods for computing left ventricular (LV) rotation.
A two-axis linear/rotary system was designed using rotary motors controlled through a digital interface, and 10 freshly harvested pig hearts were studied. Each heart was mounted on the rotary actuator with the base being rotated at different known degrees of rotation (10°, 15°, 20°, and 25°) and was passively driven by a pump with calibrated stoke volume (50 mL) at a constant rate (60 beats/min) simultaneously. Cardiac motion was scanned to acquire 2D short-axis views using a GE Vivid 7 system for assessing rotation, and 3D apical full-volume loops were acquired using a Toshiba Applio Artida ultrasound system. Full-volume 3D image loops were analyzed online with Toshiba Wall Motion Tracking software, and short-axis 2D images were analyzed offline for LV rotation in GE EchoPAC PC at corresponding LV levels.
At each state, both 2D and 3D echocardiography detected the changes in LV rotation but overestimated the rotation degrees. The biases for overestimation from 3D imaging were smaller compared with 2D imaging at each LV level. Both methods, when compared with each other, showed a linear correlation ( r = 0.84, P < .0001). Bland-Altman comparison showed 99% of data points within range, with a constant bias between both methods (adjusted values of 3D = 1.892 + 0.964 × 3D).
Although 3D echocardiography showed smaller bias, the results between 2D and 3D echocardiography were comparable.
Left ventricular (LV) twist is suggested as an important index of contractility and a potential marker of myocardial dysfunction in the diseased heart. As an index of systolic function, LV twist is computed from the net difference of counterclockwise apical and clockwise basal LV rotation during systole. Rapid unwinding of this systolic twist has also been shown to have significant contribution in early diastolic filling. An accurate quantification of twisting LV motion, therefore, can provide important information about both systolic and diastolic function of the heart.
A noninvasive imaging-based assessment of LV twist is of significant resource in the clinical evaluation of dynamic LV function. With the introduction of digital tracking of speckles in ultrasound image loops, high–frame rate two-dimensional (2D) echocardiography was tested for the computation of LV twist from short-axis (SAX) views and validated. However, selection of optimal imaging planes for such computation is quite challenging, because of limited acoustic windows and oblique orientation of the heart in the patient’s chest cavity. Despite an accurate assessment of LV rotation with speckle tracking, the out-of-plane myocardial motion during 2D acquisition is a major source of error, especially near the LV base.
More recently, three-dimensional (3D) echocardiography has been introduced and is rapidly being integrated in clinical imaging because of its enhanced display of cardiac anatomy. New-generation matrix transducers have allowed the sequential acquisition of scan-line data with electrocardiographic gating and the reconstruction of 3D loops with a programmable degree of overlapping between successive volumes to suppress the through-plane motion, with resolution high enough to track speckle motion through the volumes for the computation of mechanical functions. The ability to obtain a full-volume image loop with a single acquisition from the same level and analyze the same volume at multiple levels avoids many difficulties of 2D echocardiographic methods. Digital feature tracking has also been tested to compute LV twist from high–frame rate 3D echocardiographic image loops and validated against sonomicrometry. We sought to compare both 2D and 3D echocardiographic methods to quantify LV rotation at different levels using a custom-designed phantom.
A two-axis linear/rotary torsion system was designed using rotary stepper motors to achieve both linear and rotary motion synchronized to each other, simulating base-to-apex linear and twisting cardiac motion. The rotary stepper motors were controlled using a six-step switching technique to allow for discrete commutation of the rotor and precision control of the position of each of the axis. An Atmel AT90CAN128 microcontroller (Atmel, San Jose, CA) was used to control the stepper motors and provide a digital interface ( Figure 1 ). We studied 10 freshly harvested pig hearts. Each heart was mounted on the rotary actuator, with the base being rotated and the apex held fixed to avoid translational motion but permitting rotation. A pulsatile pump was connected to a balloon inserted into the LV cavity, and its rate of pumping was synchronized with the rate of rotation and linear motion. With each counterclockwise rotation as viewed from the LV apex, the base moved linearly toward the apex, which was held fixed, and the heart was emptied by synchronized negative suction of the pulsatile pump. The heart was filled with a positive upstroke of the pump with a known stroke volume and was synchronized with apex-to-base lengthening and clockwise untwisting, as viewed from LV apex. An electrocardiographic signal from the pulsatile pump was used to synchronize the pumping and torsion system. Linear motion of the torsion system was set to 20% of LV length for each heart. We studied different degrees of rotation (10°, 15°, 20°, and 25°) at a constant rate of 60 beats/min and stroke volume of 50 mL.
