Facebook and New York University (NYU) have announced they are exploring how artificial intelligence can be used to make MRI scans 10x faster.
The “fastMRI” research partnership, unveiled on Monday, is between the Facebook AI Research (FAIR) group and the Department of Radiology in New York University School of Medicine. The aim of the project is to see if AI can be used to reconstruct MR images in new, faster ways.
Magnetic resonance imaging (MRI) scans are one of the best ways we have of seeing what is going on inside a patient’s body. But they require a patient to lie inside a cocoon-like metal tube for anywhere from 15 minutes to over an hour as the scanner collects large amounts of data. If successful, fastMRI could turn a 50-minute evaluation into a 5-minute evaluation, according to Dr. Dan Sodickson, an imaging specialist at NYU.
“Using AI, we believe it may be possible to capture less data and therefore image faster, while still preserving or even enhancing the rich information content of MR images,” said Sodickson. “The key is to train artificial neural nets to recognize the underlying structure of the images, and what types of things tend to be clustered together, so that the nets can then fill in views that are emitted from the accelerated scan.”
Sodickson compared the image reconstruction technique to how a person’s brain is able to successfully make out objects in the dark. “When we’re looking at a scene in low light, we don’t have a complete view of an object, but we know what the underlying structure of that object is so we can very quickly and very accurately fill in that missing information without missing a beat,” he said.
NYU will provide Facebook with approximately 3 million MR images of the knee, brain, and liver to help train its algorithms. Those behind the project said all data, including both images and raw scanner data, are being fully stripped of patient names and all other protected health information.
“The benefits of this are really dramatic in the real world,” said Sodickson. “Obviously [if the time is reduced], it’s a more comfortable patient experience and it’s particularly important for sick patients or children who have trouble staying still. You also get increased accessibility in areas with MRI shortages and you can get improved image quality when you’re trying to image things that move fast, like the heart. If we can get it fast enough to replace X-rays or CT (scans) then we can also reduce radiation exposure for the population while still getting the critical medical information.”
An unspecified number of Facebook employees in Menlo Park, New York and Montreal are involved with the project, which could be extended to other medical imaging applications in the future. Each week, FAIR employees and NYU researchers will have a video call to discuss how the project is going.
Larry Zitnick, a research manager at FAIR, said there was no specific time frame: “In six months we should be able to make good progress on this. It could take less with a breakthrough, or it could take a year.”
Zitnick added that partnering with NYU could help the social media giant get the technology into practice if it proves to be successful. “If we do show success, we have an avenue to get this out into clinical practice, test it out, put it in front of real radiologists, and make sure that what we’re doing is actually going to be impactful,” he said.
But when asked if Facebook plans to release and build medical products in the future, Zitnick didn’t give much away. Instead, he said that “FAIR’s mission is to push the science of AI forward,” before going on to say that FAIR is looking for problems where AI can have a positive impact on the world.
Facebook and NYU have a long-standing relationship, with several people working for both organizations including Yann LeCun, who was the director of FAIR before he became Facebook’s chief AI scientist. “This all got started with a connection by someone working both for NYU and in collaboration with FAIR. They suggested it’d be good for us to start talking, which we did,” said Sodickson.
Facebook and NYU plan to open source their work so that other researchers can build on their developments. As the project unfolds, Facebook said it will publish AI models, baselines, and evaluation metrics associated with the research, while NYU will open source the image dataset.
Facebook isn’t the only tech company exploring how AI can be used to assist radiologists. For example, DeepMind, an AI lab owned by Google, has developed deep learning software that can detect over 50 eye diseases from scans.
DeepMind has a number of other healthcare projects but Facebook (who was reportedly interested in buying DeepMind at one stage) claims this project is the first of its kind, as it aims to change the way medical images are created in the first place, as opposed to using existing medical images to see what can be achieved.