xvr: system matches intraoperative X-rays to a 3D scan in seconds

In minimally invasive procedures like angioplasty, a doctor guides instruments through a small incision while monitoring their position on a live X-ray. The image, however, contains only two dimensions of information. It is difficult to determine exactly where the instrument is located in the patient's body or how it is oriented. Manually matching an X-ray to a preoperative CT or MRI is time-consuming, and existing AI tools have handled this task inconsistently across different patients.

A team from the Massachusetts Institute of Technology and collaborating institutions developed a system called xvr. It does not attempt to train a single model for all patients. Instead, it adapts to the specific individual. Using the patient's preoperative CT or MRI data, it generates simulated X-ray images from various angles, replicating the physical processes of image formation. The system produces approximately 1000 simulated images per second. Because the simulation relies solely on the patient's own data, it does not generate information from nothing, thereby avoiding the problem of hallucination.

From twelve hours to five minutes

Training a model dedicated to a single patient from scratch would take about 12 hours. The teams therefore used 3D scans from more than 2000 patients to pretrain a base model. For a new patient, fine-tuning takes about five minutes, after which the image matching itself completes in a few seconds.

In tests using real data from five hospitals, the system covered adult and pediatric patients as well as several dozen bone and organ configurations. xvr proved more accurate and stable than existing AI methods. It was also fast enough for emergency situations, such as in an operating room. The results were published in the journal Nature.

The practical significance is twofold. Faster and more reliable image registration helps doctors assess the position of a catheter or endoscope, reducing the risk of damage to surrounding tissues. The same technology can also be used to navigate surgical robots.

Uncertainty must also be measured

Image registration is not the only area of research focused on building trust in imaging models. On the list of medical physics preprints on arXiv, which is updated continuously, papers appeared in the same week on artifact reduction in cone-beam imaging (arXiv:2609.30169), a generative alternative to Monte Carlo calculations in photoacoustic imaging (arXiv:2609.28261), and identifiable uncertainty in CT imaging (arXiv:2609.29599). The last direction is clinically significant: a model that can indicate how uncertain it is about its own results is easier to integrate into a decision-making process.