[a-dub]

woah. woah. woah. hold on a second here... are we comfortable enough with understanding all of the behavior of deep learning models to where we can confidently put them in the pipeline for diagnostic clinical imaging?

i'm okay with using them for image analysis, but denoising and other image production tasks seems dangerous. how do you know what you're looking at is real as opposed to something that just looks convincing? (like deep neural nets are famous for producing)

[l33tman]

This project doesn't use AI to improve the image, they use it to estimate the EMI noise from the surroundings. So they're not "filling in the gaps" in the actual resulting 3D voxel volume with fantasy voxels (which I hope will never ever fly in a clinical setting).

"To tackle the EMI signals from the external environments and internal low-cost electronics during scanning, we developed a deep learning driven EMI cancellation scheme"

So it's kind of using deep learning to improve the SnR in the RF reception. Of course this could theoretically also lead to "fantasy voxels" but due to the nature of MRI decoding, I'm willing to guess that bad predictions of the EMI interference will not show up as unnoticeable alterations of realistic tissue imaging but rather as artefacts all over the volume, like you normally see in clinical MRIs that weren't taken 100% optimally.

[ortusdux]

I'm glad that this is the approach that they are taking. There have been plenty of issues with fMRI false positives due to misconfigured software.

The most famous would probably be the IG Nobel winning study that detected brain activity in a store-bought salmon:

https://blogs.scientificamerican.com/scicurious-brain/ignobe...

https://www.discovermagazine.com/mind/fmri-gets-slap-in-the-...

Later studies called into question the results of between 10% and 40% of historic fMRI studies:

https://blogs.warwick.ac.uk/nichols/entry/bibliometrics_of_c...

https://www.pnas.org/content/113/28/7900

[prefrontal]

Thanks for the kind words. I am the first author of the "Neural correlates of interspecies perspective taking in the post-mortem Atlantic Salmon: An argument for multiple comparisons correction" paper. Happy to take any questions here. A link to the original poster: http://prefrontal.org/files/posters/Bennett-Salmon-2009.pdf

[BizarroLand]

It would be nice if the cost of an MRI was so low you would typically get a cheap one as part of your yearly physical and if anything popped up they could do it again in an expensive, high powered one to verify.

[prefrontal]

I've scanned about 300 people as part of my research career. The director of the imaging center reviewed every anatomical scan. From that group of 300 we informed about three people that they had an anomaly which should be examined by a doctor.

[BizarroLand]

Yeah, that's what I'm talking about. Sure, it was 1% that needed further validation but that 1% is so much cheaper and easier to treat when its caught early vs. later on when it's noticed by the patient.

MRI's becoming commonplace, even if it were every 3 years instead of annually would be a useful tool to improve health outcomes across the board.

[dtgriscom]

You run into the most entertaining people here. Got any good fish recipes?

[prefrontal]

Marc Abrahams, organizer of the Ig Nobels, asked us for a salmon recipe to include in a cookbook they were publishing. We sent in a single page recipe for how to cook a salmon in an MRI scanner by overriding the safety protocols. That was fun to write.

https://www.amazon.com/Ig-Nobel-Cookbook-1/dp/1939385164

[jacquesm]

> The most famous would probably be the IG Nobel winning study that detected brain activity in a store-bought salmon:

A store-bought dead salmon.

I am assuming that most salmons bought in stores are dead but that particular detail is rather relevant here.

Also that had me laughing, what a great move.

[robwwilliams]

Not sure the dead salmon is relevant. That paper is focused on false discovery in FUNCTIONAL MRI. Different can of fish. Most clinical work is structural MRI.

[jacquesm]

The chances of finding brain activity in a dead salmon are a bit lower than finding it in one that is alive.

[nichos44]

Those false positives are because fmri runs countless statistical tests and the earlier "misconfigured software" wasn't running stringent enough multiple comparisons corrections. Basically the same issue in the classic "jelly bean causes acne" xkcd (https://xkcd.com/882/). The exact number depends on voxel size, temporal resolution, and experimental condition but is somewhere close to tens of thousands of tests.

The "images" that are presented in fMRI studies and that contain false positives are representing results of statistical tests (t-values, and f-values after correction) not the contents of voxels. So the false positive rate of an fMRI has very little to do with the accuracy of a voxel's content in a structural MRI.

[WalterSear]

I prefer to think of that study as evidence for life after death.

[a-dub]

the primary innovation is using deep learning to denoise the signal and the cost savings derived from being able to use a noisier signal.

whether you call it "SnR improvement" or "additive noise cancellation", it is undeniably adulteration of the signal.

looking at the supplementary information, it looks like this paper was reviewed by mr-physicists. i think it also should have been reviewed by ml experts as well.

[heartbeats]

Sure, but it's not like it has the potential for overfitting. From my layman's understanding, the process is this:

1. Measure outside interference sources 2. Measure MRI of "nothing" 3. Use ML to estimate f(interference) = noise 4. Subtract estimated noise from signal

So the noise removal process has no awareness of brains, skeletons, etc.

[scratcheee]

Exactly my thoughts too. I'm fine with a simple noise-removal pass, but if the AI is context-aware, what's to stop it saying "hmm, this brain would look more like a normal brain if I remove these tumors". Obviously, they'll test for that, but that only handles common cases they concider, it's always going to be a risk for more unusual sceanrios, and the danger with altering the data is that anyone looking at the results wont have a way to tell how dubious that data is.

Reminds me of https://en.wikipedia.org/wiki/Xerox#Character_substitution_b... which was _so much_ worse than the equivilent OCR bug because it occured at the image level, where everyone expects errors to to produce noise, not contextly sensible and sharp _but wrong_ characters.

EDIT: based on other comments below, this is thankfully not the case, the AI just understands noise, it doesn't try to "fill in the blanks" based on how brains are supposed to look.

