Landmark Localization in CT Images Using Dense Matching of Graphical Models

We present a method based on graphical models for the localization of corresponding anatomical landmarks in CT images of multiple patients for which only limited labeled training data is available. Our method mobilizes anatomical spatial relationships learnt from labeled training images in order to improve dense matching using weak landmark appearance descriptors. In this study, we report results for localization of 22 different anatomical landmarks in 20 unseen lung cancer patients and different types of anatomical constraints (none, box-range, Gaussian). The average registration error over all landmarks improved from 18.8 voxels (37.6mm) of the raw landmark descriptors to 4.2 voxels (8.4 mm) using the anatomical constraints.