A few months ago there were articles going around about how Samsung galaxy phones were upscaling images of the Moon using AI [0]. Essentially, the model was artificially adding landmarks and details based on its training set when the real image quality was too poor to make out details.
Needless to say, AI upscaling as described in this article would be a nightmare for radiologists. 90% of radiology is confirming the absence of disease when image quality is high, and asking for complementary studies when image quality is low. With AI enhanced images that look "normal", how can the radiologist ever say "I can confirm there is no brain bleed" when the computer might be incorrectly adding "normal" details when compensating for poor image quality?
> “A camera is supposed to take pictures of what it sees.”
Feels like that’s just a matter of expectations.
A phone used to be a device for voice communications. It’s right there in the Greek etymology, “phonē” for sound. But 95% of what people do today on devices called phones is something else than voice.
Similarly, if people start using cameras more to produce images of things they want rather than what exists in front of the lens, then that’s eventually what a camera will mean. Snapchat thinks of themselves as a camera company, but the images captured within their apps are increasingly synthesized.
(The etymology of “camera” already points to a journey of transformation. A photographic camera isn’t a literal room, as the camera obscura once was.)
Some of us want a record of what was, not a hallucination of what might have or could have been.
Courts, for example. Forensic science was revolutionized by widespread adoption of photography leading to a reduction of the importance given to witnesses. Who also hallucinate what might have happened.
So when I took an 8 second exposure of the aurora on Friday and then used Capture One to process the raw to make it more vivid than it was in real life - is that a record of what was?
Don’t get me wrong, I’m not super keen on AI type stuff in cameras as a whole. The line is muddy though. A smartphone camera straight up can’t capture the moon well, or at all. If it then looks more like it did in real life after processing is that better or worse than my above example?
How often do you capture auroras or other beauty shots, vs. readouts on your electricity meter, stickers on the back of your furnace, receipts, and a hundred other displays and documents you need to send someone? I definitely do plenty of the latter, and in such cases, I'd really appreciate the AI to not spice things up with details it thinks should be there.
I'm by no means against the feature. Hell, I shoot 90% of my family photos in Portrait mode on my Galaxy phone, which does some creative bluring and probably some other magic[0]. I just really appreciate being able to turn the magic on or off myself. That, and knowing exactly what the magic is[1].
--
[0] - I don't know what exactly it does, but switching from normal to Portrait mode is all it takes for my photos to suddenly look amazing, as judged by my wife, vs. plain sucking.
[1] - See e.g. "scene optimizer" in Galaxy phones. It's a toggle on normal photo mode. I have no first clue what it does, I can't see any obvious immediate difference between shots taken with vs. without that feature.
When I'm taking a picture of a receipt or sticker behind a machine, I don't actually want a literal photograph of the entire scene but just a reproduction of the text content.
Any environmental lighting, color and texture of the desk, and all other visual detail are only a distraction.
So if the camera would recognize this intent and just give me the receipt looking like it came from a scanner, that would in fact be a great improvement. So I think your example is in fact a point in favor of having AI meddle with most photos that people shoot.
There's a pretty famous example of Xerox's scanners using compression on the scans done...except the "compression" would sometimes substitute the wrong letter or number on text-based content. [1]
So I can foresee pretty straightforward problems with merely storing the text content from the images, to say nothing of any less binary "is it correct" questions.
I recently took a picture of a lizard on a granite with large grains. When I zoomed in to identify the type of lizard I saw that all the grains and some leaves on a tree had been simplified with some type of swirl. I find it unlikely those swirls were artifacts of the sensor itself. My assumption is the effect is related to compression given how often it repeated but I'm not sure.
> Some of us want a record of what was, not a hallucination of what might have or could have been
Yes, but that doesn’t imply “A camera is supposed to take pictures of what it sees”, only “cameras sometimes are supposed to take pictures of what they see”.
Some of us prefer a nice picture over a more exact record of what was; some of us will even argue that such manipulated pictures are better at capturing what was precisely because they sacrifice some of the physical reality for the non-physical essence of a moments one’s memory of such a moment.
