Neural Deformable Cone Beam CT

In oral and maxillofacial cone beam computed tomography (CBCT), patient motion is frequently observed and, if not accounted
for, can severely affect the usability of the acquired images. We propose a highly flexible, data driven motion correction and
reconstruction method which combines neural inverse rendering in a CBCT setting with a neural deformation field. We jointly
optimize a lightweight coordinate based representation of the 3D volume together with a deformation network. This allows our
method to generate high quality results while accurately representing occurring patient movements, such as head movements,
separate jaw movements or swallowing. We evaluate our method in synthetic and clinical scenarios and are able to produce
artefact-free reconstructions even in the presence of severe motion. While our approach is primarily developed for maxillofacial
applications, we do not restrict the deformation field to certain kinds of motion. We demonstrate its flexibility by applying it to
other scenarios, such as 4D lung scans or industrial tomography settings, achieving state-of-the art results within minutes with
only minimal adjustments.