Training-Free Deformable Image Registration Using Unsupervised Multi-Teacher Distillation
Abstract
Purpose
Deformable image registration (DIR) in medical imaging remains inherently ill-conditioned due to structural ambiguities and weak anatomical constraints. Although foundation models (FMs) have shown strong promise for unsupervised DIR, existing approaches typically rely on a single FM backbone, limiting robustness, domain generalization, and the ability to integrate complementary priors. In this study, we address these limitations through multi-teacher knowledge distillation to learn a unified and efficient representation for accurate DIR without task-specific fine-tuning.
Methods
We propose PRISM-Reg, a training-free DIR framework that leverages a unified student foundation model distilled from three complementary teachers: MedSAM, DINOv3, and BiomedCLIP. Given moving and fixed images, the student encoder is kept frozen and used to extract deep feature representations. DIR is performed via test-time optimization of the deformation field directly in the student feature space by maximizing similarity between fixed features and warped moving features, while enforcing deformation regularity using smoothness and Jacobian-based constraints. PRISM-Reg performance was evaluated on multi-modality, cross-institutional dataset comprising of 50 cardiac MRI and eight abdominal MR-CT pairs using quantitative metrics including Dice similarity coefficient , 95th-percentile Hausdorff distance, and the standard deviation of the log-Jacobian determinant.
Results
PRISM-Reg produced anatomically plausible deformation vector fields and demonstrated strong visual alignment between warped moving images and fixed references. Quantitatively, PRISM-Reg achieved a DSC of 0.770±0.104 on cardiac MRI, substantially outperforming the state-of-the-art baseline (0.668±0.137; p<0.01). On the abdominal MR–CT dataset, PRISM-Reg similarly achieved higher alignment accuracy, with DSC of 0.768±0.165 compared with 0.642±0.199 for the baseline.
Conclusion
By optimizing deformation fields directly in a distilled foundation-model feature space at test time, PRISM-Reg improves alignment accuracy and correspondence quality while producing anatomically realistic deformations across heterogeneous imaging modalities. The multi-teacher distillation strategy enables continual integration of complementary FM knowledge, positioning PRISM-Reg as a practical, extensible, and generalizable framework for training-free DIR.