Non-rigid registration with position dependent rigidity for whole body PET follow-up studies

A constrained non-rigid registration method is proposed for intra-patient registration of whole body images. Both the deformation field and the constraint operate on individual voxels. The deformation is implemented as a displacement vector for every voxel. The constraint is modeled as a position dependent nonlinear resistance against deformation, using a spring model. This allows us to assign a different rigidity to different pairs of neighboring voxels, based on a segmented image, assuming that such an image is available. As a measure of image alignment, either mutual information or the sum of squared differences is used. The use of mutual information makes the software more widely applicable, while the sum of squared differences is expected to perform slightly better for intra-modality applications. The method has been tuned for registration of
¹⁸
F FDG PET whole body images of the same patient in oncology. Thresholding is applied to segment regions with high tracer uptake. These regions are assigned a high degree of rigidity. As a result, the volume of lesions (and also of the bladder and the brain) will not be changed during registration. After registration, subtraction images are computed, which are expected to be helpful for the evaluation of treatment response. A first evaluation indicates that the registration reduces the time required for visual and quantitative analysis of the images.