MIT Researchers Develop Patient-Specific AI for More Precise Surgical Navigation

Summary

MIT researchers and clinical collaborators developed xvr, a patient-specific AI method that aligns X-rays captured during minimally invasive surgery with a patient’s preoperative CT or MRI scan. This registration helps clinicians locate and orient tools such as catheters and endoscopes relative to internal anatomy, addressing the difficulty of interpreting flat, real-time X-ray images. Xvr uses a patient’s 3D scan to generate thousands of synthetic X-rays through physics-based simulation, producing about 1,000 images per second. A foundation model trained on scans from more than 2,000 patients can adapt to a new patient in about five minutes, while registration itself takes seconds and reaches sub-millimeter precision. Training a separate model from scratch would take about 12 hours, so the pretrained system is intended to make patient-specific accuracy practical in urgent procedures. Tested on the largest available real 2D/3D registration dataset, spanning five hospitals, varied bones and organ systems, and adult and pediatric patients, xvr outperformed existing AI methods by an order of magnitude in accuracy and robustness. The researchers say it could support emergency interventions and robotic surgery, but further validation, faster real-time operation, and handling of moving body parts remain future work.