Researchers have created a technique that rapidly and accurately aligns intra‑operative X‑ray images with a patient’s pre‑operative 3D medical scan. By enabling precise guidance of catheters, endoscopes and similar tools, the method promises faster, safer minimally invasive procedures.
During such surgeries clinicians rely on real‑time X‑rays, but the flat images lack depth cues, making it hard to infer the exact location and orientation of instruments and raising the risk of tissue damage. Manual registration with CT or MRI scans is slow and depends on extensive training. Existing AI‑based registration models struggle with the wide anatomical variability across patients and suffer from limited annotated data, limiting their practical use.
The MIT‑led team introduced xvr (X‑ray Volume Registration), a patient‑specific machine‑learning approach. It first generates thousands of synthetic X‑rays from the patient’s pre‑operative 3D scan using a physics‑based simulation that produces realistic images at about 1,000 per second, eliminating hallucinations. An AI model trained on these simulated views can then align real 2D X‑rays to the 3D volume within seconds with sub‑millimeter precision.
Training a model from scratch for each patient would take roughly 12 hours, which is infeasible for emergencies. To accelerate adaptation, the researchers pre‑trained a foundation model on synthetic data derived from over 2,000 whole‑body scans covering diverse ages, modalities and body regions. This model can be fine‑tuned to a new patient in about five minutes while retaining the accuracy of a full‑scratch training.
Evaluation on the largest available real‑world 2D/3D registration dataset—spanning five hospitals, multiple organs and both adult and pediatric cases—showed that xvr outperforms other AI methods in both accuracy and robustness, and runs fast enough for urgent surgeries. The system also has potential to enhance robotic surgery navigation.
Future efforts will focus on real‑time deployment, broader clinical validation, and extending the approach to handle moving anatomy. The team is already collaborating with surgical‑robotics companies and clinical groups to turn the research into practical navigation tools.
Review