I also tweeted a reply, repeating it here - I'm a physicist and a former physicist colleague of mine used to work on ultrasound imaging, here's what he said:
> Ultrasound can go through bone. It just reflects a bit at the boundaries, so regular dumb ultrasound wands show behind it as a shadow. But if you can detect small signal, and do processing, yes, the information is there. And it is not crazy small signals, like 10 to 100 times lower maybe. Not a million times lower.
And on others claiming otherwise:
> Yeah, I think this is a case of "I'm an ultrasound person, and I know that ultrasound is blocked by bone". Kind of like, "I'm a NASA engineer, and I know it it absolutely not worth reusing rockets"
Also, my partner and I are expecting a child, and on Monday got an ultrasound done, in which I can clearly see the inside of my daughter's skull (including a measurement of the size of the corpus callosum). They're thin bones in a fetus I'm sure but ultrasound can obviously go through bone to some extent.
> Also, my partner and I are expecting a child, and on Monday got an ultrasound done, in which I can clearly see the inside of my daughter's skull (including a measurement of the size of the corpus callosum). They're thin bones in a fetus I'm sure but ultrasound can obviously go through bone to some extent.
Going just by Wikipedia ( https://en.wikipedia.org/wiki/Medical_ultrasound#Neonatology ), it seems that imaging inside skulls of fetuses or babies is done by imaging through the gaps in the skull (fontanelles) -- remember, infants have multiple distinct cranial bones with gaps between them; they only fuse into a single skull later. Once the last of the gaps closes up at about 1-1.5 years of age, I think then the only way to do an inside-skull ultrasound is the calculation-heavy approach that Scott discusses and that apparently has yet to be demonstrated.
I don't think that's the case here, since what you get isn't like, an image of just a cone-shaped region inside the head due to a focused beam of ultrasound going through a specific gap - it's an image of the whole baby, and you can see the entire outline of the skull and its contents.
Also, I can just see the inside of her head in every ultrasound image that happens to include a cross section of the head, regardless of whether it was an image trying to deliberately look at the brain or not. Nothing deliberate seems to be needed. I am pretty sure it is just a case of ultrasound straightforwardly going through bone, it's just the bone is not very thick in this case so it's not that big a deal.
As mentioned in the tweet, I'm pretty sure the gap at the top of her skull is a shadow, not an actual gap (even though I know such gaps really do exist). The thickest region of bone that ultrasound has to go through is where the skull surface is tangential to the ultrasound propagation direction, so just below that is where shadowing would be the strongest. The gap looks like it has a sharp edge instead of the exponential-ish decay you would expect from shadowing, which is weird, but across different images it moves based on angle to the ultrasound direction, not keeping a fixed spot on the skull so I am pretty sure it's a shadow. Anyway that's not where the ultrasound would be entering. The plane of the image doesn't even have the eye sockets in it so that's not how the ultrasound is getting in either.
So bone is losing you 10-20dB? (are we talking about signal power or signal amplitude here?) Some googling suggests a traditional scanner has sidelobe peaks at around -13dB. So an object behind the bone could plausibly get swamped by artefacts from bone elsewhere.
Unfortunately I don't know nearly enough about beamforming to say how this would translate to a pool surrounded by transducers.
Don't know if amplitude or power, but sounds like different thickness of bones are going to change transmission by orders of magnitude, so it's probably not meaningful to ask which without specifying what kind of bone!
Yes, also as a physicist, there are necessarily frequencies of ultrasound that can penetrate bone.
This reminds me of taking a radiation safety training led by an engineer; one of his quiz questions was "what is the maximum energy of a gamma ray?" His desired answer was 2 MeV (IIRC), because that's the highest energy you would see if working with terrestrial isotopes. The correct answer, as someone who works with particle accelerators and astrophysical surveys, is of course infinity.
Actually no! The reasons for this are fairly complicated, but one relatively (pun somewhat intended) simple way to think about it is that photons are massless. There is an equivalence between mass and energy, but mass is an invariant. In the photon's rest frame, it has no mass, energy, or momentum, regardless of what energy it has when moving relative to us. (This also means that photons do not experience time.)
The gamma photon is never traveling through perfectly empty nothingness, the space it's traveling through must have some positive temperature. In the case of space, about 3 Kelvin. So the gamma ray can interact with a CMB microwave to become a black hole.
Ok, as a physicist too, I have to say this. The contribution to the curvature tensor comes from the mass-energy tensor, not from the invariant mass of any particles involved. You should totally expect that a sufficiently energetic photon would cause some quantum gravity effects, which why to begin with there was that whole paranoia about "CERN might be creating black holes and will destroy the earth", of which only the second part was dumb.
(For anyone reading, the trivial proof that any such black holes are short lived, is that the atmosphere is regularly bombarded by particles that are 100000x more energetic than anything at CERN, it's just you cannot set up the kind of detectors you have at the LHC there.)
(If you're thinking "wait but that means the notion of whether something causes a black hole is observer dependent", yes, it is. That's the intuition behind the maths of Rindler metric, showing that event horizons can appear by a change of frame.)
Finally, even if you decide that quantum gravity cannot be a thing, quantum field theory itself tells you that something must break at the Landau pole energy, which one way or another would prevent photons as we understand them from existing at that energy. That's even more intrinsic than the QG argument, as it's a problem in the theory describing the photon and its interactions by themselves, nothing else required.
The fetal skull is cartilage, not bone, and it has gaps in it.
I don't know much about the physics, but I would be surprised if every doctor everywhere had missed the fact that actually ultrasound through bone worked just fine. So if it's theoretically possible, there must be some practical blocker. I said in the post that it was probably compute, and I don't see any reason to change that assessment.
Ultrasound can penetrate the skull pretty easily, the bigger problem is signal aberrations produced by the surface wave and the skull having a different sound speed then just water or the brain.
I am doing research specifically on transcranial ultrasound based loosely on these techniques at a national lab, although I regrettably have to be sparse with the details. https://pubmed.ncbi.nlm.nih.gov/40993881/
You are very right in your assessment that computation is a limitation, after we run scans and get data running the back modeling algorithms takes a lot of time even with with only a couple hundred receive elements.
The first thing I noted when I heard of this, is just how much computing power this is going to take.
This is a very interesting technology that is still a fairly young field, so I'm curious to see if midjourneys idea goes anywhere
The last ~10 years or so there has been a lot of development in water surface penetrating laser scanners (green laser), used for doing bathymetric surveys f.ex. I once watched a presentation from a guy who's phd helped pioneer this technology. As I understood it the breakthrough here was do to lots of complicated maths interpreting the laser signal subject to various refraction etc. breaking the water surface. I don't know if it is similar at all, but thought it might be worth mentioning.
Interesting, you're the doctor, but Claude disagrees (about it being cartilage, not about the gaps). I asked:
> Is the skull of a 20w gestational age fetus more accurately described as cartilage than bone?
And Claude said:
> No — that's a common intuition but it's not accurate, at least for the part of the skull people usually have in mind.
>
> The cranial vault (the flat bones: frontal, parietals, squamous occipital and temporal) forms by intramembranous ossification, meaning it ossifies directly from mesenchyme and never passes through a cartilage stage at all. Ossification centres appear around weeks 7–9, so by 20 weeks these are genuinely bone — just thin and incompletely mineralised, with large gaps between the plates. Crucially, the soft spaces between them (sutures and fontanelles) are fibrous connective tissue, not cartilage. So the "softness" of a fetal/neonatal skull, and the ability of the plates to overlap during birth (molding), comes from thin membranous bone plus fibrous gaps, not from a cartilaginous skull.
I mean I definitely don't think doctors are wrong that it can't go through bone in the sense of "gets reflected heavily with poor transmission", I'm not saying they missed some blatantly obvious fact.
