Uncertainty Estimation in Non-Rigid Image Registration
Speaker: Petter Risholm , Harvard Medical School/BWH
Registration uncertainty means the lack of complete certainty, i.e. the existence of more than one probable transformation that will align two images, which is certainly the case in non-rigid registration which is generally ill-posed and requires regularization because of noisy images, homogenous image intensity regions and the underdetermined nature of the problem. In contrast to the uncertainty in rigid point-based registration which is generally well understood, there has been little focus on non-rigid registration uncertainty in the registration community. In this talk I will introduce a Bayesian non-rigid registration framework where Boltzmann’s distribution is used to convert traditional energy functions used in non-rigid registration (i.e. similarity and regularization) into probabilities. Through the use of robust Markov Chain Monte Carlo sampling, we characterize the posterior distribution over deformations and are able to quantify the most probable deformation as well as its uncertainty. Our recent results highlight the versatility of this framework and the possible clinical impact of registration uncertainty in areas such as estimating delivered dose during head and neck radiotherapy, prostate registration for estimating dosimetric coverage during brachytherapy, estimating lung stiffness, and intra-operative registration of brain-images.