MIT Publishes Patient-Specific AI That Matches Surgical X-Rays in Seconds

MIT researchers and clinical collaborators have published xvr, an AI system intended to make X-rays captured during surgery more useful for navigation. It adapts to a new patient in about five minutes, then matches their X-rays with a CT or MRI scan in seconds; the researchers report sub-millimeter precision.

The research, published in Nature, addresses a persistent problem in minimally invasive procedures. Clinicians can use live X-rays to guide devices through small incisions, but the images are flat. Matching them to a preoperative 3D scan helps locate a tool in relation to the patient’s anatomy.

Training on one patient rather than everyone

Manual image matching can require a clinician to enter values into a computer or select anatomical landmarks on a screen. The researchers say broadly trained AI systems can struggle with differences between patients, so xvr is designed to learn from the person undergoing the procedure rather than rely on one model that must fit every anatomy.

Xvr starts with that patient’s CT or MRI scan and uses a simulation of how X-rays pass through the body to create thousands of synthetic views from different angles. It produces about 1,000 images a second. Those simulated images train a model to connect the live X-ray view with the patient’s 3D scan.

A promising evaluation, not a surgical deployment

Training a model from scratch for each person would take about 12 hours, the team says. To shorten that step, it pretrained a broader model using whole-body scans from more than 2,000 patients, then fine-tunes it for each new patient. The researchers evaluated xvr using real image-matching data from five hospitals, including adult and pediatric patients and dozens of bones and organ systems.

In that evaluation, the researchers reported an order-of-magnitude improvement over existing AI-based methods across patients, body parts, and procedures. Those are research results, not evidence of a clinical deployment. The team plans further reliability studies, faster real-time operation, and tests involving harder conditions such as moving body parts. It is also working with surgical-robotics companies and clinical groups on possible navigation or deployment tools.