The fetal ultrasound is just demonstrating that when the bone is "thin and incompletely mineralised" you do get enough transmission for regular ultrasound wands to work. This to me just shows that's it's obviously a quantitative problem and not absolute - bone five times as thick will have to-the-power-of-five as much transmission which would be extremely little, and the reflections complicate things for sure (though I am sceptical that the processing will be as intractable for AI labs as for academia given the large gap in brainpower between them - we will see).
So it's just that the statement "it can't go through bone" is too absolute - as you said, the truth of the matter is more quantitative - how much signal to noise do you have, and can you process it. Clearly adult skulls are too thick for current approaches.
I don't think the claim is doctors missed it in the sense that 'ohh look I can see' they missed it in the sense that there is a lot of shit you can even see just by shining visible light through your body if you do enough smart things to pull a weak signal from the noise and integrate over long time periods.
One obstacle is that at a certain energy, the ultrasound beams do actually start to heat up your tissue (or in the case of lithotripsy, break up kidney stones). Diagnostic ultrasounds are currently rated *well* below the amount of power output that would produce a physiologic effect. But if we find that we can ultrasound through bone but that it requires more output, this could present a barrier.
A very very common procedure in cardiology is TEE (transesophageal echo). You sedate the patient and place a cardiac ultrasound probe in the esophagus. This is most often used for posterior structures that TTE can't see well (ie the left atrial appendage) or when there's limited windows due to body habitus or ribs. The fact we go to such lengths on a *routine* basis also suggests there's not currently a simpler solution.
Sure you could use some statistical techniques to pull signal from behind the bone, but what do you realistically expect is the maximum amount of signal you could pull? Any large lesions or tumors are going to need an MRI anyway, anything smaller cannot be reliably reported to the patient as anything other than noise.
Ultrasound is already the cheapest it can get, so the only benefit to this technology is some unreliable behind the bone scanning, which will almost always need further scanning anyway.
Just make MRI machines cheaper. China is already working on ultra low cost MRI machines and scanning.
Very good write up. medical doctors are probably in the top 10% of risk adversity, I would guess. So its natural for them to be skeptical of new tech even if it is potentially useful especially once trade offs are considered.
Where do you get "top 10% of risk [aversion]," or that doctors are generally skeptical of new tech? Doctors picked up on technologies like mRNA vaccines, CGMs and insulin pumps, and wearables pretty quickly.
Overall I think this is a solid analysis, and I think the Scanner deserves between the level of hype Twitter is giving it and the amount you are giving it.
I do think it’s possible that this ends up being on net saving a large number of lives, and so relative to expected alternatives I’m very happy that an AI image company has gone into not-yet-AI imaging!
One thing that’s a material difference understated in your analysis between yearly and weekly scans. If it’s important, you can get a yearly or couple times yearly MRI.
Midjourney claims that their per-scan costs will be negligible. This implies that your spa membership will likely include ~unlimited scans.
I think that the data you can get from doing a weekly time series of your body could be substantially better for diagnosis than getting a once a year snapshot, especially for reducing false positives!
Not all false positives will be eliminated by this, but if you see something transient and worrying on an MRI it’s a pain and a real cost to somebody to get another one.
If you see something concerning on a Midjourney Scanner, you can monitor it as frequently as you would like! This makes it easier to refrain from doing potentially harmful interventions, versus having to wait 6 months for another scan, stressing about if it’s continuing to grow.
I think that alone can make it worthwhile; and if they *can* better map air and bone with their setup as they claim, then it seems clear the hype isn’t overblown. It’s not clear they can, but I believe they believe they can, and so I think it’s great that they are spending their private dollars on this bet!
"I think that the data you can get from doing a weekly time series of your body could be substantially better for diagnosis than getting a once a year snapshot, especially for reducing false positives!"
I think this is misunderstanding the threat model. "False positives" are real consisting things in the body, they just don't matter, and they can't always be distinguished from things that matter by their growth rate.
With weekly scans it wouldn't just be growth rate, but permanence. A "zit" inside the body that grows quickly but disappears 3 weeks later will be noticed and filtered out by weekly scans. Annual scans either see the zit or don't see it, but that doesn't help very much, does it?
Unless you're lumping in some fast-growing benign tumours into the "false positive" category. If you've got a benign tumour that takes 2 years to grow to the size of a watermelon, wouldn't it also be better to catch it earlier, even if it isn't cancerous?
Would this be useful for those of us who can't have MRI scans? I have a pacemaker embedded under my skin and this means I'm not allowed anywhere near an MRI scanner.
>Using this intuition pump, I don’t think Paul’s argument stands. Most pimples come after birth! Even if you got a skin scan every month and could compare the current skin scan to the previous one, you’d still be at high risk of fretting over pimples and demanding biopsies for them.
What?? This is the opposite conclusion you should be drawing from this intuition pump. We DO get regular skin scans, multiple times per day in fact, because we CAN see our skin with our eyes and therefore notice the irregularities and changes. And yet people aren't constantly running off to the doctor to get their pimples and moles biopsied. Seeing it with high frequency allows a better understanding of what is and is not concerning, which desensitizes people to weird but benign stuff happening. There are false positives and there are false negatives, but it's not some constant crisis for normal people seeing stuff change all the time, because you just get used to it.
I think the reason we don't freak out about skin irregularities isn't that we see our own skin regularly, it's that we see other people's, and have a robust cultural awareness of skin irregularities.
Yeah, maybe an awkward analogy because it combines two things:
- You can see that pimples exist
- You can look at the pimples closely with the naked eye (and eg see that they don't look like melanoma).
I'm imagining that a scanner could see that a pimple existed, but couldn't look at it and see what it was without a biopsy. I think this is closer to the real situation.
What exactly would stop us from doing all the preventative screening in the world, and then have doctors simply not inform the patient of the things that seem harmless. the patient is no worse off than they were before, and it seems to eliminate the EV negative side.
If the fear is that the threat of lawsuits will make doctors always err on the side of disclosing too much, then this seems like much less of a problem for less litigious cultures than the US, which is essentially every culture on Earth
What do you mean by "not inform"? We're talking about scans here. Presumably people will have access to the actual images produced. Certainly there'd be rioting in the streets if you tried to stop people from viewing their own scans. And many of the actually-not-cancerous growth are visible to the naked eye for a layperson.
As I said, I don't fully understand the tradeoffs - but at least one answer to your question is that I think patients have a right to their medical data by law.
I've heard about this problem with lawyers before (but haven't experienced it myself). Someone got cancer in a specific place? Look through the back-catalog for *anything* weird at that place, the slightest shadow, and sue the doctors who didn't pursue it.
There are things the doctor concludes are harmless. There are things the doctor concludes are harmful. And then there's the middle ground. Lots of room there.
> If you could guarantee that the diagnostician would be perfectly rational (because it’s an AI) and perfectly immune to patient pressure (because it’s an AI)
You can't, though. Even the most casual interaction with LLMs will tell you that they're neither perfectly rational nor immune to pressure. In fact, ChatGPT's susceptibility to bullying is famous. Some people see flattery as the primary use case for these machines, and what is flattery but submission to social pressure?
I don't think the auto-diagnostician would look like a chatbot. It might be LLM-based, but the prompts would be highly standardized. You would send it the image, it would apply the standard factual prompt, and you would never have a chance to flatter it or anything.
That doesn't guarantee that what's inside the mostly-black box is perfectly rational. It does shut down most sources of social pressure, but perfect rationality is simply not what this branch of technology offers.
