Fast non-rigid registration of medical image data for image guided surgery

With the increasing importance of merging multiple datasets of different imaging modalities for image guided surgery applications, the speed by which such datasets can be registered is crucial. This paper presents a novel approach based on fast Radial Basis Function (RBF) evaluation using the biharmonic spline (BHS) and thin Plate spline (TPS) models for non-linear registration. The new algorithm is evaluated both in standard software and hardware assisted versions. A number of experiments on medical image data are presented, illustrating substantial speedups compared to a ‘brute-force’ implementation of a radial basis function based registration algorithm. The global transformation accuracy of these techniques has been evaluated by using two sets of anatomical landmarks, one set for calculating the model parameters, and the other set for assessing registration accuracy. Finally, we demonstrate that the technique yields sub-second performance and target registration errors of about 1.2mm on the intra-subject registration of CT and MRI image datasets obtained from the Vanderbilt database, when implemented on the GPU of a standard PC with high-end videoadapter card.