That moon photo is a nice example. Smartphone cameras aren’t very good at capturing what the full moon looks like in our memory.
Pretty much all modern digital cameras are using heuristics and algorithms to construct the image you see - it's not just a sensor grid and a bitmap file and it hasn't been for a long time.
The important property is how the pixels are correlated with the physical reality being imaged - because the goal is to reason and learn about the depicted subject though information in the photo. Heuristics and algorithms for demosaicing, white balance, auto-brightness/dynamic range, lens collection, removing motion blur, etc. improve the correlation or improve our ability to see it. This is fine, though you need to be aware at time which properties of the image are to be treated as relative vs. absolute.
This is also a far cry from having your camera think, "I'm a consumer camera! Normies often shoot pictures of the Moon, so this fuzzy circle must be it; let me paste a high-resolution photo of the Moon there", or "gee, normies often shoot sportsball, so this green thing must be astroturf, and this grey blob is probably the ball", etc.
Big difference between a fancy interpolation algorithm that compiles to 500 bytes and another that takes many more orders of magnitude of space because it also contains data used to add details from what it thinks other similar photographs have
Fine, I expect an MRI to take pictures of what's actually going on in my body rather than inventing MRI-like images that can fool the radiologist into thinking I'm healthy when I'm not. Not sure why this is controversial.
In ancient and koinē Greek it meant both voice and sound (including the sound of instruments, which carries in modern instrument names like "saxophone" and so on).
They would say "ὀργάνων φωναί", "φωνὴ βροντῆς", "φωνὴ ὑδάτων" and so on for example.
Now that you mention it. I recently picked up a bottle of Red Vinegar with large pictures of red grapes on it. Naturally I assumed this was grape vinegar. How shocking it was to discover that this Chinese company was selling acetic acid mixed with food colors.
This is an excellent ponit, and I don't know where to exactly draw the line ("I know it when I see it"). I personally use "auto" (probably heuristic, maybe soon-ish AI-powered) features to adjust levels, color balance etc. Using AI to add things that are _not at all present_ in the original crossed the line into digital art vs photography for me.
I draw the line where the original pixel values are still part of the input. As long as you’re manipulating something that the camera captured, it’s still photography, even if the math isn’t the same for all pixels, or is AI powered.
But IMO it’s a point worth bringing up, most people have no idea how digital photography works and how difficult it is to measure, quantify and interpret the analog signal that comes from a camera sensor to even resemble an image.
There was the small complication of the fact that the moon texture that Samsung got caught putting onto moon-shaped objects in photos is, of course, the same side of the same moon.
> the moon texture that Samsung got caught putting onto moon-shaped objects in photos is, of course, the same side of the same moon.
Probably not exactly the same side and orientation. https://en.wikipedia.org/wiki/Libration#Lunar_libration: “over time, slightly more than half (about 59% in total) of the Moon's surface is seen from Earth due to libration”
Sort of, kind of, but not shot at the same time, and not at the same location.
I would object slightly less if they made a model (3D or AI) that captures the whole side of the Moon in high detail, and used that, combined with precise location and date/time, to guide resolving the blob in camera input into a high-resolution rendering *that matches, with high accuracy and precision, what the camera would actually see if it had better optics and sensor*. It still feels like faking things, but at least the goal would be to match reality as close as possible.
I wouldn't go that far to call it as a fraud, unless you call literally every phone-with-camera manufacturer these days a fraud. Then I agree as my trusty old nikon fullframe always catches only whats there, including noise and instability that modern phones handle easily.
People were commenting on that thread how apple phone ie mirrored only bunny within bigger picture of a bunny in the grass (thats rather hilarious 'bug'), and we all know how apple consistently removes all moles and wrinkles, changes completely skin tone and overall tonality like every single picture looks like its taken in the golden sunset hour. Ie that nasty samsung is much more truthful when it comes to this, including latest flagships.
That's outright lying too, IMHO much worse - moon is tidally locked so showing exactly same side with same features for millions of years, so they were adding details that are there, just impossible to see on non-stabilized tiny plastic lens&sensor combo in the night.
