Facebook and NYU researchers discover a way to speed up MRI scans

This post was originally published on this site

Our mission to help you navigate the new normal is fueled by subscribers. To enjoy unlimited access to our journalism, subscribe today.

Researchers at Facebook and New York University have found a way to significantly reduce the time it takes to capture magnetic resonance images, in a breakthrough with potential to transform medical imaging.

By speeding up the process, the new technology could make easier to use magnetic resonance imaging, or MRIs, on patients, such as young children, who have trouble sitting still for long periods inside the giant cylindrical machines used to produce the scans.

Currently, scans frequently take more than half an hour, and children are often sedated or anesthetized to help doctors capture good images. The new technology could also eventually make MRIs more useful in helping doctors investigate acute conditions, such as strokes, where time is critical.

By increasing the number of patients who can be scanned in a day, the technology also has the potential to increase access to MRIs, particularly in developing countries that can afford few of the expensive scanners—which can cost between $1 million and $3 million each.

Mike Schroepfer, Facebook’s chief technology officer, told Fortune that the company is committed to open-sourcing the new technology so it gains wide acceptance. “Tech has been through a lot of fair criticism in the past few years,” he says. But he said the MRI project—on which Facebook and NYU worked for two years—was an example of how technology continued to hold the “possibility to improve a lot of lives” and why he remained excited about working in the industry.

Magnetic resonance imaging uses a powerful magnetic field to align protons in a patient’s body in a certain direction and then uses radio waves to momentarily knock these protons out of alignment. It then analyzes the energy released as these protons snap back into alignment with the magnetic field. Different tissue types emit slightly different energies during this process and software converts these energy differentials into pixels of differing brightness, resulting in an image. Sometimes doctors inject patients with benign chemicals that enhance the contrast visible in the image.

The Facebook and NYU researchers achieved the MRI speedup by using artificial intelligence to reduce the number of radio pulses the scanner needs to produce an accurate image. The process works a little like “connect the dots,” where the software learns how to accurately fill in the missing pieces of information between the data points generated by the scanner.

Many researchers and companies have been trying to create A.I. software that can automatically make diagnoses from medical images. But this research is an example of using A.I. to make human radiologists more efficient without trying to replace or even augment their diagnostic expertise. It is simply about getting the MRI scanner to generate an accurate image faster. And yet, as with many deployments of existing A.I. techniques, the business impact can be transformative.

Michael Recht, a doctor specializing in musculoskeletal imaging at NYU Langone Health who participated in the research, said the entire MRI field has spent decades trying to make the scans faster. But past attempts to improve speed by gathering less data have resulted in lower quality images, making diagnoses of certain conditions, such as small ligament tears, more difficult.

In its initial research, published today in the peer-reviewed American Journal of Roentgenology, the Facebook and NYU scientists looked specifically at 108 MRIs of human knees from people with a range injuries and conditions and found they could generate the scans four times faster without any loss of diagnostic accuracy. What’s more, a panel of six expert radiologists judged the A.I.-produced scans indistinguishable from traditional scans. In some cases, Recht said, doctors preferred the A.I.-generated images because they were less likely to contain ghostly visual artifacts that tend to be caused by people being unable to hold still enough for the lengthy scanning procedure or by the circulation of their blood during the scan.

Recht said researchers now plan a much larger study at multiple hospitals to confirm their findings. They have also already begun conducting research applying the same technique to MRIs of other parts of the body, such as the brain, and believe they may be able reduce the amount of time the scans take by even greater amounts, gathering images in as little as an eighth or a tenth of the time. This is because, Recht said, the pathologies in other body parts tend to occupy a greater number of pixels in the MRIs compared to those in the knee, where evidence of damage can be just a few pixels in length. That means it requires more radio frequency pulses from the scanner to achieve a diagnostically-useful image of the knee using the A.I.-technique.

Initially, the researchers thought they could use an A.I. method similar to the one used for deepfakes—highly credible-looking fake videos, often depicting someone’s head on another person’s body or depicting someone saying something in someone else’s voice—to artificially generate the missing data, Larry Zitnick, the Facebook researcher who helped spearhead the project, said. But the scientists soon realized the method only created great looking scans that were also highly likely to be inaccurate.

So the researchers employed a variation on a different A.I. technique that is often used in video compression, in which software learns to take a high-definition image, degrade it into a lower quality representation, then re-construct the high-definition version from the lower quality representation. Except, Zitnick said, rather than doing this for a two-dimensional visual image, the A.I. system in this case is working with raw radio frequency data and, learning to fill in the missing data points, and then translating that into a three-dimensional visual image.  

Zitnick said that while the software learns some general information on how to generate MRI data, for accurate results it needs to be trained on imagery of one specific body part only. He said he envisions manufacturers creating different software modules for each portion of the anatomy that is typically scanned.

Schroepfer said Facebook had been interested in MRI technology because the company had existing expertise in computer vision—A.I. systems that can classify and manipulate images. The company has used the technology in everything from the feature that automatically tags your friends in photos on Facebook to ones that let people search for similar images on Instagram to software the company uses internally to try to prevent terrorist videos and other extremist content from being uploaded.

But, he said, the company was looking to apply that expertise to a new domain. He also said that “having an impact in the real world” and “using A.I. for good” were important motivators for Facebook’s team of A.I. researchers.

The Silicon Valley giant employs one of the world’s leading stables of machine learning experts, and keeping them happy to come to work is an important business goal, Schroepfer said. But, he also said, by working on areas outside Facebook’s usual social media businesses, the researchers were more likely to encounter fresh thinking and insights that could eventually prove useful for Facebook’s business too.

More must-read tech coverage from Fortune: