A low-cost and shielding-free ultra-low-field brain MRI scanner
nature.com
nature.com
This is fantastic. What a sentence to get to write.
I'd love for the expensive ones to be disrupted. The dream is we get some attachment for our smart phones and turn them into Tricorders.
https://www.washingtonpost.com/news/wonk/wp/2013/03/15/why-a...
Noninvasive diagnostics should not be regulated.
There are usually a few radiologists lurking here and they would have better knowledge than me (I’m an MR tech).
https://appliedradiology.com/articles/diagnosing-brain-metas...
PET scans are limited in resolution when you get down to the sub-5 mm or so range due to scanner technology and fundamental limits of the physics of positron/electron annihilation & photon emission.
A typical MRI (i.e. alas, not what this article is describing) can usually resolve something at that size and identify characteristics like diffusion restriction or contrast enhancement which can confirm metastasis.
Also, in the brain, PET scans (at least the most common, FDG, which is based on glucose) are extremely limited in utility because of the baseline high glucose metabolism of the brain, which makes it hard to distinguish from the metabolic activity of a tumor.
Have you a link to something as my understanding is that small lesions are better found with MR?
Or is this a rule that applies to high end research work and hasn’t hit clinical practice yet? Maybe a limitation of the isotopes used clinically?
FDG PET/CT is not used to stage intracranial metastases due to background brain activity significantly reducing sensitivity. You’re likely only detecting lesions 1cm or greater in the brain, or ones which demonstrate low metabolic activity and appear dark.
MRI is much more sensitive and specific and is the standard of care for staging in my practice and as per the NCCN clinical guidelines. I haven’t been to any institution or heard of one where PET is used for staging of brain metastases.
Not sure where this is coming from.
Modalities like PET can detect events from a far greater fraction of the isotopes present. But they are limited in spatial resolution by the physics. This is why MRI is said to be a very insensitive modality.
See for example the discussion about MRI vs other modalities here:
https://wikipedia.org/wiki/Molecular_imaging
which notes that MRI has a sensitivity around 10^−3 to 10^−5 mol/L whereas PET is many orders of magnitude more sensitive at 10^-11 to 10^-12 mol/L.
So what I meant was if there's an analogy that you're looking at a skyscraper, MRI can "see" which which rooms have floodlights turned on. PET could detect whether there's a single candle burning somewhere in the building, but it can't tell you which room it's in.
16 years ago he had some kind of brain cyst. Totally benign. But as a follow up, the doctor ordered yearly MRIs, much to his annoyance.
Last year those yearly routine MRIs spotted a brain tumor- before it had time to get dangerous or cause any damage to him. It was growing quickly though and was right near his eye. Quick surgery got it out.
I want to live in a world where everyone has access to that.
That said, I’d like a go. I suspect that with more samples (more time) you could get the resolution up.
It sounds like different modes of taking (or interpreting/visualizing?) an MRI.
I really liked ‚MRI made easy‘ as an introduction to MRI physics. Just google it, it’s a free Book
https://rads.web.unc.edu/wp-content/uploads/sites/12234/2018...
T2* effects increase with higher MRI main field strength. From what I can tell so far these ultra low-field scanners have to rely on spin echoes.
The whole book is available for free as a download. MRI Made Easy (… Well almost). https://rads.web.unc.edu/wp-content/uploads/sites/12234/2018...
For some reason it reminded me of the crazy project where some team used the earths field as the static field and just added gradients and the RF stuff. I can’t find the article I remember but this project looks similar and scans a capsicum rather than an apple.
https://www.researchgate.net/publication/6956005_A_practical...
It looks like the bulk of the code for training the EMI elimination model is is pretty straightforward PyTorch.
One reason Matlab is popular in academia and industry is because someone who don't know C/C++ (mechanical engineers etc.) can use something like Simulink to program real-time systems on microcontrollers and FPGAs.
I think a bigger reason MATLAB is popular is that it works really reliably, it is very well documented, it has a ton of complex algorithms and functions built in (or available in toolkits at least) and the interactive GUI works very well.
There isn't anything anywhere that comes close to MATLAB's plotting abilities. And scientists do a hell of a lot of plotting.
> The lead author, Dr. Craig Bennett, wanted to get something fresh, so he headed in to the grocery story first thing in the morning. At the fish counter, he spoke the words that will echo down the centuries as a testimony to the dedication and drive of neuroscientists throughout the ages:
>"I need a full length Atlantic Salmon. For science."
That reminds me of the day that I needed a strong lightweight cable for a silica-fiber melting/drawing apparatus. After some puzzling, I realized that bicycle shift/brake cabling would probably be perfect for the task.
I'll never forget the puzzled look at the bike shop -- "What kind of bike are you putting it on?" "I'm not, I just need some brake cable for a science experiment...." As I recall, I think we finally settled on some precut cabling for a GT Zaskar of some kind.
Similar things came up the day that I needed a valve that switched faster than our dedicated micro-switching valves. A similar light-bulb went on, and I went down to the nearby auto shop for a fuel-injector.
"What kind of car do you need it for?"
"I don't, but there are a couple of different valve-switching protocols, some that latch open and others that accept straight TTL at reasonable currents. I need one of those."
That experiment was brought to you by an injector for, I believe, a Dodge Caravan, and later, when I needed another, an injector for a Ford Mustang. Fuel-injectors are really good valves.
The solution was a $12 ride sensor from a Cadillac SUV of some type. Worked perfect!
[1] https://twitter.com/TLAlexander/status/1480332181285212160
[2] https://github.com/tlalexander/large_format_laser_cutter
[3] https://twitter.com/TLAlexander/status/1489519692712538113
Computation is rarely the problem.
