Anyone who has ever gone in for an MRI scan knows that it can be a lengthy process. Facebook researchers and doctors from NYU Langone Health think they have a technologically-advanced solution. The team says they were able to use artificial intelligence (AI) to speed up the time it takes to generate MRI scans.
It could be a major breakthrough for the healthcare world and yet another demonstration of how AI can be impactful in the sector.
Changing the Norm
MRI scans are an incredibly helpful diagnostic tool. Doctors can use them to obtain detailed images of organs, muscles, and other soft tissues. An MRI scan can be used to diagnose everything from a tumor to a torn ACL. However, patients are often forced to lie in a very still position for longer than an hour.
That can be a difficult task for those with extreme pain, claustrophobia, or sheer impatience.
MRI machines generate images of the human body with strong magnetic fields and radio waves. Hydrogen atoms in the body react to the magnetic field and emit a radio frequency, which is then picked up by the scanner. With enough data, an image can be created.
The Department of Radiology at NYU Langone Health worked with Facebook engineers on a study that involved 108 patients receiving an MRI scan of their knee. The team’s research is set to be published in the American Journal of Roentgenology.
With the typical image generation method, each scan took between eight and 11 minutes. Scans conducted with Facebook’s “fastMRI” AI system took between four and six minutes. That’s a significant reduction. It’s worth noting that a typical MRI session consists of several scans to get different views of the area.
Importantly, the study also shows that the AI-enhanced MRI scans are accurate despite being faster and requiring less data than the traditional method. Essentially, AI is able to predict what the final image will look like by referring back to its training and applying it to the abstract information it receives.
Dan Sodickson, a professor of radiology at NYU Langone Health, said, “The neural net knows about the overall structure of the medical image. In some ways what we’re doing is filling in what is unique about this particular patient’s [scan] based on the data.”
Trained for Accuracy
Of course, any use of AI in the healthcare sector immediately draws suspicion. Many worry that algorithms can’t be trusted like human caregivers. Though that perception isn’t going to change overnight, projects like fastMRI are gradually contributing to a shift in mentality.
Sodickson says, “The key word here on which trust can be based is interchangeability. We’re not looking at some quantitative metrics based on image quality. We’re saying that radiologists make the same diagnoses. They find the same problems. They miss nothing.”
Most radiologists involved with the study couldn’t tell which scans had been generated by the AI. In fact, many of them thought they resulted in clearer scans.
When faced with AI-generated scans and traditional ones, doctors made the exact same diagnoses. That means the algorithm was able to reliably generate images that accurately mimic current MRI technology in a fraction of the time.
That’s a noteworthy breakthrough. Many AI-based image generation programs struggle to create high-resolution outputs from low-resolution data. Errors are often introduced along the way as the AI tries to make something out of nothing. While the idea of an AI system “guessing” what it sees in an MRI scan is worrying, the team says it isn’t an issue.
“We don’t just allow the network to create any arbitrary image,” says Sodickson. “We require that any image generated through the process must have been physically realizable as an MRI image. We’re limiting the search space, in a way, making sure that everything is consistent with MRI physics.”
A few minutes might not seem like that big of a deal. However, in the medical world, minutes can be a matter of life and death. For instance, a patient undergoing an MRI while suffering a stroke might have just minutes to live while the blood supply to their brain is interrupted.
NYU Langone Health’s Dr. Michael Recht says, “We think in the brain we can accelerate much faster than we are able to do in the knee. And if we can accelerate… three times as fast or four times as fast, then you’re really saving significant amounts of time.”
Aside from the potential to save lives, the AI method could have other implications. By decreasing the time needed for each MRI scan, hospitals would effectively be able to perform scans on more patients. Since MRIs are so useful, they could eventually be used to diagnose even more conditions.
Recht says, “We’re getting the same or more information in less time. So we should be able to use this sequence for all types of pathology if it holds up.”
What Comes Next?
It’s easy to see why hospitals would want this sort of technology in their radiology departments. However, getting them to adopt it won’t necessarily be easy.
One thing that will help is the fact that fastMRI is completely open access and can be incorporated into existing MRI scanners. This means hospitals won’t need to invest in expensive new machinery. Sodickson notes that researchers are already in contact with the companies that manufacture MRI machines.
Getting manufacturers to implement a new solution often causes a bottleneck—in most industries, but especially in healthcare. If the fastMRI team is able to get MRI machine manufacturers on board early, the adoption of their technology could happen fairly quickly. For both patients and hospitals, that’s a good thing.
It will be very interesting to see how AI is used in the medical imaging field in the coming years. All indications suggest that it will play a significant role in many advancements. Perhaps one of the first will be the fastMRI system.