The goal of this work is to reliably and accurately localize anatomical landmarks in 3-D computed tomography scans, particularly for the deformable registration of whole-body scans, which show huge variation in posture, and the spatial distribution of anatomical features. Parts-based graphical models (GM) have shown attractive properties for this task because they capture naturally anatomical relationships between landmarks. Unfortunately, standard GMs are learned from manually annotated training images and the quantity of landmarks is limited by the high cost of expert annotation. We propose a novel method that automatically learns new corresponding landmarks from a database of 3-D whole-body CT scans, using a limited initial set of expert-labeled ground-truth landmarks. The newly learned landmarks, called B-landmarks, are used to build enriched GMs. We compare our method of deformable registration based on such GM landmarks to a conventional deformable registration method and to a "baseline" state-of-the-art GM. The results show our method finds new relevant anatomical correspondences and improves by up to 35% the matching accuracy of highly variable skeletal and soft-tissue landmarks of clinical interest.

We propose a method for reliably and accurately identifying anatomical landmarks in 3D CT volumes based on dense matching of parts-based graphical models. Such a system can be used to establish reliable correspondences in medical images which can be useful on their own or as part of more complex processing e.g. atlas building. We propose and investigate novel methods for efficiently optimizing parameters of appearance models for landmark localization in 3D images. We also investigate the trade-off between the number of model parameters and registration accuracy. We present results for the localization of 22 landmarks in clinical 3D CT volumes of cancer patients and optimization of part-specific patch scales. Over-fitting is likely due to an intrinsically high variability of the data and a limited labeled training and test set, here 83 scans, so we employ a rigorous bootstrap analysis to validate the results. The average mean and maximum registration error over all landmarks is reduced by 31% and 25% for the optimized model, compared to an empirically determined baseline. Additionally, we show a significantly improved performance over standard methods as the number of free parameters increases from an isotropic patch scale shared by all parts, to specific anisotropic patch scales learnt for each part in the model.