We propose a novel scheme for annotation tracking in long liver ultrasound sequences. It is based on a variational non-linear image registration method using Normalized Gradient Fields, extended by a moving window strategy based on registrations to the provided annotation on the first frame. By this we achieve robustness against error accumulation, while handling large deformations at the same time.
The method is evaluated on 21 datasets with up to five annotations as contribution to the MICCAI CLUST14 challenge. We achieved a mean tracking error of 1.31 mm with a standard deviation of 1.63 mm, while running at close to real-time speed, exceeding acquisition rate in ten cases with up to 44 frames per second on standard hardware.