Description:
Researchers at the University of Virginia have developed a novel motion estimation algorithm based on a continuous representation of discrete-sample images, with accuracy better than one 100th of a sample. A novel strategy is employed to reduce computational cost and minimize memory requirements. In addition to rigid body motion, local stretching/compression and shear, or other complex deformations, can be quantified.
Motion estimation is critical to many modern signal-processing algorithms. In medical imaging, for example, motion estimation is used to align images in extended field of view applications, estimate blood or tissue motion, and estimate radiation force or mechanically induced displacements for elasticity imaging. In all of these applications, the sampled nature of real-world images makes precise estimation of sub-sample displacements computationally costly at best and impossible at worst.
While cross-correlation and similar pattern matching techniques developed for continuous signals can be readily applied to sampled, real-world data, these approaches cannot perform motion estimates with errors of less than one-half a sample. A handful of strategies have been developed in an attempt to circumvent this limitation, and while each of these algorithms can be applied to reduce systematic errors in motion estimation, they also entail a greatly increased computational cost and significant bias errors. Spline-based image registration has also been described in the literature; published techniques, however, are limited by the use of a separable spline model and by a lack of quantification of intrinsic bias errors. It has been shown that the performance of tissue elasticity imaging can be significantly improved by the application of 2-D companding, but the computational burden of this technique has limited its application.
The present invention, "Multi-dimensional Spline based motion Estimator" (MUSE), provides a multi-dimensional approach to precise and accurate estimation of motion from sampled data. Simulated and experimental data demonstrate that MUSE outperforms other motion estimators with maximum bias of 0.002 samples and intrinsic variance of 0.003 samples. Computational complexity is significantly reduced with the MUSE algorithm. Straightforward extension to include companding and shear estimation make this algorithm particularly attractive for tissue elasticity imaging applications.