Making somebody 20 years younger, much prettier and changing their overall look on most important feature we humans have, doing it by default without any real option to turn it off, does a lot of long term body-perception damage in young folks.
> A camera is supposed to take pictures of what it sees.
You wouldn't like the picture of what it sees. The lens is just not big enough. Even the pro raw and other features that phone introduced apply processing.
Yeah but that's too hard, and you can just use "AI" to make cool photos instead. Who wants an actual camera when you can have something "like" a camera at half the price?
> Bad example. Autofocus makes changes to the light that goes into the camera, not just the data that comes out.
Arguably so do the AI things (i.e. they take multiple shots at different exposure and composit them).
The point is, we have ceeded manual control of photography to automated systems a long time ago. Most people are happy with that choice, as phone cameras serve a cultural function and are not scientific instruments.
All images taken with digital cameras have been filtered by a pipeline of advanced algorithms. Nobody ever looks at "what the camera sees". What kind of savage would look at an image before demosaicing the Bayer pattern? (Except from the people who work in demosaicing, of course.)
What a goofy point to be raised as many times it as it has been in this thread. All of that stuff serves the purpose of more faithfully emulating actual human vision.
Actual human vision works a lot more like the AI stuff then you think. Human vision is famous for filling in details that aren't there based on what you expect to see.
This is one aspect about machine learning models I keep discussing with non-technical passengers of the AI-hype-train: They are (in their current form) unsitable for applications where correctness is absolutely critical.
I don’t know enough to make absolute statements here, but deep learning models can beat out human experts at discerning between signal and noise. Using that to guess at data and then hand it off to humans gives you the worst of both worlds. Two error probabilities multiplied together. But to simply render a verdict on whether a condition exists I’d trust a proven algorithm.
Yes, pattern recognition is one of the applications ML shines at. Now the question was about using ML to extrapolate between sparse pixels and how much humans can rely on the added detail.
The goal would be to find a way to make ML extrapolate only pixels that really describe actual really present features and never imagining detail that wasn't there in the first place. Now I am no expert at the matter, but what I know of deep learning models they are really good at the latter as they basically make statistic guesses on what would be plausible.
Getting a plausible guess on what looks like a convincing answer works really well for answering a question. But the problem at hand is more like predicting the words someone said based on the first and last word in a sentence. Imagine a criminal case where the evidence is fragmented like that: I am pretty sure a LLM could give a convincing prediction here, but I am not sure how much you could rely on that prediction being reflective of what was actually said. I certainly wouldn't feel comfortable with a conviction the result of that prediction even if it was reflective of the ground truth in 90% of times.
There are a lot of models that are simply good at that without hallucinating nonsense. LLMs are a specific thing with their own tradeoffs and goals. If you have a ML model that says how much does this microscope photo look like an anomaly in this persons blood on a scale from 0-100 it can certainly do better than a human.
So if an AI fantasized your face into the extrapolated pixels of the evidence for a documented murder case you would be happy with the conviction, because on average it might be somewhat correct?
I don't wanna hurt anybodies feelings by stating that AI isn't a magical wand that makes everything better — but every technology has use cases at which it excels (e.g. pattern recognition) and use cases for which it is fundamentally unsuitable. If you try to screw on a nut using a hammer, that doesn't mean hammers suck, it means the user has a wrong idea what a hammer is capable of.
The point is: Don't be that person if you can avoid it.
The state of the art MRI stuff uses "compressed sensing" -- essentially image completion in some domain or another. Presumably, carefully designed to not hallucinate details or one would hope.
There isn't necessarily a particularly neutral choice here: the MRI scan isn't in the pixel domain, artifacts are going to be 'weird' looking-- e.g. edges that move during the scan ringing across the whole image.
