Gradient based image registration using importance sampling

Analytical gradient based non-rigid image registration methods, using intensity based similarity measures (e.g. mutual information), have proven to be capable of accurately handling many types of deformations. While their versatility is largely in part to their high degrees of freedom, the computation of the gradient of the similarity measure with respect to the many warp parameters becomes very time consuming. Recently, a simple stochastic approximation method using a small random subset of image pixels to approximate this gradient has been shown to be effective. We propose to use importance sampling to improve the accuracy and reduce the variance of this approximation by preferentially selecting pixels near image edges. Initial empirical results show that a combination of stochastic approximation methods and importance sampling greatly improves the rate of convergence of the registration process while preserving accuracy