Inertial demons: A momentum-based diffeomorphic registration framework

Abstract

Non-linear registration is an essential part of modern neuroimaging analysis,from morphometrics to functional studies. To be practical,non-linear registration methods must be precise and computational efficient. Current algorithms based on Thirion’s demons achieve high accuracies while having desirable properties such as diffeomorphic deformation fields. However,the increased complexity of these methods lead to a decrease in their efficiency. Here we propose a modification of the demons algorithm that both improves the accuracy and convergence speed,while maintaining the characteristics of a diffeomorphic registration. Our method outperforms all the analysed demons approaches in terms of speed and accuracy. Furthermore,this improvement is not limited to the demons algorithm,but applicable in most typical deformable registration algorithms.