I don't think we know what's in the black box here. It could be an equivalent relatively unopinionated regularizer ("the pixel domain will be locally smooth, to the extent it has edges they're spatially contiguous") or it could be "just look up the most similar image from a library and present that instead" or anywhere in between. :)
But they're using it to eliminate stray or environmental EMI from the RF signals. That might not create fake stuff at the voxel level. Depends on the specifics.
the future is already here, GE has put this into production. if you have remove the onerous constraint of being correct you can make some really crispy images! 9/10 radiologists. that was literally what the FDA approval process was, surveyed a bunch of radiologists to see what they preferred. no adults in the room.
The interesting, indeed concerning thing, is that problem is not only applied to medical machines and mobile phones but to zillions of daily used wearable devices such as smart watches, brain eeg (e.g. Muse), and others without adverting users that what they see (e.g. HRV) couldn't be interpreted easily by a computer program.
Not saying that we humans are always better but saying that we are believing in number and conclusions from apps created as-is.
It is just weird that papers like this can be published. "Deep learning signal prediction effectively eliminated EMI signals, enabling clear imaging without shielding." - this means that they have found a way to remove random noise, which if true, should be the truly revolutionary claim in this paper. If the "EMI" is not random you can just filter it so you don't need what they are doing. If it isn't random, whatever they are doing can "predict" the noise, they even use the word in that sentence. They are claiming that they can replace physical filtering of noise before it corrupts the signal (shielding) with software "removal" of noise after it has already corrupted the signal. This is simply not possible without loss of information (i.e. resolution). The images that they get from standard Fourier Transform reconstruction are still pretty noisy so on top they "enhance" the reconstruction by running it through a neural net. At that point they don't need the signal - just tell the network what you want to see. The fact that there are no validation scans using known phantoms is telling.
I'm a professional MR physicist. I genuinely think the profession is hugely up the hype curve with "AI" and to a far lesser extent low field. It's also worth saying that the rigorous, "proper" journal in the field is Magnetic Resonance in Medicine, run by the international society of magnetic resonance in medicine -- and that papers in nature or science generally nowadays tend to be at the extreme gimmicky end of the spectrum.
A) Many MR reconstructions work by having a "physics model", typically in the form of a linear operator, acting upon the required data. The "OG" recon, an FT, is literally just a Fourier matrix acting on the data. Then people realised that it's possible to I) encode lots of artefacts, and ii) undersample k-space while using the spatial information using different physical rf coils, and shunt both these things into the framework of linear operators. This makes it possible to reconstruct it-- and Tikhonov regularisation became popular -- so you have an equation like argmin _theta (yhat - X_1 X_2 X_3.... X_n y) + lambda Laplace(y) to minimise, which does genuinely a fantastic job at the expense, usually, of non normal noise in the image. "AI" can out perform these algorithms a little, usually by having a strong prior on what the image is. I think it's helpful to consider this as some sort of upper bound on what there is to find. But as a warning, I've seen images of sneezes turned into knees with torn anterior cruciate ligaments, a matrix of zeros turned into basically the mean heart of a dataset, and a fuck ton of people talking bollocks empowered by AI. This isn't starting on diagnosis -- just image recon. The major driver is reducing scan time (=cost), required SNR (=sqrt(scan time)) or/and, rarely measuring new things that take too long. This almost falls into the second category
The main conference in the field has just happened and ironically the closing plenty was about the risks of AI, as it happens.
B) Low field itself has a few genuinely good advantages. The T2 is longer, the risks to the patient with implants are lower, and the machines may be cheaper to make. I'm not sold on that last one at all. I personally think that the bloody cost of the scanner isn't the few km of superconducting wires in it -- it's the tens of thousands of phd-educated hours of labour that went into making the thing and their large infrastructure requirements, to say nothing of the requirements of the people who look at the pictures. There are about 100-250k scanners in the world and they mostly last about a decade in an institution before being recycled -- either as niobium titanium or as a scanner on a different continent (typically). Low field may help with siting and electricity, but comes at the cost of concomitant field gradients, reduced chemical shift dispersion, a whole set of different (complicated) artefacts, and the same load of companies profiteering from them.
Would it be easier to deploy devices like this to developing counties without the infrastructure to support liquid helium distribution? I imagine a much simpler device WRT exotic cooling and distribution of material requirements is a plus. Couple that with the scarcity and non-renewable nature of helium, maybe using devices like this at scale for gross MRI imagery makes sense?
