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

Researchers have developed a new artificial intelligence technique that can rapidly and accurately match X-rays taken during surgery with a patient's preoperative 3D medical scan. This innovation could greatly improve the precision and safety of minimally invasive surgeries, which often rely on real-time X-rays to guide devices such as catheters and endoscopes through tiny incisions.

However, the flat nature of X-rays makes it difficult to determine exactly where the surgical tools are located and oriented within the patient's body, increasing the risk of complications.

The new system, called xvr (short for X-ray volume registration), uses an AI model specifically tailored to each patient to match their X-rays with 3D scans in just a few minutes, with sub-millimeter precision. This is a significant improvement over existing AI methods, which often struggle to align images robustly for all patients, rendering them impractical for use. The MIT researchers tested xvr across various patients, body parts, and medical procedures, and it outperformed other AI techniques by an order of magnitude.

The lead author of the paper detailing xvr, Vivek Gopalakrishnan, explains that clinicians typically use X-rays in minimally invasive procedures to visualize the surgery from any angle. However, aligning real-time X-rays with preoperative MRI or CT scans is crucial for guiding surgical tools safely. This process, called registration, helps clinicians determine the location and orientation of the tools in relation to the patient's anatomy.

Traditionally, manual registration methods are slow and cumbersome, requiring clinicians to guess the position of surgical instruments by inputting numbers or clicking landmarks on a screen. To address this, the researchers created xvr, which first generates thousands of synthetic X-rays for each patient using a physics-based simulation of the X-ray process. These synthetic images are then used to train an AI model that can accurately align the patient's real-time X-rays with their 3D scan in a matter of seconds.

While xvr's registration model is highly accurate, training a new model from scratch for each patient would take about 12 hours, making it impractical for emergency situations. To overcome this challenge, the researchers pretrained a more versatile AI system, known as a foundation model, using xvr. This foundation model can quickly adapt to each new patient, allowing the registration process to be completed in a much shorter time frame.