Fast random walker image registration using precomputation

Shawn Andrews, Lisa Tang, Ghassan Hamarneh

Abstract: In this paper, we introduce an extension to the random walker image registration method designed to increase the speed at which a registration is performed. Our method involves precomputing data from one of the images being registered while anticipating the acquisition of the second image, and then using this precomputed data to approximate the final transformation once the second image becomes available. The precomputation scheme incorporates a parameter controlling the trade-off between registration speed and accuracy that can be tuned when the registration is being performed. Our results show that with precomputation, random walker image registration runs 3 to 10 times faster on volumetric images with only 3% to 20% loss in registration accuracy.