Rapid patient-specific neural networks for X-ray to volume registration
Abstract: Advanced navigation techniques in image-guided interventions and surgical robotics require the rapid and precise alignment of three-dimensional (3D) preoperative volumes (such as computed tomography and magnetic resonance imaging) to two-dimensional (2D) intraoperative images (such as X-ray fluoroscopy) 1,2 . However, existing 2D/3D registration methods fail to generalize across the broad spectrum of fluoroscopy-guided procedures: intensity-based optimizers require per-individual hyperparameter tuning 3,4 , whil…
United States Naval Medical Bulletin Vol. 28, Nos. 1-4, 1930 by U.S. Navy. Bureau of Medicine and Surgery. Public domain
On 2026-09-16, researchers described xvr, a self-supervised framework that trains patient-specific neural networks from a patient's own preoperative CT or MRI using physics-based simulation, then uses gradient-based optimization to align 3D volumes to 2D intraoperative X-ray fluoroscopy. A foundation model pretrained on thousands of whole-body scans allows 5-minute fine-tuning to any anatomical region.
The work matters because fast, accurate 2D/3D registration underpins navigation in fluoroscopy-guided interventions and surgical robotics, where prior methods failed to generalize or required extensive manual labels. By reporting high accuracy in seconds across anatomies, modalities, and hospitals in the largest real-fluoroscopy evaluation to date, the study suggests broader clinical adoption is feasible, though uncertainty remains about performance beyond rigid alignment and in settings without high-quality preoperative volumes.
- Foundation model pretrained on thousands of whole-body scans then fine-tuned per patient using physics-based simulation from that patient's own preoperative scan.
- Eliminates manual annotation and per-individual hyperparameter tuning required by prior intensity-based and deep-learning registration methods.
- Evaluated as largest real fluoroscopy 2D/3D registration study to date across diverse anatomical structures, modalities, and hospitals.
Self-supervised patient-specific neural network with physics-based simulation aligns intraoperative X-ray fluoroscopy to preoperative 3D volumes in seconds with order-of-magnitude accuracy improvement across anatomies and hospitals.
The rundown
The framework, called xvr, combines patient-specific neural networks with gradient-based optimization and uses physics-based simulation to create training data from the patient's own CT or MRI, removing manual labeling. A foundation model pretrained on thousands of whole-body scans enables adaptation to any region in 5 minutes.
The authors position xvr against intensity-based optimizers that need per-individual tuning and deep-learning methods that need extensive labeled data and stay constrained to trained anatomy. They report open-source release to make the approach accessible to clinical and research communities.
Sources
- Peer-reviewedNature2026-09-16
How should this claim be treated?
ace
The debate