Continuous Ultrasound Speckle Tracking with Gaussian Mixtures

Abstract

Speckle tracking echocardiography (STE) is now widely used for measuring strain, deformations, and motion in cardiology. STE involves three successive steps: acquisition of individual frames, speckle detection, and image registration using speckles as landmarks. This work proposes to avoid explicit detection and registration by representing dynamic ultrasound images as sparse collections of moving Gaussian elements in the continuous joint space-time space. Individual speckles or local clusters of speckles are approximated by a single multivariate Gaussian kernel with associated linear trajectory over a short time span. A hierarchical tree-structured model is fitted to sampled input data such that predicted image estimates can be retrieved by regression after reconstruction, allowing a (bias-variance) trade-off between model complexity and image resolution. The inverse image reconstruction problem is solved with an online Bayesian statistical estimation algorithm. Experiments on clinical data could estimate subtle sub-pixel accurate motion that is difficult to capture with frame-to-frame elastic image registration techniques.