Facebook and NYU Set Out To Develop AI-Powered 5-Minute MRI Scan 30
Researchers at NYU Langone Health and Facebook's artificial intelligence division have teamed up to develop an AI model that uses less data and creates images faster than traditional MRI techniques, according to a Wall Street Journal report. From a report: The goal of the project is to create a five-minute MRI as an alternative to the 20 minutes to an hour it takes for current MRI machines to scan a patient, Michael Recht, MD, told the publication. Dr. Recht is professor and chair of New York City-based NYU Langone Health's radiology department and also a co-author of the research project. The combination of AI and MRI technology aims to construct images with less data rather than diagnose a medical condition. The project uses different technology and standards than those used to create AI-generated or synthetic media. Because it centers on constructing MRI scans, Facebook said the project must create images "that are accurate to the ground truth," compared to synthetic media, which usually needs to create a believable image, according to the report.
For the experiment, researchers created 108 patient images using standard MRI techniques as well as a second set of images in which some of the image data was thrown out. Facebook's AI model was then applied to construct the images with less data. Researchers used commercially available MRI machines, and data was collected from patients from various points of their bodies. Six MRI readers reviewed both sets of images, and readings were spaced out across a four-week period to ensure the readers could not recall important details from previous sets. Dr. Recht told the Journal that all six engineers concluded that the quality of the AI model-generated images was "as good [as] or better" than the conventional images. The AI system still needs regulatory approval, but NYU Langone is now using it to treat patients as part of an institutional review board study, according to the report.
For the experiment, researchers created 108 patient images using standard MRI techniques as well as a second set of images in which some of the image data was thrown out. Facebook's AI model was then applied to construct the images with less data. Researchers used commercially available MRI machines, and data was collected from patients from various points of their bodies. Six MRI readers reviewed both sets of images, and readings were spaced out across a four-week period to ensure the readers could not recall important details from previous sets. Dr. Recht told the Journal that all six engineers concluded that the quality of the AI model-generated images was "as good [as] or better" than the conventional images. The AI system still needs regulatory approval, but NYU Langone is now using it to treat patients as part of an institutional review board study, according to the report.
I'd prefer dying (Score:4, Interesting)
...than being scanned my Facebook.
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
does it fit in a salt shaker
Fine print (Score:4, Interesting)
Great.. now I need to ensure the fine print of any future MRI scan doesn't include a passage for Facebook to freely harvest and utilize this data.
Who am I kidding. They'll make me create a Facebook account during the check up.
Re: (Score:2)
They'll create it for you if they haven't already.
Not used to diagnose conditions? (Score:2)
The combination of AI and MRI technology aims to construct images with less data rather than diagnose a medical condition.
Er.. what's the point then? Or is it an MRI machine for Facebook to invade your physical privacy rather than just your personal data?
Re: (Score:3)
It's not very clearly worded. What they mean is the AI is used to construct an image, and then a human does the diagnosis.
As opposed to using the AI to perform the diagnosis, which is something we have seen people trying to use AIs for.
Re: (Score:2)
5 minutes in the MRI machine vs 20. It means more throughput, less time and exposure for the person, and overall lower costs.
Re: (Score:2)
Less exposure is a non-issue. You can go to 9.1T with no ill-effects, and even at 12T you just get hallucinations. Which is probably what the patient will try for later on that day anyway.
The question is accuracy. MRI scans, as popularly done, are low accuracy due to the low resolution they're working at. What we need are high accuracy, high resolution scans that are done faster. If you could do a 12T scan in 5 minutes, then the ill-effects might very well be negligible.
Re: (Score:2)
Except I'm sure they will charge more for it because of "technology".
Robotic surgery is faster and better for some things than a human but it costs a lot more.
Re: (Score:2)
It links your brain in as an app, so that other humans and advertisers can post directly to it.
Advances or regression? (Score:1, Troll)
Facebook? (Score:5, Informative)
I guess this is news because, Facebook. But there has been work on this several several years. It's called MRI compressed sensing. A lot of work has been done with cardiac MRI in particular as it takes the most time. With compressed sensing the scan can be done in 10 minutes rather than an hour. But it takes 12 to 24 hours to reconstruct the images into something usable.
The time it takes is a major issue in the US as in the past a typical cardiac MRI can take an hour or more. While just about any other body part can be scanned in 5 to 15 minutes. In the past scanning a knee got reimbursed at the same rate as a cardiac exam, it becomes a no brainier for which makes a hospital the most money. Pediatric CMR is where most of the work has been done since those scans and take 1 to 2 hours currently.
CT scans went through this years earlier as the need for radiation dose reduction was a major factor. No time was really saved in the scan, but the radiation exposure was greatly reduced. This gave very noisy images that were not diagnostic quality. But after running them through the necessary algorithms, they could be used and the radiation dosage was cut down significantly.
Re: (Score:1)
Re:Facebook? Buzzword Bingo? (Score:2)
Re: (Score:2)
Re: (Score:2)
That'd be my concern, such a system is no good if it's just filling in the gaps in a creative way so as to make scans that look real rather than scans that have the necessary accurate representation of the patient's body part. And it has to do this consistently, it's likely no good if it's accurate 90% of the time but hides important information 10% of the time.
MRI compatibility with Cochlear implants (Score:3)
Could be quite useful for people with Cochlear implants who want to minimize the time they're in the MRI machine. While any vaguely modern Cochlear implant is rated for typical 1.5T MRI's with magnet in place(except for Advanced Bionics, who were way behind in MRI compatibility), some patients report pain and discomfort anyways.
Good Luck In North Carolina (Score:2)
https://reason.com/2018/07/31/... [reason.com]
The state of North Carolina requires medical service providers to go through an arduous application process to prove that they need an MRI machine before they are allowed to buy or lease one. Need, mind you, is determined not by how many patients are asking for services, but rather by how many MRI's the state's Department of Health thinks an area requires.
lowkey ENHANCE! (Score:2)
The combination of AI and MRI technology aims to construct images with less data
researchers created 108 patient images using standard MRI techniques as well as a second set of images in which some of the image data was thrown out. Facebook's AI model was then applied to construct the images with less data.
Isn't this the equivalent of motion based image enhancement? Should we be worried this tech is from Facebook?
How does this compare to... (Score:1)
How does this compare to existing methods like https://www.max-planck-innovat... [max-planck-innovation.de] ?
I believe I've heard the people around Frahm have also looked into AI, but don't know if this is already part of FLASH2.
Extremely dangerous! It *makes the data up*! (Score:2)
It does not "construct" the data! The data comes from nothing! Meaning it is *guessed*. Aka *made up*!
No way in hell I or any doctor will judge whether you got cancer from *made-up* data!
IMO, not telling the doctor it's made up, while a cancer was not visible as a result, should be interpreted as *murder* by a judge.
With enough AI... (Score:1)
With enough AI, you can hallucinate most anything.