Combining Automatic Landmark Detection and Variational Methods for Lung CT Registration

This paper proposes a novel method for image registration of lung CT scans. Our approach consists of a procedure for automatically establishing landmark correspondences in lung CT scan pairs and an elaborate variational image registration scheme. The landmark information is incorporated into the registration scheme as pre-registration using the landmark-based Thin-Plate-Spline (TPS) method. The TPS displacement field is improved by an additional minimization of an objective function consisting of a Normalized Gradient Field distance measure, a volume term, and a curvature regularizer. As a special property, landmark correspondences as established by the TPS registration are guaranteed to remain within a user-defined tolerance during the variational registration step. The new method, called LMP (LandMark Penalty), is applied to the 20 publicly available DIR-Lab data sets and compared to state-of-theart methods. Particularly on the challenging COPDgene data sets, LMP stands out with an average landmark error of 1.43 mm.