You can, but due to all of its hallucinations, on average it is far less than "perfectly rational". But to be fair, the ultrasonic scanner would probably employ some kind of classifier rather than an LLM, so it might be less prone to hallucinations. No matter what though, neural networks are stochastic, and are thus not "perfectly rational". To be fair, the same is true of humans.
Randomness is also introduced during training, as it is performed via stochastic gradient descent. But you are right, by and large the irrationality of NNs is not merely due to randomness; I was merely pointing out that *perfect* rationality is unachievable due to randomness.
FWIW, I see hallucinations pretty regularly. I do a fair number of highly specific Google searches, and when you put in words with no obvious connection to one another the search-support AI can get a bit desperate with its interpretations.
Which is one reason I try to ignore the thing, but my eyes read automatically.
Being good at math - or good at applying logical rules in general - is not the same as being perfectly logical. As legions of irrational mathematicians and philosophers can demonstrate.
I'm very good at math and I know a fair bit about probability but I don't necessarily make good decisions.
I'm not trying to say that these machines are useless. But we have to stop assuming that our various irrationalities are somehow inextricably tied to our meat. Artificial emotion is as real as artificial intelligence. Artificial stupidity, artificial insanity, artificial bigotry...we see it all, because silicon-based minds can do anything carbon-based minds can do.
I think the point in this case is you can tune the classifier until you get your preferred ratio of false positives to false negatives, then freeze it and not worry about whether the patient is paranoid or the doctor is scared of lawsuits.
I think this is an interesting development from Midjourney. But the discussion misses the point a little bit: the premise of AI is that collecting massive amounts of data allows you to sift out emergent patterns. It's nothing like the scientific method, there are no hypotheses and no false positives in the collection and the analysis. I can imagine a world where you take a billion scans, run a vast amount of future-learning algorithms on the data and then infer lots of currently non-obvious results. Imaging brains and lungs from complex refraction and reflection patterns is one possibility. Classifying polyps into dangerous or not while they're still tiny seems perfectly doable, or recommending physio to people before they fall and break a hip, or detecting plaque buildup before a stroke or heart attack and treating it then. Why should we wait until there are symptoms to treat, rather than acting pre-emptively? The only reason currently is the lack of data - maybe this supplies it. More likely, the result will be something no-one even thought of yet, because we're still at Henry Ford's 'faster horses' stage of developing the technology. And most likely, it'll be of no use at all.
On a different dimension of questions about this technology: How different will their "using AI to reconstruct images from ultrasound data" be from the currently-existing using AI to do the equivalent of a CSI TV-show "enhance" to significantly upscale images?
The AI will presumably be trained on a bunch of samples to produce images that look like the samples and are reasonably consistent with the incoming data. However, one of the problems with using current AI imaging models to do this sort of thing (in, for example, the upscaling case) is that it does not distinguish what is solidly supported by the data from what is a merely plausible guess that is not directly contradicted by the data. As a result, it's very easy to use the current technology to produce things that are almost entirely guesswork with a very loose connection to reality.
That may be fine for artwork in the hands of someone who is expecting a mostly-fictional guess, but that is not at all what one wants in a medical scanner.
Do the people who are going to be creating this device have the necessary background and understanding to distinguish what's data-supported and what's guesswork in the created images? And are they going to design the UI and output of the machine in such a way that medical technicians can also distinguish the two?
If Midjourney is serious about building this technology instead of running a full scam, then the answer to your "How different" question is "very".
The difference is that "what is solidly supported by the data" vs "what is a merely plausible guess that is not directly contradicted by the data" is a much steeper gradient when you have a large enough amount of sensor data. You verify if the image/physical model matches the sensor data by solving some differential equation (the forward-modeling). And in the other direction, traditionally you either are lucky enough to process the data with some numerically cheap way, in which case you won't need much of AI's help, or have to guess an initial model, solve the equation, compare the answer to the data, and do this many times intelligently to minimize the misfit (the inversion). Most of these problems are ill-posed, so even traditional methods only give you a "plausible guess" in some sense, but that's typically good enough, and there are ways to determine how uncertain the inversion result is.
AI methods can step in at different points. The training data would be data you generate from physics-based forward modeling, and you can train the AI model either in the forward direction so that it can be a fast surrogate forward model (typically solving the differential equation is the slow part of the process, so if AI can do this, it'll already be a huge speedup), or in the other direction so that it can produce an inversion directly.
The structure of the problem is so different that I'm struggling to analogize the space to generating artwork from prompts, but how it should work is not a mystery. To add a bit of context, I work in an adjacent field where some of my colleagues are doing precisely this kind of work, training AI models to replace traditional physics-based inversion. And to be sure, most of these people are physicists and applied mathematicians, so I kind of share your concern (as well as the concern of a few other commentors) from a different angle, that is, if a lot of the expertise required is physics, what is Midjourney specifically bringing to the table? I don't have an answer, but the optimistic stab-in-the-dark guess is that they got especially clever with designing neural nets that process images somewhere along the way.
For the intuition pump about skin conditions that you had to scan for: I think if you run a filter where you only bother about something if it persists and maybe grows over the course of several months, I think you cut the false positive rate down far enough that it might be worth investigating any remaining hits so long as you could do so not-too-invasively and not-too-expensively. Don't know if this analogy would hold for internal organs.
> One possible answer is convenience. Midjourney claims these ultrasounds could be much faster and more comfortable than an MRI (which involves ~60 minutes in a giant scary metal tube that sometimes kills you if you forget to take off your jewelery).
Is it possible to do what they're trying to do with a CT scanner? I feel like you're underselling just how inconvenient an MRI is. It makes loud noises (that are loud even *with* the hearing protection they give you) and you gotta remain still for a long time. CT scans are much easier to undergo.
> This is an interesting take. If you could guarantee that the diagnostician would be perfectly rational (because it’s an AI) and perfectly immune to patient pressure (because it’s an AI), would that switch the sign of primary imaging-based screening from negative to positive?
Perhaps you can make a perfectly rational diagnostician, which is immune to patient pressure, but you can't make a *system* of diagnosticians that is immune to patient pressure because the patient can go elsewhere.
Huh. I have no idea about the medicinal background, but I always thought the advice against population-wide scanning was simple Bayes, i.e. if the incidence of a condition is low enough, even the test with almost perfect specifity (on an order of magnitude that makes any sense) would have a large false discovery rate, just mathematically. Am I off-base with that?
That’s the usual textbook logic yeah, but there are other knobs you can play with to make the usual logic maybe not apply. Maybe if your test gives a number between 0 and 100, instead of freaking out wherever the number is above 50, you wait until it’s above 90, or even 95, which reduces the false discovery rate to maybe something manageable. Or maybe there’s a cheap follow up test that’s easy enough that you’re okay with the screening test having a crazy false discovery rate.
I am satisfied to read this post right after finishing Zvi's weekly AI roundup, which dropped earlier today, and similarly highlighted the MidScanner as an interesting development. The future moves fast, and getting an Actual Doctor's Opinion was exactly the follow-up I was hoping to go look for, so now I don't need to.
Got introduced to DEXA scans years ago by someone with pretentions of high-powered athleticism (who predictably ignored all the fine-grained recommendations, that's a true stereotype), and although I don't have any "real reason" to use them, it's still an interesting experience? Filling out more data on one's character sheet is intrinsically satisfying, and routine medical care doesn't give most of that information. There's certainly some kind of, eh, moral medical hazard to beware of...but that seems like more of an issue with scrupulosity and poor intuition than a fault of the information-giving source? Personally I just found it really cool to get hard proof that, yes, regularly exercising was paying good muscle mass dividends, even if I'll never have a washboard stomach or similar highly-visible signifiers. Which is the reverse intuition from pimple biopsies: it's Good, Acktually to know that good things are happening behind the scenes, when undeniable proof like six-pack abs is already so far along you no longer need the confirmation. Flywheel effects are quite motivating!
