New AI Lines Up Live X-Rays With 3D Scans During Surgery

When surgeons operate using images, they rely on two different views of your body: flat, real-time X-rays taken during the operation, and detailed 3D scans taken beforehand. To use both together, the flat X-ray has to be lined up perfectly with the 3D picture — a process called 2D/3D registration. Get it wrong and the surgical tool on screen isn't where the tool actually is. Today that alignment is slow: older software often needs to be retrained for each individual patient before it works well.

A research team introduced XPos3R, an AI model that skips that preparation entirely. Think of it as a translator that speaks both languages — flat X-rays and 3D volumes — instead of only one. The model uses an asymmetric design, meaning the two halves of the system are built differently on purpose, one for each type of image. To train it without needing mountains of rare hospital data, the team generated millions of synthetic medical images and organized them by body anatomy.

The results are striking. A single pretrained version of XPos3R beat the older patient-specific methods on both accuracy and reliability. With a few seconds of fine-tuning at the time of use, it pushed the 3D positioning error under 4 millimeters — smaller than the width of a pencil eraser — and the image alignment error under 1 millimeter. The work was accepted to ECCV 2026, a major computer vision conference.

The catch: this is a research paper, not a product. It hasn't been tested in real operating rooms at scale, relies heavily on synthetic training data, and would still need clinical trials and regulatory approval before hospitals could use it. But as a proof of concept, it suggests the slow, patient-by-patient setup step may soon be a thing of the past.