Non-rigid registration algorithm with spatially varying stiffness properties

Non-rigid registration algorithms have been proposed over the years to register medical images to each other. One class of applications for these algorithms is the automatic segmentation of structures and substructures using a predefined atlas. But these algorithms have been limited to image volumes without gross abnormalities or pathologies and have thus been of limited use for applications such as the automatic segmentation of radiation sensitive structures for radiation therapy planning. The algorithm we present in this paper is an extension of a non-rigid registration algorithm we have previously developed (the adaptive basis algorithm). This extension permits the use of the algorithm for the automatic segmentation of medical images even when structures have been displaced substantially. The algorithm automatically adjusts the stiffness of the transformation to permit larger displacements over regions that are known to be very compliant and smaller displacements over regions that are known to be less compliant. The stiffness map can be defined once and for all in an atlas. The algorithm has been tested and evaluated on head image volumes with large ventricular enlargements and head image volumes with large space-occupying lesions.