The AI used here as I read it is a generative approach trying to specifically compensate for EMI artifacts rather than a physics model and it likely wouldn’t be doing macro changes like sneezes to knees, no?
Zero-boil-off "dry" magnets have been widely used for the last decade -- we engineered away the thousands of litres of liquid helium in exchange for bigger electricity bills and some added complexity (and arguably cost). They basically put the cryocompressor/cold head on a large heatsinked plate and use helium gas as a working fluid to cool it and through conduction the rest of the magnet. The supercon wire has a critical T/B/Ic surface and (to my knowledge) they essentially accept worse Ic in exchange for higher Tc.
The cold head vibration can introduce a bit more B0 drift per day, but it's not practically a problem.
Regarding artefacts, one of the other reasons that MRI rooms are expensive are the Faraday cages. They do help. Not just in terms of noise floors but because there tends to be a lot of intermittent RF transmission from people like paramedics. Did you know a) that the international mayday frequency is 121.5 MHz, b) that overhead helicopter flights may transmit with kW of RF on that frequency, c) that the larmour frequency of protons at ~2.9T is 121.5 MHz, d) that Siemens "3T" magnets are routinely around 2.9T, and e) the voltage of the signal you detect in MR is micro to millivolt at best? I've seen spurious peaks in spectra from this.
The DL method the paper talks about "may work", but as the OP says this is deeply unsatisfactory for a whole host of reasons and is, in my overly sarky opinion, a bit like fixing a wall with rising damp by putting a television in front of it showing a beautiful, high resolution picture of a brick wall in the same colour.
Perhaps a standard bit of kit for an imaging room ought to be a receiver at the operating frequency outside of the room that can pause the sequence when a potential jammer is active, and log the event so that you could potentially make a report to the relevant authorities (perhaps encourage them to keep the transmitters off near your facility).
Pausing the sequence is also not so much of an option when contrast was just administered either, I guess.
(I suppose if the signal weren't so hot that it was saturating the ADCs there might be some opportunity to subtract it off... but that's starting to sound like another ten thousand phd-educated hours of labour mentioned up thread)
Except there are other uses for an MRI and something that doesn’t require super conductors would be pretty awesome and deployable to places that lack the infra to support a complex machine depending on near absolute zero temperatures and the associated complexities.
> We conducted imaging on healthy volunteers, capturing brain, spine, abdomen, lung, musculoskeletal, and cardiac images. Deep learning signal prediction effectively eliminated EMI signals, enabling clear imaging without shielding.
So essentially, the neural net was trained to what a healthy MRI looks like and would, when exposed to abnormal structures, correct them away as EMI noise leading to wrong diagnostics?
I won't be very dismissive of this approach and probably deep learning has a strong role to play in improving medical imaging. But this paper is far, far from sufficient to prove it. At a minimum, it would require mixed healthy / abnormal patients with particularities that don't exist in the training set, and each diagnostic reconfirmed later on a high resolution machine. You need to actually prove the algorithm does not distort the data, because an MRI that hallucinates a healthy patient is much more dangerous than no MRI at all.
Seems like a huge and obvious red flag to me indeed. I can't imagine how the authors managed to not even mention the issue in the abstract. If the model is trained on healthy scans, well, yes, it will spit out healthy scans. The whole point of clinical radiology is to get enough precision to detect (potentially subtle) anomalies.
I don't think that (necessarily) says what you think it says.
You can read that as saying that the DL eliminated the background noise rather than saying that the system was conditioned on images of healthy people. From that it may well have been conditioned on just an empty machine or neutral test samples.
If so, there may be a good reason to suspect that it isn't likely to create artifacts that look like or mask anatomical structures.
You can read it like that, but they surely didn't prove it works like that and the burden of proof is squarely on them.
Realistically, the training set is most likely MRIs of similar tissues and would be naturally biased towards healthy structures. Even the remotest possibility of a hallucination should be addressed and disproved for such an application but they make no mention of it, just "OMG magic ENHANCE button!".
