This paper proposes a novel training-free pipeline to address the need for large training data, a key issue with existing learning-based methods for medical image registration. It is based on a fixed DINOv3 encoder and optimization of deformation field test time in feature space. It generates accurate and regular deformations on two representative benchmarks (Abdomen MR-CT and ACDC cardiac MRI), outperforming existing methods. In particular, on Abdomen MR-CT, our pipeline achieved the best performance, achieving an average Dice similarity (DSC) of 0.790, a 95th percentile Hausdorff distance (HD95) of 4.9±5.0, and a log-Jacobian standard deviation (SDLogJ) of 0.08±0.02. On ACDC cardiac MRI, we also improved the average DSC to 0.769 and reduced SDLogJ and HD95 to 0.11 and 4.8, respectively. This suggests that operating in a compact underlying feature space at test time could be a practical and general solution for clinical registration without additional training.
