New AI technique could make minimally invasive surgeries safer and more precise

Researchers at MIT and collaborating institutions developed an AI technique that matches X-rays captured during surgery with a patient's preoperative 3D medical scan, according to MIT News. The system, named xvr (X-ray volume registration), adapts to each patient in about five minutes and performs the matching in seconds with sub-millimeter precision, the report states. It reportedly outperformed existing AI methods by an order of magnitude across a range of patients, body parts and procedures.

The technique addresses 2D/3D registration, the alignment clinicians use to relate real-time X-rays to preoperative CT or MRI scans during minimally invasive procedures such as angioplasty. The report describes manual registration as slow and burdensome, and says existing AI models struggle to align images robustly across patients with diverse anatomy. The researchers instead built a model tailored to one specific patient.

According to the account, xvr takes a patient's 3D scan and generates thousands of synthetic X-rays from many angles using a physics-based simulation, producing about 1,000 images per second. Training a registration model from scratch for each patient would take roughly 12 hours, so the team pretrained a foundation model on whole-body scans from more than 2,000 patients, enabling adaptation in about five minutes with the same accuracy.

The team tested the model on what the report describes as the largest available dataset of real 2D/3D registrations, with data from five hospitals covering bones and organ systems in adult and pediatric patients. The report says the model also could be used to improve robotic surgery technologies, and that future work includes making xvr faster for real-time deployment, further reliability studies, and handling more complex scenarios such as moving body parts.

Vivek Gopalakrishnan, a postdoc in MIT CSAIL and lead author, said the work is now being pursued with surgical robotics companies and clinical groups. Polina Golland and Neel Dey are co-senior authors, with additional co-authors from Harvard Medical School, St. Luke's Marion Bloch Neuroscience Institute, Shriners Children's Hospital, Brigham and Women's Hospital, Boston Children's Hospital and Harvard. The paper appears in Nature. Funding included the National Institutes of Health and several MIT programs and funds.