A parametric study on processing parameters for two-dimensional cardiac strain estimation: an in-vivo study

In the last years, two-dimensional (2D) myocardial strain estimation has been evolving as a method that could replace todays 1D strain estimation and solve several of the limitations of this technique. The 2D strain estimation is typically based on 2D displacement tracking of spatial features in the gray-scale (GS) or radio-frequency (RF) data using either 1D or 2D kernels. All approaches have specific advantages and disadvantages but a trade-off between accuracy and computational cost generally has to be made. Although some studies have been presented in the literature on the effect of these parameters, these have been limited to simulated or tissue mimicking phantom setups. The aim of this study was therefore to evaluate the influence of these parameters on the strain estimate in an in-vivo setting in order to enable an optimization of computation time. Methods: B-mode IQ data sets were acquired from the inferolateral wall of 4 open chest sheep. 2D displacements were calculated in each data set based on a sum of squared differences estimator using different parameters. Myocardial radial and longitudinal strain were simultaneously estimated in the inferolateral wall using a previously described methodology based on 2D motion gradients. Three segment-length sonomicrometry crystals gave a continuous reference for the longitudinal and radial strain. Results: No statistically significant differences were found for the different kernel sizes in neither axial or azimuth direction. Similarly, neither sampling frequency nor GS vs. RF tracking had a significant influence on the mean error. However, for all parameters longitudinal strain errors were signficantly smaller than the radial ones (p < 0.01). Conclusion: It has been shown in the literature that under ideal conditions the variance in the strain estimator is highly affected by parameters as window size and sampling-frequency. These results did not reproduce in the in-vivo setting indicting that other sources of noise is more dominant. Short 1D kernels might thus be advantageous in 2D strain estimation since they give a similar accuracy at a reduced computational cost.