Researchers have long looked for ways to speed up magnetic resonance imagining (MRI). Two years ago, Facebook’s Artificial Intelligence Research group (FAIR) began collaborating with researchers at New York University’s Medical School on artificial intelligence that held the promise of reducing the time it takes to get an MRI from an hour to as few as 15 minutes. The group is now releasing the results of its first study, which shows that the quality and usefulness of these AI-fueled MRI scans are not detectably different from those of traditional MRI. Called FastMRI, the technology stands to make MRI more accessible and affordable if adopted broadly.
The study, conducted by NYU researchers, took conventional MRI scans of the knees of 108 patients, pulled out some of the raw data, and created new MRI scans using AI to produce an image based off this new, limited set of information. Six radiologists reviewed all of the scans and made clinical recommendations for each. The study found that overall, radiologists made the same recommendations for the original scan as they did for the fastMRI scan. Researchers say that the radiologists preferred the quality of the FastMRI over the traditional and only one was reliably able to discern the difference between the two. The study says that the results suggest the medical community may be keen to adopt FastMRI technology if and when it becomes available.
MRI machines, used for taking pictures of soft tissues inside a person’s body, are often donut shaped. The hole at the center produces both a magnetic field and radio frequency pulses that, introduced at different angles, produce a picture of the interior of the body. Depending on the complexity of the scan, it can take over an hour to get an image. This limits the number of people who can get a scan in a day and also keeps the cost of a scan high. There are often redundancies in the various snapshots that the MRI takes to get the final big picture. The FastMRI cuts out about 75% of these data points to speed up the process and uses artificial intelligence to fill in the gaps. The AI also reduces the noise in the image, giving FastMRI a clearer picture than traditional MRI, which is why the radiologists liked it more.
This study was designed to determine if FastMRI images could be used interchangeably with traditional MRI. Prior to the study, there was a concern that this technology could introduce artifacts into images that weren’t really there or possibly omit information. This study suggests that FastMRI is free of these potential problems.
“We’re able to train the deep learning models that we built, using MRI knowledge—how MRIs are gathering information in the first place,” says Nafissa Yakubova, the researcher at FAIR who worked on the algorithm.
Work in progress
Though promising, the study was limited in a few ways. It only used one type of MRI machine and only tested the technology’s ability to replicate the soft tissue around a knee. Dr. Dan Sodickson, a researcher and professor at NYU Langone Health, says he’s working with other vendors and research institutions to set up additional multisite studies and further validate the FastMRI technology. The technology is open source, making it easy for others to access. He’s also looking at expanding the FastMRI to other regions of the body.
“This process should be relatively generalizable, and we have some evidence that it is, but we want to explore any fine-tuning that’s needed to accelerate, for example, in head MRIs or in other areas,” he says.
Other institutions have tried to speed up MRI in the past. In 2019, Boston University researchers developed a copper and plastic device that helped MRI machines produce clearer images faster. Though potentially cheap to produce, the device was ultimately impractical, a study concluded. NYU Langone Health has been working on using artificial intelligence to make MRI imaging faster since 2016, but they didn’t team up with Facebook’s AI Research unit until 2017.
Other than the potential to make a vital tool of modern healthcare available to more people, why is Facebook working on an application of AI that’s so far afield from its own business? Rather than seeking profit or to productize the MRI tech itself, Yakubova says that FAIR can learn from this project because MRI imaging is such a complex undertaking. “The bar for medical imaging is higher,” she says. “For example, in the knee, an abnormality you’re looking for, it could be just a few pixels large. You’ve seen, maybe, AI that generates faces of celebrities pretty realistically, but if you start zooming in, you see some details [that aren’t rendered] right.”
When health is involved, you can’t afford that kind of inaccuracy, but the insights Facebook gains could be useful for computer-vision efforts closer to home. “Medical imaging offers new challenges for advanced AI vision, so I think that’s a lot of lessons learned,” says Yakubova.