Image Registration Techniques for Independent Acquisitions of Cone Beam Computed Tomography Volumes

This work is centered on the alignment of a volumetric dataset acquired through Cone Beam Computed Tomography (CBCT) technology. Monomodal registration is useful for comparing different acquisitions of the same anatomical district for monitoring a pathology progression or regression, as well as for stitching together CBCT consecutive volume segments, usually when a large region of interest does not in fit the device field of view. Several methods were studied, both intensity and feature-based. Gradient-free techniques and evolutionary algorithm class, in particular genetic algorithms, were investigated. Results were analyzed to establish which approach is more efficient and accurate. Convergence speed represents a known issue of this evolutionary algorithms that was handled through the choice of an adequate stop criterion. Results were presented over a dataset, where a known rigid transformation matrix is applied, with the aim of comparing the estimated transformations with the actual ones.