Localization of 3D Anatomical Structures Using Random Forests and Discrete Optimization

In this paper we propose a method for the automatic localization of complex anatomical structures using interest points derived from Random Forests and matching based on discrete optimization. During training landmarks are annotated in a set of example volumes. A sparse elastic model encodes the geometric constraints of the landmarks. A Random Forest classifier learns the local appearance around the landmarks based on Haar-like 3D descriptors. During search we classify all voxels in the query volume. This yields probabilities for each voxel that indicate its correspondence with the landmarks. Mean-shift clustering obtains a subset of 3D interest points at the locations with the highest similarity in a local neighboorhood. We encode these points together with the conformity of the connecting edges to the learnt geometric model in a Markov Random Field. By solving the discrete optimization problem the most probable locations for each model landmark are found in the query volume. On a set of 8 hand CTs we show that this approach is able to consistently localize the model landmarks (finger tips, joints, etc) despite the complex and repetitive structure of the object.