The screening MRI cost/risk conclusion here in Taiwan is the opposite of the AMA consensus. Not by a little, it’s wholesale, and not literally but practically mandated by labor law.
Basically everyone in this hood with a full time professional job at a company with >10 people gets a non-MRI full body set of exams every two years. And doing this is a legal obligation that falls on individual citizens here to get done and to provide the results to the HR department if you join a company.
There are thousands of companies here that have enough profit or funding to spend the extra $1000/year, so that people can get full body MRI with initial follow up CT scans if anything from the MRI warrants a higher res look. And basically every company of any type has a majority of the deal with the health exam companies so that if an employee wants to go for the more extensive testing, They can pay an out-of-pocket fee that’s about half of what the retail rate is for the incremental, more sophisticated tests.
Is there any evidence that these scans deliver health benefits, or that these benefits exceed the costs, including the financial and stress costs of "false positives" that Scott's post highlighted? I'd like to think the answers are yes. In the UK we delude ourselves that we have the best healthcare system in the world, so we obviously don't have anything like this!
I asked all the llms and the median conclusion was basically “if you a clear minded and wealthy private individual of middle or late age, it is defensible but by no means obvious or necessary”
Can’t be done for the whole population for financial means, so it isn’t surprising that we have a good explanation to give doctors for why it isn’t beneficial for genpop
I don't think this is a scam (though I do think Midjourney are hyping it up too much, but that's normal).
But I am surprised at the lack of scepticism from the AI community, including from quite sensible people, about something which seems quite improbable on the face of it (an image model company pivoting into medtech and immediately inventing a revolutionary new technology). This makes me think we are probably going to see actual scams emerge out of the AI space soon (and in fact I wouldn't be surprised if at least one neolab is dodgy)
This reminds me of Demis Hassabis' story about how he was explaining AI to Mark Zuckerberg, and Zuckerberg seemed appropriately impressed and excited, and then they started talking about random silly technologies that wouldn't work (eg the metaverse?) and Zuck seemed equally impressed and excited about all of them.
I'm increasingly worried that many people who are excited about AI aren't actually smart enough to realize it's a good technology, they just have a "technology always works" prior (admittedly a refreshing change from the usual "technology is always bad" prior) and this will become increasingly concerning as their successful AI ventures elevate them higher and higher.
The LJ-99 room temperature superconductor thing a couple of years ago seems a good example of this. Lot of people got very excited by something that the scientific priors were heavily against
I think we're just beginning to see the scams arriving, as ever criminals have been in advance of the rest of us in adopting new tech in order to better victimise people.
I agree the most likely result is that this will be part of a spa/boutique private health clinic treatment package, not widely adopted in general hospitals (imagine trying to fill and empty baths for 15-30 patients a day) - a new spin on "taking the waters".
At least this AI pivot shouldn't do as much harm as radium:
Medically, radium was usually injected or taken in pills. It was used to treat a wide range of ailments including hair loss, impotence, atherosclerosis, high blood pressure, rheumatism, gout, sciatica, nephritis and anaemia. This led to a craze for radium-based products, and radioactivity in general, during the 1920s and 1930s. It was added to a wide range of commercial products including: wool for babies, water dispensers, chocolate, soda water, male supports, foundation garments, condoms, toothpaste, suppositories, cigarettes, cleaning products, boot polish, fertilisers, luminous paints and cosmetics. All of these commercial applications had a common theme; that the rays given off by radium had a ‘vitalising’ effect on the human body.
Cosmetics
A number of companies put radioactive materials into cosmetics or made their cosmetics radioactive using the process of induction; i.e., making them temporarily radioactive by exposing them to a radioactive source. As the original work on radium was carried out by the French scientists, Marie and Pierre Curie, it is not surprising that this was more common in France than elsewhere. England also had a number of cosmetic companies that used radium but the same cannot be said of the United States."
It's the latest, cutting-edge science, it must be safe and effective!
I wouldn't be so quick to dismiss this. No doubt your clinician-followers are smart and competent - but they might also be a tad cynical and burnt out.
The Midjourney Scanner page specifically mentions dolphin echolocation as an analogy. Dolphins use lower frequency sound waves that can penetrate more deeply than medical ultrasound, and image bone. Yes, longer wavelengths sacrifice higher-resolution, but with with a few tricks borrowed from super-resolution microscopy (which sees objects smaller than the diffraction limit of light) it might be possible to "break" the limit of longer wavelength sonar.
A new high-resolution, full-body, non-ionizing imaging modality could compliment MRI - and with some fancy AI analysis could reduce false-positives and actually prove useful. Who knows?
(I especially like that there's a fully open-source seismic tomography software stack, using GPU compute, originally developed by (and for) geophysicists, now being used for medical imaging: https://www.stride.codes )
Fascinating! Worse in what sense? Just a less well-developed field? Or more difficult/limited in some fundamental physical way? (Stricter constraints on non-destructive frequencies and amplitudes, perhaps..?)
Even if this were an obviously good product, I would wonder whether Midjourney was positioned to pull it off. They will need to bring in a lot of expertise in ultrasound sensing/medical devices/manufacturing and build a functioning team quite quickly. This stuff is all slower and more painful than developing a software product.
If I were them, I would be looking for an existing innovative medical devices company to partner with/acquire/buy a dept from.
Doctors are famously bad at probabilistic reasoning. Any new testing technology that has a different statistical threshold for clinical action than what they are used to is going to hopelessly confuse them. It will take decades for new generations of doctors to act sane when exposed to new kinds of tests with high rate of false positives.
In this case, usually insurance companies can save you by acting like designated party poopers and refusing to cover the treatments and additional tests that the doctors will ask for just in case. But it won't save the people who pay out of pocket, and it will be very poor consolation for the occasional patient who turns out to be a false-false-positive or the CEO who gets murdered in retaliation.
> If you could guarantee that the diagnostician would be perfectly rational (because it’s an AI) and perfectly immune to patient pressure (because it’s an AI)
Is this some kind of hypothetical AI we don't yet have? Because AFAICT LLMs aren't perfectly rational—they may be less biased than many humans, but they do have their own systemic biases—and they are also famously, uh, mune to client pressures.
I don't think client pressures are a problem - I don't imagine this auto-diagnostician as a chatbot you can talk to, just as a pipeline you submit the scan photos to that has a factory prompt and returns its opinion.
I also think you can quantify/standardize/avoid AI biases better. For example, I bet there are a lot of tough probability questions that Claude would get right 100% of the time, and randomly selected doctors much less often.
𝘮𝘶𝘯𝘶𝘴 (n.): "public service", tax in whatever form it's assessed (money, unskilled labor, military service, putting on a festival...), or a donation to the public coffers.
𝘪𝘮𝘮𝘶𝘯𝘪𝘴 (adj.): describing someone who doesn't have to pay taxes.
Would this be useful for babies? I had a child in the nicu, who got quite a few ultrasounds of heart, brain, and abdomen. Also a CT, which required general anesthesia and thus intubation, and still didn't hold still well enough to get the extra clarity over bedside ultrasound that the cardiologist wanted.
It would work for babies, but I'm not sure it would be better than existing tech. I think whether it could replace CTs would depend on the exist CT indication. I also don't know how still the baby would have to be in the water bath.
As far as using AI to normalize interpretation of other types of scans, just a couple months ago I underwent what Mass General described as an AI-assisted MRCP (magnetic resonance cholangiopancreatography--that is, an MRI focusing on the bile ducts) as part of the treatment for my primary sclerosing cholangitis.
