Improving Lung Registration by Incorporating Anatomical Knowledge: A Variational Approach

In this work, we present a novel approach for the registration of CT lung images. Therefore, we incorporate additional segmentation information yielding a significant improvement of accuracy, robustness, and reliability. The main idea of our approach is rather general and not limited to the case of lung registration. We describe a generic method for incorporating segmentation information into a variational image reg-istration framework. Assuming that segmentation masks are available for reference and template image (e.g. masks separating lung tissue from the background), the method drives the registration process towards exact alignment of the masks. Furthermore, we extend the classical variational setting by an additional term that controls change of volumes and in particular guarantees non-singular deformation fields. Both extensions can be combined with arbitrary distance measures and regularizers, and therefore can be adapted to arbitrary registration tasks.