Interactive Model-based Image Registration

We present an interactive technique for the registration of captured images of elastic and rigid body parts in which the user is given flexible control over material specific deformation properties. Our method can effectively handle arbitrary stiffness distributions and it achieves an accurate matching without sacrificing the physical correctness of the simulated deformations. The algorithm consists of three steps, which are performed iteratively until an optimal spatial mapping is determined: First, the optical flow is used to predict an initial image transformation. Second, a priori knowledge of the deformation model is used to refine the predicted field. A physics-based filter operation generates a transformation that is consistent with the model of linear elasticity. Third, the process is repeated using the displaced template image. To achieve accurate image deformations we employ implicit multigrid solvers using finite differences (optical flow) and finite elements (linear elasticity). The robustness and accuracy of our method is validated using synthetic and real clinical data composed of heterogeneous materials exhibiting different stiffness characteristics.