As far as accuracy goes, its conclusion that the liver was heavily scarred and the bile ducts constricted was at least supported by the doctor who later attempted an ERCP (endoscopy into the bile ducts to try to clean them out manually), who couldn't even access the left side and who told me there wasn't any point in trying because the scarring was so bad.
I was a little confused when I didn't see it mentioned in any of the comments in the post. Maybe it's just a new enough technology that not many places have it? Proximity to a facility like Mass General was part of the reason I moved to Rhode Island in the first place. I did find this study from 2025, so it's not unheard of as a concept:
I would also like to register my annoyance at the lack of high-quality audio recordings of MRI procedures. There is one channel on YouTube which has a few recordings, and even names the different scanning processes by sound, but none like what I've heard in the MRIs I've had (and between my Ulcerative Colitis and my PSC, I've had a few, though I don't go out of my way to get them).
Seriously, there are parts of even a regular abdominal MRI which wouldn't need any editing beyond cutting the number of pulses to a multiple of four in order to make some truly great EDM. About two thirds of the way through the MRCP, I started grinning like an idiot because the rapid clicking, without various high and low pitches cycling on and off at regular intervals, was so perfectly musical.
Worth noting I always bring a set of foam earplugs to go on under the headphones (I always have a bag of them on me when I go out), but since moving out of Arizona I've never needed to use my own because the facilities have always offered them. This turns MRIs into what I imagine guided meditation must be like: wearing loose, comfortable clothing; lying on a soft, padded table; and listening to breathing instructions for some 20 minutes.
I appreciate applying the principle of charity and arguing about the merits of the proposal as presented by Midjourney. But I just can't ignore that it's Midjourney specifically proposing this. A company, whose main area of expertise is creating very plausible and realistic looking images that do not correspond to an actual state of the physical world, is claiming that they've pivoted to solving a very different set of technical challenges to build a system that (presumably) still outputs a realistic looking image.
This feels like if the guy famous for magic tricks involving humans hidden in ingeniously constructed boxes announced his pivot to building mechanical chess robots. It's not like there's zero overlap but it's a bit too much of a coincidence.
This was the first thing I thought of. If your entire business is making up fake images, becoming a tomography service is a bizarre thing to do.
Also, I don't really see the overlap, in this case _or_ in the hypothetical case of the guy who's famous for stage magic pivoting to chess robots. The problems aren't the same. The approaches aren't the same. There's no useful transfer from doing one thing to doing the other thing.
"even when entrepreneurs set up their own full-body MRI clinics that don’t cost the system anything, the medical system recommends against using them"
the system can hardly say "this is dangerous when ordinary people do it but it's harmless to billionaires paying out of pocket"
besides, demand is demand. building your own MRI still raises demand for everything that goes into building an MRI, hiring a radiologist still raises demand for radiologists, etc.
1) This is all entirely hypothetical. We have no idea what this tech actually does or what the use cases might be.
2) Detecting cancer is relatively easy. Proving that that detection has a positive effect on people's lives is hard. Prostate cancer is common yet most cases cause no harm while others are fatal. Figuring out how to deal with that is still an ongoing issue.
There is vast literature on failed screening attempts. People might want to look at same arguments about the NHS-Galleri trial, a major cancer screening test which failed its most critical trial.
I didn’t know you were a psychiatrist. Psychiatry seems to be a unique branch of medicine that has so far avoided the march of instrumentalism in medicine. There are no probes or blood tests or any physical objective correlates of mental states such as depression.
Are you also seeking an instrument to observe the mind?
With regards to the whole body MRI and things that are “obviously bad” it’s helpful to think in terms of exactly what you are trying to catch
H. Gilbert Welch divides cancers into birds, rabbits, and turtles.
Birds have already flown away, widely metastatic cancer that either respond to immunotherapy anyway or kills you. It doesn’t matter how early you catch them
Turtles were never going to cause an issue anyway. They grow too slowly.
So what you need to catch are the rabbits — cancers that can be surgically removed before they escape and become incurable. What cancers can an MRI even plausibly detect? We have better, more reliable screenings than MRI for breast, skin, colon, cervical, and lung cancer.
MRI is bad / mediocre at catching GI cancers and uterine cancer early
So really what you are specifically searching for is early pancreatic cancer that’s surgically resectable, ovarian cancer, renal cancer, brain cancer, throat/neck cancer, thyroid cancer and bone cancer
Many, many screening strategies have failed for pancreatic, thyroid and ovarian cancer. They’re too aggressive and hard to see. However this is legitimately a place where MRI could be useful. Renal cancer and prostate cancer are usually turtles. Brain cancer is almost always either a turtle or a bird. Bone cancer is extremely rare. Neck and throat cancer is rare aside from smokers
What about non cancerous diseases? Spinal bone findings are famously worthless in asymptomatic patients. Simple blood work is better for fatty liver and CKD. We still don’t know how to treat most things a brain MRI can see. MRI is a bad test to catch blockages in arteries.
So overall the utility of whole body MRI in non smokers is limited to possible early screening for pancreatic, ovarian, thyroid and maybe prostate /renal cancer, all of which have long and failed histories of screening tests in clinical trials. That’s the benefit you’re trying to achieve. And it’s an impossible benefit to get without accepting a LOT of false positives
There is a screening program for pancreatic cancer in Czechia, but for people with history of pancreatitis only. Allegedly this makes the program efficient.
Most head and neck cancer these days are HPV related. Most thyroid cancer is a turtle. Screening programs detect the slow growing and miss the very rare fatal cases.
I think doctors are working on a principle of don’t bother with a test if the result is irrelevant to what you do. If the rate of false positives is high enough that you’re just going to assume any positive result is a false positive, the test is not worth doing.
(The implication would seem to be — only run the test if you have evidence on other grounds that the patient is likely to have the thing you’re testing for)
> Using this intuition pump, I don’t think Paul’s argument stands. Most pimples come after birth! Even if you got a skin scan every month and could compare the current skin scan to the previous one, you’d still be at high risk of fretting over pimples and demanding biopsies for them.
Not getting this. As I see it, the skin comparison support's PG's argument.
The reason we don't freak out about observed anomalies on the skin is because they are routine, familiar, and understood-harmless.
The reason we do freak out about observed anomalies in deep imaging scans is because they are surprising, foreign (to experience), and not understood at all (again - according to patient experience). But the surprise is because we rarely look - not because the anomalies are rare.
If we had frequent high res scans that frequently showed blips coming in and out, and if our siblings and neighbors and parents had the same, then our reactions would settle pretty quickly to treating these more like pimples than potential cancers to be cut out.
Incidentilomas are real and a real problem. But it is also very difficult to access medical care and run the gauntlet in order to get any testing. Something like a connective tissue disorder, or autoimmune, or rare musculoskeletal diseases are impossible in many locations to get any medical help. Everything requires a code, with the doctors working for the insurance companies. If getting the code is difficult for a doctor or requires arguing, or extra work, the aren't going to often do it. Their time and the hypocratic creed gets ground down by insurance bureaucracy and $20 an hour sparsely staffed call centers, designed for the grind. In other words I feel there is much real estate between incidentilomas and having some ability for patients to self help, with ai tools, and hopefully medical oversight.
Isn't the world better off with more investment (of time and dollars) towards innovative medical scanning technology than another - increasingly commoditized - image generation model?
Does weighing in with a bucket of cold water (as an influential voice, particularly in the target group of people who might decide to make those sort of pivots & investments) promote or deter others from following that same path?
(There is also a point around - if you believe AGI is coming in <10 years - and you understand the "bitter lesson" - is a technology that brings 1,000x the resolution & data that exists today likely to be useful or useless?)
