Patient-Agnostic Pretraining Boosts Patient-Specific 2D/3D Registration
A novel framework for intraoperative 2D/3D registration enhances efficiency through patient-agnostic synthetic pretraining. Detailed in arXiv:2607.23343, this technique initially pretrains a model on synthetic digitally reconstructed radiographs (DRRs) derived from various CT volumes to capture transferable pose-sensitive features. Subsequently, it fine-tunes the model for a specific patient using only a small set of synthetic projections from the target CT. This strategy significantly lowers the computational expenses associated with developing individual patient-specific models from the ground up, tackling a major barrier to practical application in image-guided procedures.
Key facts
- Intraoperative 2D/3D registration aligns preoperative CT volumes with intraoperative X-ray or fluoroscopic images.
- Recent learning-based methods show promising accuracy in patient-specific settings using DRRs.
- Training a separate patient-specific model from scratch for each patient is computationally inefficient.
- The proposed framework uses patient-agnostic synthetic pretraining and spherical similarity learning.
- Pretraining on synthetic DRRs from multiple CT volumes learns transferable pose-sensitive representations.
- Adaptation to a new patient requires only a limited number of synthetic projections from the target CT.
- The method aims to improve efficiency for practical deployment in image-guided interventions.
- The paper is available on arXiv with ID 2607.23343.
Entities
Institutions
- arXiv