If the noise exists only on a certain frequency then the model would learn a passband filter of sorts and won't necessarily filter out abnormal structures. But they'd need to verify that.
I can’t access the full paper, but from the abstract, is it accurate that they’re using ML techniques to synthesize higher-quality and higher-resolution imagery, and that’s the basis for their claim that it’s comparable to the output of a conventional MRI scan?
Do clinicians really prefer that the computer make normative guesses to “clean up” the scan, versus working with the imagery reflecting the actual measurements and applying their own clinical judgment?
I can say that most radiologists would not want a computer trying to fix poor scan data. If the underlying data is bad, they would have recommend an orthogonal imaging abnormality. "I don't know" is a possible response radiologists can give. Trying to add training data to "clean up" an image would bias the read towards "normal".
Spot on. When I can't interpret a study due to artifact, I say that in my report.
Let's say there's a CTA chest that is limited because the patient breathed while the scan was being acquired, I need to let the ordering clinician know that the study is not diagnostic, and recommend an alternative.
If AI eliminates the artifact by filling in expected but not actually acquired data, I am screwed and the patient is screwed.
To nitpick, wouldn't it by definition bias the read toward normal? I suppose the problem is more that you don't want to bias it to normal if it wasn't.
As a practicing radiologist, I think this is great. We can have AI enabled MRI scanners hallucinating images, read by AI interpreting systems hallucinating reports!
I'm a radiologist and very sceptic about low-field MRI + ML actually replacing normal high-field MRI for standard diagnostic purposes.
But in a emergency setting or especially for MRI-guided interventions these low-field MRIs can really play a significant role. Combining these low-field MRIs with rapid imaging techniques makes me really excited about what interventional techniques become possible.
> This machine costs a fraction of current clinical scanners, is safer, and needs no costly infrastructure to run (2). Although low-field machines are not capable of yielding images that are as detailed as those from high-field clinical machines, the relatively low manufacturing and operational costs offer a potential revolution in MRI technology as a point-of-care screening tool.
I don't think this machine is being billed as replacement to high-field machines.
> I don't think this machine is being billed as replacement to high-field machines.
Countries where health regulation is less developed are likely to see misrepresentation where this form of MRI will be equated to full-field MRI by snake oil salesmen.
What is it about lower fields that means you cannot get a good image? Interference? Tissue movement in longer exposures? Why can't the device just integrate over a longer period of time?
It's just the physical reality of nuclear magnetic resonance. SNR scales with B^(3/2), since the signal scales with B^2 and the noise scales with root B.
This means going from 0.05T to 1.5T boosts your sensitivity ~150x. Measurement time scales with sensitivity^2, so you'd have to measure 20k x longer.
The application of a system like this could be as augmentation to imagers like CT and ultrasound. Because of its up resolution techniques and lower raw resolution (2x2x8mm), it might not be used for early cancer detection. But it looks really useful in a trauma center or for guiding surgery, etc. These same techniques could also be applied to CT scans, I could see a multi sensor scanner that did both CT and NMRI use super low power, potentially even battery powered.
Regardless, this is super neat.
> We developed a highly simplified whole-body ultra-low-field (ULF) MRI scanner that operates on a standard wall power outlet without RF or magnetic shielding cages. This scanner uses a compact 0.05 Tesla permanent magnet and incorporates active sensing and deep learning to address electromagnetic interference (EMI) signals. We deployed EMI sensing coils positioned around the scanner and implemented a deep learning method to directly predict EMI-free nuclear magnetic resonance signals from acquired data. To enhance image quality and reduce scan time, we also developed a data-driven deep learning image formation method, which integrates image reconstruction and three-dimensional (3D) multiscale super-resolution and leverages the homogeneous human anatomy and image contrasts available in large-scale, high-field, high-resolution MRI data.