Somewhat paradoxically, if "superintelligent" AI is coming in 10 years, then anything you do today -- be it technology, art, politics, etc. -- is useless, unless it directly impacts said AI. Either the AI is going to be malicious, in which human civilization immediately ends and nothing you've ever done will matter; or it will be benevolent, in which case paradise on Earth is immediately ushered in, human civilization as we knew it immediately ends (to be replaced by something else), and nothing you've done up to that instant will matter.
But also, "don't criticize technology if it's in a good/important field" is a great way to have all of your good/important fields full of constant scams, so much so that nobody ever tries to do good work in them because all of the investment money flows to scammers.
I'm not calling this tech a scam! I think there are some scenarios where it might find niche uses. But I'm pretty concerned about people in my orbit talking about how now all other medical scans are obsolete, or how this proves that the medical cartel was keeping the miracle technologies from us before.
But also, people are constantly freaking out about the possibility of there being "hype" in AI. I think in order to keep actually important things credible, it's important to push back against some of the actually-hyped things so that people have some trust that if Silicon Valley is saying something will be very important, it might actually be very important.
Alright, I like this for Midjourney, I think it's a cool experiment, I doubt this will come to a hospital near you any time soon.
Let's start with the medical response. I don't want to question doctor's medical expertise. They know their field, they're highly trained, and most importantly they're the ones directly talking with the patient and bear ultimate responsibility. But...these people cannot manage tech systems. If there's doctors in the audience, sorry, but you would still be using paper documents if we let you, Congress had to pass the HITECH act in 2009 to mandate you to use electronic records systems on pain of losing Medicare/Medicaid reimbursement and the first thing you did was walk straight into the arms of the Epic EMR monopoly which...charges a lot for not the greatest quality software and stifles innovation because they can just copy any new innovation and integrate it into your workflows which they monopolize. Like, I'm sorry, I can just walk into any hospital or health system in the country, see the software you use, and know that this is not your area of competence.
The only company I've ever seen crack this is Abridge with the AI scribes. Like, I know doctors are down on this, doctors are down on every new technology that could impact their patients and they have a long track record of failure, both in adoption and what they do adopt when legally forced to.
As for Midjourney, I think the tech is definitely there and I think the goofy spa thing is probably the right way to go. AI imaging is definitely not my area of expertise but I think the following dataset probably doesn't exist anywhere and is extremely valuable:
#1 Monthly or quarterly full body scans for a couple thousand people, preferably more than ten thousand.
#2 A full list of all their medical outcomes.
Like, I'm confident we have datasets where we suspect there's cancer, we send the patient to get an MRI, and then we determine whether there's cancer. That's a proven technology. But there's very specific and that's not really what we want. We want a full-body scan of everything, on a semi-regular basis, and then all outcomes over the subsequent years. We don't care whether it's cancer or arthritis or some muscularskeletal degeneracy thing or a fractured bone or whatever. The dream is you get a full body scan and then it diagnoses everything. And the best way to get that is....well, scan a bunch of people semi-regularly and then track their outcomes over time.
And how would I sell that, how would I build this dataset? Well, I can't go to hospitals, they can't figure out binary sepsis diagnosis, they'll never do this. But, like, what if I put a goofy golden spa where:
"Our spa will have hot tubs, saunas, cold plunges, and cozy rooms with pools of golden light which softly scan your body.
It should be a place you love going, whether it's by yourself, or with friends. It should be available 24/7. The scans are a side-effect. You barely think of them when going to the spa. But suddenly, you have a huge library of data about your health."
Like, what if we just sell a ritzy spa experience to rich technophiles in San Francisco? There's got to be tens of thousands of those. Midjourney can probably do this at a loss for awhile, undercost it and burn some VC cash, because if they've got weekly whole body scans for a substantial population and then they can get access to those patient's medical records and outcomes (which this patient population would probably sign off on)...woof. That's an insanely valuable dataset. That's a dream.
And there's a big, big feedback loop. Start by selling a spa experience to rich techies, cool, maybe all it is for now is a better Dexa scan. But once your population starts having health issues, once you've gotten weekly scans of, say, 70 people who got breast cancer and you start really being able to flag it in the early days, you can publicize some wins and start to get users outside the rich techie crowd...woof.
Because I think the fundamental is right. I don't think anyone seriously doubts the following: if you had a weekly/monthly dataset of full-body scans for hundreds of thousands of people and then were able to track their health outcomes, you'd be able to predict a wild number of bad health events.
TLDR; I don't think MidJourney is selling a mature product right now, I think they have a MVP that could really scale if they get users/get data. I'm not surprised that doctors and medical professionals are down on this; they have a really bad track record on technology adoption.
Thank you for basically writing what I was thinking.
I would add to this that I think people tend to have a strong bias towards first principles reasoning in cases like these, which gets in the way of being appropriately empirical. An argument that "all X is good for is Y" is usually not a great reason to avoid experimenting with other uses for X, and this is doubly true when you're talking about data generation. Ultrasound might have nothing to offer here, but we need way more, not less, of this type of experimentation. Whether that justifies any of the annoying punditry or hype Scott encounters in his milleu is a different matter.
> All of your organs are constantly growing little pimple-like things that don’t matter. Depending on the location and type, we call these cysts, polyps, fibroids, lipomas, adenomas, hamartomas, etc. We are bad at distinguishing these from dangerous tumors - and so far the studies show that when we try to do it in the entire population on a mass scale, we cause more problems than we solve. That could change in the future! But it’s a separate bet from whether ultrasound tomography will be good or not, and the ultrasound tomography won’t be very useful unless we solve this problem too.
How would this change in the future, if NOT by scanning more regularly?
I (not a doctor) am absolutely willing to buy the argument that omni-scanning is _impractical_. Maybe if we start today, it eventually saves 1 billion QALYs a year in early detection and treatment, but only after the upfront payment of 100 billion QALYs in medical anxiety and expense.
But practicality (and the Midjourney specific stuff) aside, it just seems that... diagnostic data should be commutative?
Say modern medical has a rule, "If you're coughing up blood, and the MRI shows a lump in your lungs, you have lung cancer."
Patient A presents with blood-coughing. They're ordered an MRI, which shows a lump in their lungs. The doctor says, "It's probably lung cancer, here's your costly and uncomfortable but life-saving treatment plan."
Patient B shows up for their appointment. Due to a clerical mistake, before seeing the doctor, they're sent in for an MRI, which shows a lump in their lungs. The doctor says, "Unfortunately, our medical system is bad at handling uncertainty, so we're gonna have to recommend you have the same costly and uncomfortable treatment as Patient A, even though you're probably fine." Patient B says, "Well, I have been coughing up blood lately..."
Surely the doctor updates on learning this information, even if their diagnosis isn't allowed to reflect it.
> if you tell your hospital orderlies “please transport this frail 90-year old to my giant water-filled tank, and lower them in slowly” they will stab you with your own scalpel.
There's another SF company doing ultrasonic phased-array stuff. Their goal is stimulating lucid dreaming though, here: https://www.prophetic.com/. Might find it interesting. I wonder if this means that these ultrasonic phased arrays have just recently gotten cheap enough to use, or that LLMs like to recommend them to founders, or something.
I would love a deep(er) dive from Scott about mainstream medicine's arguments against full-body scans and the counterarguments. I too have been struggling to wrap my mind fully around the issue and have noticed the frequent conflation of purely medical considerations with economic ones from many anti-screening people. At the same time, I understand the medical concerns aren't trivial.
In addition to raising the thresholds for action or developing new technologies, one could imagine AI-driven improvements to the analysis of existing imaging types. And perhaps there are ways to reduce the downsides of false positives (e.g., additional, more focused scans before any biopsies). Anyway, the point is I'd vote for a "Much More Than You Wanted to Know" and further exploration/untangling/speculation about all of this.
