Abstract: Three-dimensional (3D) image registration is a computationally intensive problem which is commonly solved in medical imaging. The complexity of the problem stems from its size and nonlinearity. In this paper we present an approach that drastically reduces the problem size by using adaptive mesh refinement. Our approach requires special and careful discretization of the variational form on adaptive quad/octree grids. It further requires an appropriate refinement criterion. We show that in some cases this approach can reduce the computational time by a factor of approximately 10 or so in two dimensions and 5 in three dimensions compared to the nonadaptive approach.