MetaPose: A meta-learned framework with dual-branch frequency-aware network for fast and accurate 2D/3D image registration

During pelvic fracture reduction surgery, fast and accurate 2D/3D image registration is of great clinical importance for intraoperative navigation and bone fragment reduction. However, conventional learning-based registration methods usually require training a separate registration network for each patient, which limits their generalizability. To address this issue, we propose a 2D/3D image registration method based on Model-Agnostic Meta-Learning (MAML) and a differentiable X-ray rendering optimization mechanism, while introducing a Dual-branch Frequency Perception Network (DFPNet) to enhance the capture of local details. This framework is trained via meta-learning on an open-source pelvic dataset, yielding a pose estimation network with robust generalizability, and achieves sub-millimeter registration accuracy for new patients through quick adaptation and rendering-based test-time optimization. A systematic evaluation was conducted on public and private pelvic fracture datasets, and the experimental results demonstrate that the proposed method significantly outperforms existing mainstream registration methods in terms of mean mTRE, registration success rate, and robustness, indicating promising clinical applicability.