Well, it seems AI has now reached the point of all new technologies when it's peddled as a miracle cure/quack nostrum. The same as radium (stunning new discovery can only be good for what ails you!) and electricity:
Here is a question right up the alley of the "rationalist community".
There is a famous named law about the observation that most journalistic output about your own area of expertise will be inaccurate so you should assume that most journalistic output is inaccurate (period) and that you just do not notice this outside your own area of expertise. Crucially the journalistic output sounds plausible to the layman. You have heard this one.
So now suppose you learned that a relevant AI player was trying to move into your area of expertise and you judge the product market fit to be ... well let's say less than perfect. Should this, by analogy to the journalism example, make you, in the parlance of this community, "update" your beliefs about product market fit in other potential applications of AI?
An acquaintance was late to board games this morning because his flight back from SF coming back from the release party was delayed and he needed some extra sleep. He doesn't work for Midjourney but for a supplier. Things that came up in conversation is that system apparently involves racks of Nvidia GPUs and they're not trying to store all the data they collect but process it as they go.
I agree that the Ultra-Ultrasound sounds like it might be a bad idea. But the general fact that a big (or medium sized) AI company is explicitly pivoting to medical scanning & diagnosis is very exciting for me. I have chronic "non specific" neck & lower back pain, which can be pretty debilitating when it's bad and affects many aspects of my life even when it isn't. I've seen so many different specialties and gone through so many different front & second line treatments. I've even been to the Spine Center at a pretty famous medical system, where the first doctor I met opened with, "There's a lot we don't know about the spine."
So besides being a seemingly very good fit for the things that AI does well (synthesizing a lot of information and recognizing patterns), medical diagnosis is also pretty much the only near to mid term AI capability that I personally care about. Claude is fun & saves time in some of my work, but is very much not worth the risk of AI-launched nukes IMO. An AI that can tell me how to not be in pain every day for the rest of my life, though... I'd take a little nuke risk for that!
Quoting Dr. Harris from the article: "To the lay person, it seems obvious that finding a cancer on imaging earlier is always better, but in many cases, it’s not only not better it can be worse! Especially on a population basis when you account for medical misadventures that may ensue from unnecessary or earlier or futile treatment."
I dislike this kind of framing, especially with the "lay person" condescension. From an epistemological standpoint, it is always better (basic information theory). It can only be worse because it is scaffolded within an inferior decision making system. The medical profession seems particularly confident in taking their current consensus best practices as fact, and one gets the feeling everyone within the profession is constantly getting a memo of what they're supposed to say. When talking with a medical professional (not all, but an annoyingly high percentage of the time), you always feel feel like you're talking to a generic representative of the entire profession, not a single human being.
As a trivial aside, I learned at some point that birthmarks are not—as the common wisdom suggests—perfectly benign. Pigmented birthmarks carry approximately the same cancer risk as ordinarily acquired skin moles, which is to say very little risk. However, that risk is definitely relatively much higher than any equivalently sized region of 'unblemished' skin.
All that to say: If you bother to monitor ordinary moles for malignancy, then you should also monitor your birthmarks for malignancy.
I had quite a large one on my right forearm and it went on me a few years ago very quickly. Two different melanomas, two different genetic signatures. Apparently very unusual; one of them was gotten rid of. It was a flat discoloration. The other one formed like a mole and I didn't notice it soon enough
It feels unlikely to me that this particular device is going to be useful in any meaningful way, but the medical community's fear of overmeasurement feels very counterproductive to the future of medical research. More preventative scans of healthy people seems obviously good in the long run; the trick is getting past the harm of acting on incidentalomas in the short run. Not all radiologists are opposed to this sort of thing - I used to work at one of these full body MRI companies and we worked with plenty of enthusiastic ones - but presenting this data needs to be done very intentionally.
It may be a faux pas to link to my own post from years ago, but https://ostro.ws/post-healthcare has some more thoughts about how we can massage the messaging to make more measurement less dangerous.
When a solid body is immersed into water bubbles form on the surface. They are a problem. Even if you degass the water - which you have to do for good acoustical measurements - the skin will bring trapped air and fresh bubbles.
Supposedly the images are from a demo unit. It may be that there are enough gaps between such bubbles to allow for the propagation of the waves between them in a similar way as with infant skulls.
I have no idea how the images were generated by a company whose expertise is in generating images of things that don’t exist.
The problem with bubbles is that they resonate and appear much larger than their physical size. You don’t need them to be placed closely to completely dominate the reflection.
This is why, for example, cavitation-free propellers are a must for stealth submarines. Bubbles light up on sonar like a giant disco ball.
Lol on the image generation part. I’m on the “it can be real but very disappointing side” of the fence here. Certainly possible for them to be faking everything but that seems unlikely. We’ll see
> When a solid body is immersed into water bubbles form on the surface. They are a problem.
There is a big difference between a problem, and a showstopper. There are all sorts of potential solutions to this. From chemicals that disrupt bubble formation, to underwater brushes or Jacuzzi jets to physically dislodge the bubbles to fancy algorithms that remove bubbles from the data.
You don’t work in ultrasound I take it. It always looks like there are “all sorts of potential solutions” in a domain you are not familiar with. For example, brushes make bubbles. Jets make bubbles. Do I need to go on?
They mention a “consumable” that provides an interface. A brief search didn’t find any technical details. So I don’t know.
But look at their device. This is no casual spa full-body scan. Dedicated bath for a small part of the body. Took years (decades, really) to make work, and still only as a supplemental to normal mammography. I stand by my assessment: Midjourney is full of it.
According to their pamphlet, the disposable Sequr Breast Interface only touches the end of the breast, by the nipple and uses gentle suction to "center, shape, and steady [the breast] during the image acquisition process". The majority of the breast is just in water.
Made a comment on a reply, but I had another thing I want to add here.
For one transcranial ultrasound is a thing that's already done and being pursued by different labs and research groups. I'm doing work specifically on transcranial ultrasound tomography, very similar to this at a national lab: https://pubmed.ncbi.nlm.nih.gov/40993881/
The thing I wanted to say that I haven't heard talked about enough, is the biggest advantage of this technology is that you can determine the sound speed of different artifacts, allowing much more quantitative analysis over the current qualitative "It looks like x", which is how it is done now.
This is not the first time a company has tried to bring ultrasound tomography to the market, there was a notable company who actually did bring a machine that would use ultrasound tomography to do the same job mammograms do, and no one bought it. For a couple reasons, mainly that most doctors feel that mammograms are the final form of that technology, but also just because that market is so saturated.
After doing some research and talking to colleagues more versed in the field. I do not think they are currently doing tomography as stated. I think they are doing compound B mode imaging. Tomography can penetrate and image through bone. Compound B mode imaging functionally cannot. I think this is a significant difference that they haven’t talked about yet. They are using butterfly Ultrasound-on-a-chip transducers which do not give you the raw rf data required to do the inversion to do tomography. They have built in beam forming and processing. I think it’s just helical B mode imaging. Which is backed up by the fact that the guy they hired to do this did his PhD in helical B mode imagining.
That difference really hampers the usability of this technology for me. Hopefully they will address this soon.
Additionally, this group published a paper doing this on a rat which they later redacted because it anatomically made no sense. Alluding to them feeding information into an AI which output garbage.
Compared to the SoftVue ultrasound tomography scanner, the Midjourney scanner has several orders of magnitude more transducers/sensors, orders of magnitude more equivalent computational power (presumably using AI to greatly increase the efficiency), and they may be attempting full waveform inversion. If it works, it's potentially a much better image than your standard ultrasound, and could potentially see through bone. With the 360° scanning, they can potentially scan around air pockets even if they can't see inside the air pockets.