The idea sounds great, but the examples they provide aren’t encouraging for the usefulness of the technique:
> The brain images showed various brain tissues whereas the spine images revealed intervertebral disks, spinal cord, and cerebrospinal fluid. Abdominal images displayed major structures like the liver, kidneys, and spleen. Lung images showed pulmonary vessels and parenchyma. Knee images identified knee structures such as cartilage and meniscus. Cardiac cine images depicted the left ventricle contraction and neck angiography revealed carotid arteries.
Maybe there’s more to it that
I’m missing, but this sounds like the main accomplishment is being able to identify that different tissues are present. Actually getting diagnostic information out of imagining requires more detail, and I’m not sure how much this could provide.
This is remarkable. 1800W is like a fancy blender, amazing to be able to do a useful MRI at that power.
For anyone who is unaware, a standard MRI machine is about 1.5T (so 30x the magnetic strength) and uses 25kW+. For special purposes you may see machines up to 7T, you can imagine how much power they need and how sensitive the equipment is.
Lowering the barriers to access to MRIs would have a massive impact on effective diagnosis for many conditions.
This reminded me of the recent request for startups proposal by Surbhi Sarna “A way to end cancer”. The proposal states that we already have a way (MRI) to diagnose cancer at very early stages where treatment is feasible but cost and scaling need to be tackled to make it widely accessible.
Something like this low power MRI could be a key part of enabling a transformation of cancer treatment.
I have a hard time picturing the radiologist whose reputation and malpractice rely on catching small anomalies being comfortable using a machine predicated on inferring the image contents.
There are some non-ML based approaches for ultra low field MRI that are starting to work: https://drive.google.com/file/d/1m7K1W--UOUecDPlm7KqFYzfkoew... . You can still add AI on top of course, but at least you get a better signal to noise ratio to start with!
It may miss some scans because there could be special cases which the model wasn’t trained with and would predict a different result/error. Maybe it’s acceptable in places where you may not even get a chance to be diagnosed
With a voxel size of 2x2x8mm^3, this would do what X-rays/CT's do now, and a bit more (but likely not replace high-energy MRI's? I'm not understanding how they rival high-energy accuracy in-silico, but that's how the paper's written)
In the acute setting, faster and more ergonomic imaging could be big. E.g., in a purpose-build brain device, if first responders had a machine that tells hemorrhagic vs ischemic stroke, it would be easier to get within the tPA time window. If it included the neck, you could assess brain and spine trauma before transport (and plan immobilization accordingly).
That's because people on Hacker News don't know shit. Everyone here who has a superficial understanding of deep learning and believes they know what an MRI is comes up with extremely strong opinions, but it is very obvious that noone here has even read the full article beyond the abstract (or even the title). There are legitimate questions and concerns you can raise about this article, but not a single one of them is found in this comment section.
This paper is actually the culmination of a series of developments in the field over the past decade. Anyone who was following the subject was not surprised by it. Yours and all the other comments here are nothing but a testimony to the presumptuousness of HN. It's also funny how posts about modern physics usually just yield admissions that people know nothing of the subject, but when the topic is AI or medicine, everyone is suddenly a research scientist.
I can't read the full article but low-T MRI is potentially a big deal IMO because a 0.05T magnetic coil can be air or water-cooled but higher T-magnets (like 1.5 and 3T MRI magnets) have to use superconducting wire and thus must be cooled to sub 60K temperatures (even down to sub 10K) using Helium refrigeration cycles. I worked for a time at a company that made MRI calibration standards (among many other things).
helium refrigeration cycle equals:
- elaborate and expensive cryogenic engineering in the MRI overall design.
- lots of power for the helium refrigeration cycle.
- requirements for pure helium supply chain, which is not possible in many parts of the world, including areas of Europe, North America, etc.
This problem has already been solved by the MRI machines at your local airport. Cost and performance are not the issues. For $25 your luggage gets an MRI that automatically differentiates between organic molecules in seconds. How easily could this be adapted for free annual screenings at the mall?
But that's not the objective, and so your research is doomed
The medical equipment industry will not suffer fools who don't understand 'regulatory capture' and 'rent seeking.'
