How AI Is Accelerating MRI Scans

Imagine you’re at the hospital, in pain and anxious about what diagnosis might be coming. Your doctor orders an MRI scan, which means waiting for an available time slot so you can lie perfectly still inside the scanner’s narrow tube for up to an hour. As the MRI machine gradually collects data, you have only the intermittent crackling and banging of the electric current in the scanner’s magnetic coils for company. If you move during the scan, the image might not be clear enough to be useful, which would mean booking another appointment to come back and do the whole thing over again.

MRIs are often the best tool for diagnosing problems with organs, muscle, and other soft tissue. But even with recent advances, it takes a significant amount of time for the scanner to gather the necessary data. That is difficult for anyone, and it can be impossible for the very young or the seriously ill.

The time it takes to complete an MRI scan doesn’t just make the patient experience more grueling. It also limits how many people can be scanned in a given day. And some types of tissue are in constant motion as the scan takes place, so an image that takes a long time to generate can sometimes be too blurry to be useful. What’s more, when doctors need information quickly, they often must use other technology instead of waiting for the MRI scanner to do its work. X-rays and CT scans are much faster, but unlike MRIs, they expose the body to ionizing radiation. And with some types of tissue, MRIs can reveal more detail than alternatives.

Facebook AI researchers have partnered with doctors and medical imaging experts at NYU Langone Health to solve this problem and advance artificial intelligence research. We are using AI to create complete images from far less raw data. Since collecting that data is what makes MRIs so slow, this has the potential to speed up the scanning process significantly. So one day in the hopefully not-too-distant future, you might spend just a few minutes in the scanner’s tube to generate a crystal clear image.

After two years of work on this fastMRI initiative, Facebook AI and NYU Langone have reached an important milestone. A new clinical study to be published in the American Journal of Roentgenology shows for the first time that fastMRI images are interchangeable with those of regular MRIs. The study focused specifically on knee scans, and we are now working to extend the results to other parts of the body.

“This is an important step toward the clinical acceptance, and utilization of AI-accelerated MRI scans,” said Dr. Michael P. Recht, Louis Marx Professor and Chair of Radiology at NYU Langone Health.

Radio Waves, Magnets and Math

To understand fastMRI’s approach, it helps to first review how MRIs work.

To create the image that your clinician or radiologist reviews, the MRI machine uses magnetic fields that interact with hydrogen atoms in the body’s soft tissue and vital organs. Those atoms then emit electromagnetic signals that act like beacons, indicating where in the body the atoms are located. The signals are collected by the scanner as a sequence of individual 2D frequency measurements, known as k-space data.

Once all the data is finally collected, the system then applies a complex mathematical formula — an inverse Fourier transform — to that raw k-space data to create detailed MR images of the knee, back, or brain, or other area of the body. Without a complete set of data points, the math can’t pinpoint exactly where every signal comes from.

AI Meets MRIs

The fastMRI team used an entirely different way to create an image — one that requires far less raw data. The researchers built a neural network and trained it using the world’s largest open source data set of knee MRIs, which was created and shared by NYU Langone Health and as part of the fastMRI initiative. (The fastMRI data used in the project, including scans used for the study, are from the open-source dataset that NYU Langone created in 2018. Before open-sourcing the data, NYU Langone ensured that all scans were de-identified, and no patient information was available to reviewers or researchers working on the fastMRI project. No Facebook user data was contributed to creation of the fastMRI data set.)

The fastMRI research team removed roughly three-fourths of the raw data in each scan and then fed the remaining info into the AI model. The model then learned to generate complete images from the limited data. Importantly, the images produced by the AI model didn’t just look like generic MRIs; the AI-generated images matched the ground truth image created by the standard slow MRI process. Imagine taking only 250 pieces of a 1,000-piece jigsaw puzzle and then completing the entire image in a way that doesn’t just look plausible; it also exactly matches the complete puzzle shown on the box. That’s a rough approximation of what the fastMRI team was able to do with their model.

The fastMRI approach is different from other attempts to use AI in medicine. Often these algorithms aim to automate the review of medical images to try to spot potential problems, like a doctor would. But fastMRI does not try to do medical experts’ jobs for them; rather, it creates a complete image out of sparse information. Radiologists and clinicians can then use the fastMRI image just as they normally would. The only difference is that the patient spends less time in the scanner’s tube.

How Radiologists Put FastMRI to the Test

The researchers behind fastMRI had to make sure their model didn’t sacrifice accuracy in the pursuit of speed. Just a few missing or incorrectly modeled spots in an image could mean the difference between finding a torn ligament or a possible tumor, and giving patients an incorrect all-clear report.

GIF comparing a traditional MRI to fully sampled k-space

The clinical study to be published in the American Journal of Roentgenology  demonstrates that fastMRI’s AI model does indeed produce images that are just as accurate, useful, and reliable as those from a standard MRI. The study shows that fastMRI can generate “diagnostically interchangeable” MRI images of knee injuries while using about 75% less raw data from the scanning machine. In fact, the expert radiologists who participated in the study were unable to distinguish the AI-accelerated images from conventional ones. (More details are available in the study and this post on the Facebook AI blog.)

GIF comparing fastMRI and undersampled k-space

Building Toward Faster MRIs for Everyone Who Needs Them

Today’s clinical study is an important step forward, but there are many more advances to come. Next, Facebook AI and NYU Langone researchers want to show that fastMRI works just as well with other vital organs, such as the brain. FastMRI has also published its data, models, and code so that other researchers can build on their work and contribute new ideas. The fastMRI team hopes this open approach will speed progress and lead to new ways to use AI to accelerate MRI scans. What’s more, since we have shared our models openly, MRI manufacturers are free to test fastMRI with their machines right now, and to bring the resulting advantages quickly to patients.

There’s more to do for fastMRI. But one day soon, AI-accelerated MRIs may benefit millions of people around the world.

To help personalize content, tailor and measure ads, and provide a safer experience, we use cookies. By clicking or navigating the site, you agree to allow our collection of information on and off Facebook through cookies. Learn more, including about available controls: Cookie Policy