A novel measure to express tracking quality in ultrasound block matching

Speckle de-correlation is a major problem in block matching based ultrasound methodologies as it limits the accuracy of the tracking result. It would be of benefit to have a quantitative measure expressing the local tracking quality as it would allow discarding unreliable motion estimates. We hypothesized that kernels showing sufficient gray scale pattern would more reliably track than kernels with more homogenous gray scale distributions. The aim of this study was to test this hypothesis in-silico. Ultrasound B-mode sequences were simulated from a kinematic model of the carotid artery. Two-dimensional motion was estimated using block matching with the normalized cross-correlation function as similarity measure. For each kernel, two measures of tracking quality were stored: the normalized cross-correlation coefficient (Ccc) and a measure of the amount of edges inside the kernel detected using a canny filter and counted on a pixel-by-pixel basis. As such, a quality measure (Cedge) between 0 (no edges) and 1 (nothing but edges) was obtained. Axial and lateral strains were subsequently obtained by linear regression in regions of interest (ROIs) with best/worst mean tracking quality scores. The root-mean-squared-error (RMSE) was significantly lower in regions with low Ccc (worst ROI) compared to ROIs with high Ccc. However, more edges in the kernel did indeed result in better overall tracking (lower RMSE). Thus, the proposed edge-detection method showed to be a better tracking quality measure than the commonly used Ccc.