[robwwilliams]

The ML denoising is within-sample across voxels—-or so I presume from similar work in small animal MRI. And you can always have the “with” and “without”. I do not see any problem if a radiologist is in the review process.

[robertlagrant]

> a simple noise-removal pass

Even that is inventing data, no?

[scratcheee]

Yes, but context is key.

Denoising can on average improve the result, but sometimes it will be wrong.

Spotting when it goes wrong is potentially a difficult task, but generally the difficulty scales pretty clearly with the difficulty of understanding the original image anyway. If you can't spot when a denoising filter has screwed up, chances are you wouldn't have spotted anything interesting in the original image anyway.

But once an AI is context-aware things get way more complicated - it will try very hard to produce an image that doesn't _look_ wrong. Even if it goes wrong, it can go wrong and still succeed in managing to make an image that looks correct, it just no longer matches the real brain that was scanned. Perhaps it decided a tumor was just a smudge on the lense, and invented some brain to go behind it. An operator expecting to see brain and seeing brain wont think anything of it. When the patient dies, they may look back and say "wow, that tumor didn't exist at all just 3 days before! that should be impossible!".

tldr: Having an ai that might make mistakes is one thing, having an ai that can just invent exactly the data everyone is expecting to see is dangerous.

[heartbeats]

(reposting my comment) From my layman's understanding, the process is this:

1. Measure outside interference sources 2. Measure MRI of "nothing" 3. Use ML to estimate f(interference) = noise 4. Subtract estimated noise from signal

So the noise removal process has no awareness of brains, skeletons, etc.

[tshaddox]

Hopefully the system would be trained to accurately convey relevant medical information rather than to generate an image of a brain that looks normal.

[xattt]

Therein lies the dilemma of this technology: would a scanner that might sometimes substitute information be better than no scanner at all?

[jart]

How do you know? They're both based on neural networks. JBIG2 was also responsible for the Pegasus FORCEDENTRY thing.

[VikingCoder]

That's a valid risk.

You're asking a cost-benefit question.

The cost of an invalid diagnosis is indeed high.

The cost of no diagnosis at all is also high.

This device will not replace the MR at your local hospital. It will be the first MR device in hospitals that have never had one before.

[lostlogin]

Veterinary MRI may be another application.

[VikingCoder]

Good point. Cargo inspection? Luggage inspection? I dunno.

[lostlogin]

There are a lot of applications that are surprising. I’ve scanned for salmon farmers (is that the term?) who want to check they are breeding good fish and are looking at spine alignment.

I’ve scanned logs for forestry managers who want to look at something in their trees.

I’ve scanned old hearts that have been sitting in formalin for decades.

All are probably better at higher field strength but maybe some of that can be compensated for by scanning for longer? A log isn’t going to move, and a dead fish scan is likely only limited by the time it takes for it to rot.

I’d be scared of scanning unknown things, it might be a low field MRI, but it’s still a big magnet.

[petra]

It's better than nothing. Let's start with those situations, and slowly build a database proving or disproving this technology.

[jcims]

>However, MRI accessibility is low and extremely inhomogeneous around the world. According to the 2020 Organisation for Economic Co-operation and Development (OECD) statistics, there are approximately 65,000 installations of MRI scanners worldwide (~7 per million inhabitants)

Given that my little podunk hospital in the midwest seems to have roughly 5x the worldwide average number of MRI machines, totally agree.

[lostlogin]

MRI is also a massive revenue generator. That's a key reason they buy them.

[axg11]

Whether you’re comfortable with it or not - it’s already happened and in production. Look up Subtle Medical and GE AIR Recon.

[lostlogin]

Literally just purchased this - Installed tomorrow: https://www.siemens-healthineers.com/magnetic-resonance-imag...

[thfuran]

What's the worst that could happen?

https://www.dkriesel.com/en/blog/2013/0802_xerox-workcentres...

[mattkrause]

Amen!

I have seen an alarming number of talks where someone proposes to algorithmically add Gado contrast or turn a T1 into a T2 image. In a few very specific contexts, this makes sense (e.g., aligning a T1 taken in one session with a T2 taken in another). Otherwise though, it seems dangerous to mistake a "real" image with the expected image given another one.

[lostlogin]

If reducing gadolinium dose is the aim, a more prompt following of the literature, radiologist request (rather than surgeon demand) and weight based dosing would drastically reduce dosage. A moaning radiographer, what a surprise!

[phkahler]

I'm kind of thinking the opposite. Image analysis is where you don't want AI. Noise removal is further upstream (I'm assuming) and if it fails wouldn't it cause significant artifacts (blur for example) in the images?

It would be helpful to see results with and without this correction, or even with varying degrees of it.

[MauranKilom]

Agreed, but I believe "using AI for inference from sparse observations" is unfortunately a thing already.

[VikingCoder]

That's not "unfortunate."

Sparse observations save lives. A quicker MR. Less X-Ray exposure.

It's totally valid to worry about validation, but to the degree you can validate image processing algorithms of any kind - AI or otherwise - they absolutely save lives.

[aidenn0]

NMRI use microwaves, not x-rays.

[lostlogin]

Yes, and when a quick MRI is available, it can remove the need for a CT. Fast brain protocols are now less than 5 minutes. This makes things practical that weren't before.

[VikingCoder]

I was talking about MR and CT. Applies to PET, too.

Image processing saves lives.

[lostlogin]

It's not unfortunate, it improves acquisition times and image quality - I use it daily.

[fluidcruft]

Deep learning reconstructions are already marketed/sold in high-end commercial scanners.

[iancmceachern]

Yes. This is past tense, other companies are already doing this, in the clinic.

[lostlogin]

A link: https://www.siemens-healthineers.com/magnetic-resonance-imag...