It's all about sensors and actuators.
One of the best college projects ever was to make a "drink pouring machine". You have a half-dozen bottles of various alcohols and other liquids, a circular table, and a glass. Now, pour a martini. Now, pour a gimlet. etc.
What a nightmare! The liquids have different viscosities and consequent pour rates. And shutoff is rarely clean. And don't rotate too fast or you will mess up your glass position. It goes on ... and on ... and on.
Everyone who did that project came out ... changed.
Sports physio / trainers would kill to be able to do regular MRI on their athletes. Being able to do pre/post workout imaging, and the kind of training programs this level of visibility would unlock, are quite exciting.
Assuming you can generalize from brain to whole-body, I think you could sell one of these to every major sports team in the country, and making it a non-medical device (ie skipping the FDA) would let you iterate much faster. A couple more halvings in price and this is accessible to every sports physio office and gym in the country.
Very cool!
This might have some interesting industrial applications as well (checking composites for defects, food packaging, etc). There might be some other applications that are currently unknown as MRI machines are too expensive (eg: checking beams in bridges).
I interpret musculoskeletal MRI studies in my practice. Small joint (what this would be used for) ligaments and tendons are incredibly difficult to see and assess even at ultra high res studies performs on 3T magnets.
With the spatial resolution of this you would be way better off using ultrasound.
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)
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.
Image processing saves lives.
"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.
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...
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.
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.
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.
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.
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.
Given that my little podunk hospital in the midwest seems to have roughly 5x the worldwide average number of MRI machines, totally agree.
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.
Even that is inventing data, no?
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.
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.
It would be helpful to see results with and without this correction, or even with varying degrees of it.
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.
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.
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.
https://www.dkriesel.com/en/blog/2013/0802_xerox-workcentres...
The real barrier to entry is cost. Building out a permanent magnet system like this is likely an order or two cheaper though.
That being said it's very hard to make any even slightly novel machine (MRI or otherwise) that isn't so close to an existing patent that a judge would dismiss a suit out of hand.
Right now, we (physicians) discourage people from getting tested outside of guidelines because we don't know what to do with incidental findings. But you could imagine that as a society, we would like to detect and understand these things, rather than just remain ignorant to them.
Inexpensive technology like this could be perfect for performing large-scale studies with repeated sampling of volunteers over time, to gain information that can help the next generation.
...which is exactly why the comment you're replying to says that physicians discourage them. That's missing the point; noisier devices are indeed not going to be great at improving the existing applications. But there's a whole world of other possibilities out there as long you don't try to substitute questionable data for good. Like monitoring over time, or between-patients studies where you get additional significance from large numbers, or even just fishing expeditions where you see what the cheaper and more deployable stuff is capable of. Not everything needs the best and only the best.
> Rather see a focus on new magnet technologies to reduce cost without loss if already marginal clinical MRI resolution.
Why not both? The work required is going to be pretty different.
And chaining them together is a time-honored technique: use the quick cheap thing to detect reasons to dig in with the fancy stuff. Your base rate may be low, but if the quick check is negative then the adjusted probability might drop it below some other cause that you'd be better off looking into.
Data is good, just don't fuck it up.
They already do this with DEXA body scans.
Examples: https://bodyandbone.com/mobile-dexa-services https://body-comp.com/testing/mobile-dexa/ https://www.bodyspec.com/
My current gut feeling is like: If 0.055 Tesla can create this kind of image quality, what could we possibly expect at 1.5 or more?
Edit: Compressed sense/sensing is not some AI/Machine learning thing. It's pretty neat though, a PR video here. https://www.siemens-healthineers.com/magnetic-resonance-imag...
I skimmed through the article, but still didn't comprehend how much something like this might actually cost.
Siting a conventional MRI is pretty expensive (often requires a new build 6 figures for sure) and operation costs can run up to even 5 fig/month for powerful ones.
They could probably get one of these out the door for approx 100k. Clinical scanners are typically 10x+ that.
Siting cost would be next to nothing, and operating costs low too.
Additionally, the scanner cost is only part the price. There is the Faraday cage, chilling, room setup, building strengthening, scanner install and shipping cost, peripheral equipment (compatible monitoring, injectors, compatible beds and chairs etc). It probably comes in at a doubling of the cost of the actual scanner.
While reducing the cost of the install and running will help a lot, the staffing is the larger cost in radiology, as techs and radiologists are expensive.
Costs will vary hugely depending on where you are in the world, but it isn't cheap anywhere.
Staffing is an interesting one (which I ignored, but good point you can't really) - lots of potential deployments of a small machine like this probably don't look anything like a US standard imaging suite, and aren't going to be staffed the same way. If you run all the numbers in detail you get big variations here, depending on set up.
Scanner location is absolutely a driver in where people get their imaging done. People want good parking and a nearby location.
If the building is being purpose built it is a lot easier, as retro fitting steel isn’t cheap.
I guess this implies not?
That's not particularly different from normal MRIs, and the achieved resolution is not that much worse than normal MRIs. The scans have lower contrast (and repeated/longer scans is one way to improve that) and using for functional imaging will make that worse, but honestly it doesn't seem to have suffered very much at all.
EDIT: I just want to point out that the original subject title of the post on HN was "A low-cost and shielding-free ultra-low-field brain MRI scanner Using AI" ... and the Using AI part of the post title was subsequently removed.
"There are still no OCR tools that work at human level in most applications"
and also from my personal experience working with this technology every day. There are many more mistakes in OCR even with printed material than might be expected.
There is a major problem with Xerox Scanners and the 8's and 0's issue I reference.