Motion Tracking for Medical Applications using Hierarchical Filter Models

A medical intervention often requires relating treatment to the situation, which it was planned on. In order to circumvent undesirable effects of motion during the intervention, positional differences must be detected in real-time. To this end, in this thesis a hierarchical Particle Filter based tracking algorithm is developed in three stages. Initially, a model description of the individual nodes in the aspired hierarchical tree is presented. Using different approaches, properties of such a node are derived and approximated, leading to a parametrization scheme. Secondly, transformations and appearance of the data are described by a fixed hierarchical tree. A sparse description for typical landmarks in medical image data is presented. A static tree model with two levels is developed and investigated. Finally, the notion of 'association' between landmarks and nodes is introduced in order to allow for dynamic adaptation to the underlying structure of the data. Processes for tree maintenance using clustering and sequential reinforcement are implemented. The function of the full algorithm is demonstrated on data of abdominal breathing motion.