Based on a stochastic mutual information type matching and RPROP as stochastic optimizer, an interactive image-based registration of a CT volume onto two 2D images provided by a megavoltage system is presented. The matching process is based on semi-automatic pre-segmentation, an approximate 2D-2D matching with precomputed virtual projections (DRRs) followed by an accurate 3D-2D matching step. Our sample-based approach requires only a fraction of computed DRRs for 3D-2D. A simultaneous computation of the DRR rays and their perturbations in 6 dimensions speeds up the rendering process by a factor of 6.8. The complete registration process takes 5.6 +/- 2.3 seconds on a 3 GHz Pentium IV PC, being the fastest non-parallel approach for this sort of application the authors are aware of.