The NYU School of Medicine’s department of radiology and Facebook recently announced a new collaborative research project focused on using artificial intelligence (AI) to make MRI scans up to 10 times faster.
Daniel Sodickson, MD, PhD, vice chair for research in radiology and director of the Center for Advanced Imaging Innovation and Research at NYU School of Medicine, is one of the specialists leading that research. He spoke with Radiology Business about what led to the collaboration, why it’s so important to make MRI faster and how radiologists have adapted to AI over the years, among other topics.
Radiology Business: How did NYU end up teaming up with Facebook on this project?
Daniel Sodickson, MD, PhD: Here at NYU, we have been working on accelerating MRI by any means available. In 2016, we described some of the first uses of deep learning for image reconstruction from accelerated data acquisitions, and now that is an exploding area in MR research. There were three abstracts on that topic at the 2016 annual meeting of the International Society for Magnetic Resonance in Medicine—one of them was ours—and closer to 100 in 2018. So this is a fast-moving field very much in its “Wild West” phase and in need of real AI expertise combined with real physics and biology expertise.
A colleague at NYU connected us with the Facebook Artificial Intelligence Research (FAIR) group, and the challenge of reconstructing fast MR images from limited data really appealed to them, both because it raised fundamental questions for AI and because it addressed a problem with a significant impact. It became clear early on in our conversations with FAIR that this would be a really synergistic partnership.
AI is being used in so many different ways in healthcare right now. Why did NYU and Facebook decide to focus on using it to improve MRI scans?
You hear a lot these days about AI copying human performance or even replacing humans. One appealing thing about this project is that it is not about competitive data mining with AI. It’s not about putting AI in a competition with radiologists. Rather, it is about generating new possibilities for medical visualization.
AI can affect more than just interpretations of data—what we are seeing is that it can actually help us gather those data in new ways. That’s something we’re all very excited to explore.
In broader terms, why is it so important to boost the speed of MRI scans? What kind of impact will that have on healthcare as a whole?
Speed is one of the fundamental currencies of imaging in general. The faster you can go, the more information you can get. In biomedical imaging, information is health, and MRI is a remarkably rich source of information that can quite literally save lives. It can generate multiple complementary views of the body, looking at anatomy, function and cellular microstructure. The problem is, though, that gathering all of that information takes time. The exams typically take anywhere from 15 minutes to an hour or more, which can be a real impediment when dealing with patients with chronic illness or children who struggle to stay still for extended times.
Our collaboration’s proposition is that we can make MRI ten times faster. Such an acceleration would have multiple benefits. It would create a more comfortable patient experience, but it would also increase MRI accessibility in areas where there is limited access to scarce and/or oversubscribed MR machines. Faster MRI can also mean improved image quality—we’re talking about increasing the shutter speed, so that we can freeze out motion and see anatomy more clearly. This could also lead to reduced radiation exposure. MRI itself does not involve ionizing radiation, but if we can make it fast enough, we may be able to use it instead of x-rays or CT in certain cases. For example, when a patient comes in with complaints about their knee—a patient who would typically get an x-ray, which is negative in a majority of cases, followed by an MRI—we could skip the x-ray entirely and go straight to MR.
NYU and Facebook have indicated that there are plans to open-source this research as the work moves forward. Can you tell me a little bit about that decision?
One of the goals of our research, in addition to trying to solve the problem of reconstructing from under-sampled data, is to provide resources that allow other bright minds to tackle the problem. The idea is that we will open-source both our methods and the architecture we are using so that other researchers can work with them.
We’ll also open-source the dataset, giving people access to highly specialized data that are generally difficult to gather but is essential for the success of modern AI techniques like deep learning. If we create a repository that many people can use, we can improve research all around and lift up the entire field.
If your primary goal of improving MRI times is a success, what would be the next steps for this research?
CT is an obvious next step for this research. CT is already lightning-fast, but with AI, we could decrease the radiation dose associated with CT. We could also look at other modalities such as PET and ultrasound.
The bigger vision here is that AI can help medical imaging move away from emulating the human eye—taking snapshots of a scene—to emulating the brain, which processes a stream of data to derive actionable information. And once you have such data streams in mind, you can begin to reimagine imaging hardware itself. Things like this are beckoning to us on the horizon, even though they aren’t the immediate goals of this specific collaboration.
How do you think radiologists’ opinion of AI has changed over the last few years? Are you seeing more concern about it potentially replacing specialists? More acceptance?
I think the fear of being completely replaced by AI is abating in radiology. Things are much more collaborative as opposed to being competitive. AI can be used to help radiologists to do a better job in their core diagnostic and therapeutic responsibilities, and it can also take care of some of the routine tasks radiologists don’t necessarily like doing anyway.
There is obviously a lot of hype right now about AI in radiology, and I think the tendency is to either resist drinking the Kool-Aid or to run in the same direction that everyone else is running to avoid being left behind. I would argue we should not be afraid of this hype, nor should we follow it blindly. We are the curators of important unsolved problems and unmet needs in imaging. We should embrace the new tools that AI offers and, ideally, help to direct them toward areas of genuine value for human health.