Graph Cuts-Based Registration Revisited: A Novel Approach for Lung Image Registration Using Supervoxels and Image-Guided Filtering

摘要

This work revisits the concept of graph cuts as an efficient optimization technique in image registration. Previously, due to the computational burden involved, the use of graph cuts in this context has been mainly limited to 2D applications. Here we show how combining graph cuts with supervoxels, resulting in a sparse, yet meaningful graph-based image representation, can overcome previous limitations. Additionally, we show that a relaxed graph representation of the image allows for 'sliding' motion modeling and provides anatomically plausible estimation of the deformations. This is achieved by using image-guided filtering of the estimated sparse deformation field. We evaluate our method on a publicly available CT lung data set and show that our new approach compares very favourably with state-of-the-art in continuous and discrete image registration.