MIT Develops AI System xvr for Fast Submillimeter X-Ray and 3D Scan Matching

Scientists at the Massachusetts Institute of Technology and other institutions have developed an AI system called xvr that can accurately match intraoperative 2D X-ray images with preoperative 3D CT or MRI scans within seconds, achieving submillimeter precision. The findings were published in the latest issue of Nature.

During minimally invasive procedures such as angioplasty, doctors rely on real-time X-ray images to track surgical instruments, but 2D images make it difficult to determine the exact position and orientation of the tools. Instead of training a one-size-fits-all model, xvr adapts to each individual patient. It generates a large number of simulated X-ray images from different angles based on the patient's own CT or MRI data, while simulating the physical process of real X-ray imaging, producing about 1,000 simulated images per second. Because the simulations are grounded in the patient's actual 3D scans, the system avoids the hallucination issues common in generative AI.

To reduce personalized training time, the team pre-trained a foundation model using 3D scans from more than 2,000 patients. For a new patient, the system can be personalized in about five minutes and then complete image matching within seconds.

In tests on real data from five hospitals, covering adult and pediatric patients as well as dozens of bone and organ systems, xvr outperformed existing AI methods in both accuracy and stability, while running fast enough for emergency surgery. The technology could help clinicians better determine the position of catheters, endoscopes, and other instruments, potentially lowering the risk of tissue damage and supporting surgical robot navigation.