Surgeons performing minimally invasive procedures often rely on X-ray images for instrument navigation. However, a flat image doesn't always provide a complete understanding of a catheter or endoscope's position and orientation within the body, increasing the risk of complications. Standard AI solutions designed to speed up this process often struggle with the diversity of human anatomy.
A new system called xvr, developed at MIT, addresses this problem. It adapts to each patient in just five minutes by aligning X-ray images with pre-operative 3D scans (CT or MRI) with sub-millimeter accuracy. This is tens of times better than existing AI methods, according to the developers. The system generates realistic synthetic X-ray images based on the patient's 3D data, eliminating AI "hallucinations."
"Most Americans live more than an hour away from a center where emergency intervention, such as for a stroke, can be performed. An hour in such a situation is critically long. Simplifying these procedures by combining 2D and 3D information makes highly specialized life-saving operations more accessible to a wider population," notes Vivek Gopalakrishnan, the lead author of the study.
If previously precise navigation required decades of training to learn to "see" in grainy 2D images, xvr makes X-ray images more informative. This means that complex but vital procedures can become safer and more accessible, even for patients far from major medical centers. The system does not require manual data entry or clicking on anatomical landmarks, significantly reducing the surgeon's time and effort.