Full-volume 3D images were acquired from an apical window using a Toshiba Applio Artida ultrasound system (Toshiba, Tokyo, Japan) at a rate >22 volumes/sec. Frequency was optimized at 3.5 MHz, and scan range angle was set to 70° × 70° over six beats gated by the electrocardiographic signal generated from the pulsatile pump to acquire a complete dynamic image loop with uniform resolution quality throughout the depth of image. We used a Toshiba ultrasound system for the acquisition of 3D data sets, because it was the only system with released software to analyze 3D image loops for LV rotation at the time of the study. Similarly, a GE Vivid 7 Dimensions ultrasound system (GE Healthcare, Milwaukee, WI) was used to acquire 2D SAX views at two different LV levels with a 10S probe. The frame rate of acquisition was set at 80 to 90 frames/sec, and frequency was optimized to achieve uniform resolution quality throughout the depth of the image. Apical 2D SAX views were acquired from a distance equal to 20% of LV length from the apex, and basal 2D SAX views were acquired from a similar distance from the base of the heart. We used the GE Vivid 7 system for the acquisition of 2D data sets because a majority of published studies of LV rotation and twist have used this system.
The 3D image data were analyzed for LV rotation using the new speckle tracking–based motion-detecting Wall Motion Tracking (WMT) program from Toshiba. One cycle of complete rotation and reversed motion was selected for resolution quality and complete inclusion of the LV wall. The WMT program shows 3D image data in two long-axis and three SAX 2D views that correspond to conventional two-chamber and four-chamber and parasternal SAX views (basal, middle, and apical). The program allows users to adjust the orientation of these 2D planes and draw an initial contour to define endocardial and epicardial borders on both long-axis views at the reference end-diastolic frame ( Figures 2 A and 2 B). SAX planes are created by spline interpolation. The center of gravity of each SAX plane is created from the average of points defining SAX contour, and rotation center is defined by the center of gravity of each SAX plane in each frame. The software divides the LV wall into 16 segments and computes the degree of rotation globally and for each segment automatically, while tracking apex-to-base longitudinal motion ( Figure 2 ). We computed planar rotation for the LV base and apex for comparison with 2D echocardiography by calculating the mean of their segments at each plane. The 2D image data were also exported to a Windows workstation for offline analysis of LV rotation using the 2DSR program embedded in EchoPAC (GE Healthcare). One cycle of complete rotation and reversed motion was similarly selected from image loop for rotation analysis. The 2D program requires the user to select an appropriate region of interest on the reference frame of image loop. To calculate LV rotation, a center of gravity is defined by averaging kernels within the region of interest around the circumference of the SAX view. The shift of selected points in the region of interest is calculated as angular displacements around the center of gravity. After processing the speckle motion in successive frames of a dynamic 2D grayscale image, the software calculates and displays both segmental and global LV rotation values. The direction of rotation with reference to the position of the probe is also shown ( Figure 3 ). We used global rotation values at each 2D SAX planes.
We used different statistical tools to compare our results of LV rotation from both 2D and 3D echocardiographic methods in this controlled phantom study. We assessed the discrepancy between these two methods through the quantities of bias and mean square error. Nonparametric Wilcoxon’s signed-rand test was performed on the difference between echocardiography-derived LV rotation and reference measurements. In addition, a linear mixed effects model was used to correlate repeated measurements of LV rotation by both methods (2D and 3D echocardiography) at different states of the same pig heart. To assess the agreement between both 3D and 2D methods, we used Bland-Altman plots. We also used concordance correlation coefficient to avoid the bias in estimating the variance that may happen when using Pearson’s correlation coefficient. The concordance correlation coefficient combines measures of both precision and accuracy to determine how far the observed data deviate from the line of perfect concordance, that is, the line at 45° on a square scatterplot. Last, we determined intraclass correlation coefficients to access the reliability or consistency between these two methods (3D vs 2D echocardiography). P values < .05 were considered significant.
The results illustrated were derived from the images that were selected for their resolution quality and inclusion of complete LV walls within the pyramidal full-volume 3D and 2D cine loop. Planar rotation for each LV level was derived by averaging the angular displacement of each segment at that level for both 2D and 3D loops. Results are expressed as mean ± SD.
