Patient-Agnostic Synthetic Pretraining for Efficient Patient-Specific Intraoperative 2D/3D Registration

The paper proposes a patient-agnostic synthetic pretraining framework for intraoperative 2D/3D registration. It pretrains on digitally reconstructed radiographs generated from multiple CT volumes, then adapts to a new patient using only a limited number of target-patient projections. The method combines segmentation-free domain randomization, spherical similarity learning, and differentiable Levenberg–Marquardt optimization. The stated goal is to reduce the computation and data required for patient-specific training while preserving registration accuracy across anatomical datasets.