FMIR, a foundation model-based Image Registration Framework for Robust Image Registration
Fengting Zhang, Yue He, Qinghao Liu, Yaonan Wang, Xiang Chen, Hang Zhang
TL;DR
FMIR addresses cross-domain generalization challenges in medical image registration by coupling a foundation-model-based feature encoder with a general registration head trained on a single dataset. It processes 3D medical volumes via a slice-based 2D foundation-model encoder and uses a multi-scale pyramid to predict a deformation field \( \mathbf{u} \) through the transformation \( \boldsymbol{\phi}(x)=x+\mathbf{u}(x) \). A channel regularization strategy, along with PCA-based inference, reduces reliance on dataset-specific priors while preserving cross-domain performance, enabling strong in-domain accuracy and robust out-of-domain generalization across benchmarks. The approach demonstrates plug-and-play compatibility with multiple foundation-model backbones and provides a practical pathway toward generalizable, resource-efficient registration foundation models.
Abstract
Deep learning has revolutionized medical image registration by achieving unprecedented speeds, yet its clinical application is hindered by a limited ability to generalize beyond the training domain, a critical weakness given the typically small scale of medical datasets. In this paper, we introduce FMIR, a foundation model-based registration framework that overcomes this limitation.Combining a foundation model-based feature encoder for extracting anatomical structures with a general registration head, and trained with a channel regularization strategy on just a single dataset, FMIR achieves state-of-the-art(SOTA) in-domain performance while maintaining robust registration on out-of-domain images.Our approach demonstrates a viable path toward building generalizable medical imaging foundation models with limited resources. The code is available at https://github.com/Monday0328/FMIR.git.