When the model was rotated, the LV base showed similar rotation being attached firmly to the rotary actuator. Moreover, apical levels of the left ventricle showed much lower degrees of rotation because the apex was essentially held fixed to avoid translational motion. Thus, there was a base-to-apex gradient in segmental rotation. Nonetheless, detected segmental rotation at all LV levels showed a strongly positive linear correlation with the actual degree of rotation imposed by the motor device. With increase in actual rotation at motor device, the degree of rotation at the LV base, middle, and apex also increased accordingly and was detected by both 2D and 3D echocardiographic methods ( Table 1 ). The degree of rotation at the apex was smaller than that at the base and middle of actual heart rotation (both P values < .001), because it was partially restricted by the stabilizing apex. There was no statistically significant difference between the degree of rotation of the LV base and actual rotation at the motor device (all P values > .05). Each segment and the global rotation had good correlations with the actual rotation at the motor device (base r = 0.96, middle r = 0.93, and apex r = 0.68, all P values < .001 for the 3D method; base r = 0.95, middle r = 0.85, apex r = 0.85, all P values < .001 for the 2D method).
|Reference rotation at motor device (°)||Measured rotation at LV base (°)||Measured rotation at LV middle (°)||Measured rotation at LV apex (°)|
|10||11.25 ± 0.85||13.45 ± 2.55||8.19 ± 0.75||9.95 ± 1.85||7.20 ± 0.55||8.15 ± 2.61|
|15||16.86 ± 1.35||17.13 ± 1.58||12.83 ± 1.84||13.36 ± 0.99||9.25 ± 3.30||11.14 ± 2.18|
|20||21.29 ± 1.79||23.25 ± 1.62||17.65 ± 1.42||18.16 ± 1.28||12.82 ± 3.49||15.35 ± 2.63|
|25||26.18 ± 1.86||27.09 ± 1.41||23.15 ± 1.90||24.59 ± 2.14||14.59 ± 4.21||19.25 ± 3.75|
We assessed the discrepancy of 2D and 3D echocardiographic methods by comparing the results obtained from each method at the LV base (attached firmly to the rotary actuator) to the reference values (actual rotation at the rotary actuator). Both methods overestimated the LV rotation at each state of rotation, but the 3D method showed relatively lower bias (1.18 ± 2.15) compared with the 2D method (2.09 ± 3.15). The discrepancy between both methods was further examined through the quantity of mean square error. The 3D method showed a smaller mean square error (35.98) than the 2D method (38.66), indicating that the 3D method had small discrepancy. We also performed Wilcoxon’s signed-rank test on the difference between echocardiography-derived rotation at LV base and reference measurements at each state of rotation and constructed the 95% confidence intervals of each difference ( Table 2 ). Most confidence intervals for 3D measurements contained zero, indicating no difference between the 3D method and reference; we can infer that 3D measurements were similar to reference values. Similarly, because confidence intervals for most 2D echocardiographic measurements did not contain zero, we can infer that 2D measurements were different from reference values. At 15° rotation, both the confidence intervals for both 3D and 2D echocardiography contained zero, but the confidence intervals for 3D echocardiography were much narrower than those for 2D echocardiography, indicating higher precision in 3D measurements compared with 2D measurements. Similarly, at 10° rotation, both the confidence intervals of both 3D and 2D echocardiography did not contain zero, but the confidence interval for 3D imaging was narrower than for 2D imaging, indicating that the 3D method was more accurate than the 2D method.
|Reference rotation (°)||95% confidence interval of difference between measured and reference rotation|
|3D echocardiographic rotation|
|10°||(0.1500 to 1.6750)|
|15°||(−2.0700 to 0.3950)|
|20°||(−1.0700 to 1.5700)|
|25°||(−1.1850 to 1.7200)|
|2D echocardiographic rotation|
|10°||(1.7500 to 4.0700)|
|15°||(−1.4850 to 2.1100)|
|20°||(0.1049 to 2.5001)|
|25°||(0.5351 to 2.8300)|
The comparison of 2D and 3D echocardiographic rotation measurements using the linear mixed effects model was adjusted by the location (base, middle, or apex). The P value (.1961) indicates that either method (3D or 2D) is not significantly different. We could also say that both methods are relatively comparable. We also obtained the P value (.0014) of the adjusting location factor. This result indicates that the rotation values were significantly different from one location (i.e., base, middle, or apex) to the other LV levels.
The concordance correlation coefficient estimates at the base, middle, and apex of the hearts were 0.923, 0.927, and 0.608, respectively ( Figure 4 ). Although the concordance correlation coefficient at the apex was slightly weaker than those at the base and middle, this coefficient is still in the reasonable range. The 95% confidence intervals for these coefficient estimates were 0.877 to 0.959, 0.875 to 0.957, and 0.411 to 0.752, respectively. These results indicate that the 3D and 2D methods were concordant. The estimated intraclass correlation coefficients were 0.931, 0.928, and 0.614 for the basal, middle, and apical locations of hearts, respectively. The bootstrap sample technique was used to construct 95% confidence intervals to draw an inferential statistical conclusion. The intervals were 0.8965 to 0.9544, 0.9032 to 0.9537, and 0.4627 to 0.7547 with respect to the three estimated intraclass correlation coefficients. None of these three confidence intervals contains zero, so we can infer that both the 3D and 2D methods were consistent.