Those hospital machines are expensive and rare for reasons that have very little to do with cost or performance
Wow, this seems like it could be a DIY project! I know people are complaining about the AI stuff but look at the images before AI enhancement. They look pretty awesome already!
8mm slice thickness isn't particularly at odds with what is commonly done on commercial machines, though usually there is a second transverse scan (which can't be readily fused due to patient movement).
But even if it were, plenty of interesting structures are many centimeters in size, a thousand fold decrease in costs from eliminating cryogenic / high power magnets could be very useful.
the structures are many centimeters, but I assume that the sort of anomalies you'd be looking for in a clinical scan aren't going to be that large.
if you had a fracture/tumor/damage-of-some-type that's small enough to fit between those slices and you didn't get the slices lined up just right the scan would miss it, no?
I think this could be useful as a starting point for diagnostics - a cheaper, lower-power device massively lowers the barrier to entry to getting an MRI scan, even if it's not fully reliable. If it does find something, that's evidence a higher-quality scan is worth the resources. In short, use the worse device to take a quick look, if it finds anything, then take a closer look. If it doesn't find anything, carry on with the normal procedure.
Is cost of machines really barrier? I can get MRI for $400-$500 as a self payer (Eastern Europe, i.e. if i just wanted it, not that doctor would say he wants it).
I read a paper few years ago about utilization rate, machine/service cost, how many machines per citizen/hospital... They were running day and night. Cursory glance at other countries also reveal sensible prices.
Unless it gets to a point of ultra sound machine(i.e. machine in a the consulting room a doctor can use in 10 minutes), I don't think it will decrease price much.
> applying machine learning to the output of a lower-power MRI device
So, we get worse SNR data from the device and then enhance it with compressed knowledge from millions of past MRI images? Isn't it like shooting the movie with Grandpa's 8mm camera and then enhancing and upscaling it like those folks on YouTube do with historical footage?
Awkward, I was expecting to be reading an article about Tesla moving into the medical industry.
Who's idea was it to name the unit of magnetic flux density after a car company?
This is worse than that time I ended up reading an article about a shallow river crossing
Low power MRI can be a salvation to people who have some metal inside their body. Of course imaging those parts might be still impossible, but maybe other parts can be imaged
"The lower-power machine was much cheaper to manufacture and operate, more comfortable and less noisy for patients, and the final images after computational processing were as clear and detailed as those obtained by the high-power devices currently used in the clinical setting."
300-1800W power draw seems impressive! It looks like standard machines are using something on the order of 25kW while scanning, which certainly sounds prohibitive for less developed infrastructure.
Sounds like that's more about using a cryocooler to minimize the helium used-- but presumably that requires keeping the coils in a particularly hard vacuum to adequately insulate them.
There is some research towards operating at liquid hydrogen temperatures -- but hydrogen has its own logistical challenges.
Not a big deal at all. You can make an "MRI" with literally zero magnetic field strength like this. Just make an AI hallucinate the entire super crisp image!
This is what this paper is basically doing it seems. "Look how clear the image is!" yea, because it's not real, it's AI generated garbage.
Medical imaging devices and medical devices in general are a racket. There are only a few companies and they are legal and lobbying departments first and foremost. This isn't the first time radical and radically cheaper prototypes have been proposed, but the unsolved bit it actually convincing anyone to buy.
A colleague had a device and a veteran adviced him to 10x the price.
You need a specialist to understand an MRI image. Maybe software will advance enough to change this, but it will be a slow progress. Also, carterls are definitely a thing. Radiologists will fight the software part.
Could you use the deep learning to improve the device to reduce the need for deep learning to fill in the gaps from traditional devices? Essentially teaching an algorithm to build a better, more simple imaging device in a pid loop?
Needless to say, AI upscaling as described in this article would be a nightmare for radiologists. 90% of radiology is confirming the absence of disease when image quality is high, and asking for complementary studies when image quality is low. With AI enhanced images that look "normal", how can the radiologist ever say "I can confirm there is no brain bleed" when the computer might be incorrectly adding "normal" details when compensating for poor image quality?
[0] - https://news.ycombinator.com/item?id=35136167