I want to make two broad points which tie into each other:
1. The over-diagnosis problem is rooted in trying to do medicine with population level data. What we really need is more individual tracking / continuous monitoring with high quality sensors of physiology.
2. The Midjourney tank is a step in the wrong direction. It's still a huge device, does a full body scan at low sensitivity, and they are aiming for 50k devices worldwide. What would help is almost the opposite -- an inexpensive and highly sensitive scanner that we can manufacture in the millions to billions, and hand to everyone for home use.
----
Anybody who's learned to ride a bicycle knows that over-steering is dangerous, and the surest way to take the system off-track. The correct method when controlling any complex system is always to acquire high-fidelity sensory input, at very low latency, but act on it sparingly.
If the average person can learn how to ride a bike, then I'm confident they can also internalize the same in the context of medicine; I propose the hurdle is that the epistemology of medicine is commonly misunderstood.
Contrary to the "Anna Karenina principle": every healthy human body is different, and well-regulated in its own way. It is not a machine that is supposed to be a perfect replica of some template human.
Here is where both technologists and doctors commonly go wrong: they implicitly adopt the latter perspective, and think that leveraging more data means comparing over the population to average out deviations. The natural consequence of this is to start "fixing" deviations, and this is the territory of dangerous over-steering. Perhaps because statistics attunes us to average over individual differences, any mention of data unfortunately makes people instinctively jump to this "cross-sectional" corner. Tests which are point-in-time and compare one human to others are problematic in this respect. This is especially acute when the body is measured along many many axes: like a full body scan.
The USAF study of pilot anatomy proves this point: no pilot is average in every attribute.
If we start from first principles and ask how to improve the quality of medicine, we end up in a very different corner of design space: the "longitudinal" corner.
Good doctors understand that biological systems operate on the opposite of the Anna Karenina principle. What they are actually doing, when helping a patient, is bringing all their knowledge and experience to bear on understanding the specific individual in front of them, and then steering them with the gentlest nudge necessary. The most useful data for this is actually high-res longitudinal tracking of this individual, so the doctor can help troubleshoot how *this physiology* might be going off homeostasis, and nudge it back. This is also why good doctors always focus on specific symptoms and troubleshoot very narrowly and in a hypothesis-driven manner; the goal is to fix the problem at hand and get back on track, rather than "solving" every deviation from population average.
Correspondingly, it is a huge mistake to think of the role of preventive scanning as comparing an individual's physiology to the human template. Every human is largely functional as they are, and there is likely zero immediate clinical signal in deviations from the population average. Doctors get up in arms about increased scanning because they think this is what they are supposed to look for (after all, textbooks explain how the typical human body behaves, and medical practice is always emphasizing population level-guidelines, and the statistics education we beat everyone with always emphasizes averaging).
So what is the role of preventive scanning and big data? Recall that you can't ride a bicycle by looking at the road immediately in front of you and reacting to every little bump. To steer well you've got to look farther ahead and react only to those perturbations which could compound to take you off-course if left unattended. The true role of preventive scanning is to have early warning of things that might upset homeostasis further down the line: high-resolution and low latency. And the role of big data is to help develop good models for how likely the *combination of observed deviations* is to upset homeostasis in the longer term i.e. to triage which deviations are likely benign and what situations need escalated attention… much like what a good doctor does. Given that doctors cannot attend to everyone everywhere all at once, we need all the diagnostic tracking possible to improve how we make this decision. Today this process is ad hoc and completely implicit, which means a large occurrence of both false positives and false negatives.
Therefore, the insights necessary to help any individual are very unlikely to come from a point-in-time observation; what is needed is an extended duration monitoring of the full physiology. Eg. "Yeah, we observe a few small blobs in that tissue over there, but as such they don't seem to be growing over the last year. But oh you are experiencing weight loss; hmmm. Have you additionally observed X/Y/Z? The possible cause could be related to this other indication on this other tissue, if A were causing B were causing C which then manifests in symptoms X, Y, Z." Notice how that sounds far more like a consultation with a good doctor, rather than the trope of let's find and fix (a la Procrustes) every deviation so you can fit the population average.
This clarifies that the most important axis of resolution for diagnostic information is time, and not population. What would help is highly-sensitive medical imaging that can track each the personal physiology of each human over extended periods.
The US Air Force concluded that designing the cockpit for the average pilot resulted in more accidents and fatalities, and their solution was to design a cockpit that every individual pilot could adapt to their anatomy.
----
Unfortunately the Midjourney tank is not enough of a step forward in the technology to enable the kind of imaging abundance we need. It's still a huge device, does a full body scan at low sensitivity, and they are aiming for 50k devices worldwide. That's not very different from the total number of MRI machines in the world today, and how many MRI scans has the typical person had?
We need an innovation in both the sensitivity and the form factor: an inexpensive scanner that will be more sensitive than the ultrasound technology which powers Midjourney, in a device which we can manufacture in the millions to billions, and hand everyone for ubiquitous use.
PS: If I were to venture into a corollary example of over-pathologization by fitting to the population average, I believe mental health is manifesting a crisis due to exactly this mistake. But that is a tangent to the discussion at hand, and Scott has a better understanding of the ground reality than I do, so that would be an interesting branch for another time
>Correspondingly, it is a huge mistake to think of the role of preventive scanning as comparing an individual's physiology to the human template. Every human is largely functional as they are, and there is likely zero immediate clinical signal in deviations from the population average.
This is absolutely not true. There are many measures of human anatomy and physiology for which deviance from average in one or both directions predicts shortened life. Heart size is one of many. So is walking speed.
Yes, the current system uses deviations from population average as a diagnostic signal, but have you considered that might not be the best thing to do? That's the point I'm making.
Hi: I need help from people who love science! Can data center cooling loops recycle electricity from condensation as MIT's groundbreaking 2013 experiments suggest?
As data centers face an unprecedented energy and cooling crisis driven by AI infrastructure, billions are being spent on managing thermal phase changes—specifically evaporation and condensation.
In my latest work at Decoding Science, I’ve outlined a simple backyard citizen-scientist experiment that suggests we fundamentally misunderstand electron behavior during these phase changes. Standard solid-state models predict a neutral net current during condensation on these surfaces, but geometric manipulation reveals a measurable electrical anomaly.
If this anomaly scales, it means the massive humidity and condensation loops already running inside data center cooling infrastructure could be harnessed to harvest electricity directly from waste heat. Perhaps equally significant is how understanding evaporation will facilitate the cooling process.
I’ve laid out the simple, reproducible experiment to prove this effect. Would you be interested in a brief look at how rethinking electron kinetics during condensation could impact the future of data center efficiency?
Did a pancreatic MRI last month (GLP1 caused enzymes to shoot up, looks harmless). And of course didn't want to wait 3 days for the results, so I took the CD, bought a $10 CD reader and loosed Claude Code on the data.
It ... worked? Installed a bunch of software to analyze it, did a bunch of things to try to compensate for not being a human, and in the end found about as much as the final report - which isn't as exciting as it sounds, because it rounded up to "nothing".
Note that I already have a project with all my clinical history and genetic data. As one does. So the value here is in correlations - you can have Fable level LLMs pour over all your data in the background, for what will soon cost $20 per month. Correlate genetics, lab results, MRIs, occasional Midjourney scans, your daily data (Oura, weight scale, Apple Watch) and latest research.
Currently this takes a certain personality to not obsess over health (or, in my case, to enjoy obsessing over health). But raw Opus is already pretty discerning. Fable with a proper harness can definitely do this in a practical manner.
I'm not a physicist, just a medical student going into neurosurgery with a PhD in neuroimaging, so I don't know much about ultrasound, but we regularly do high-intensity focused ultrasound (HiFu) procedures, so at least some ultrasound waves obviously can penetrate bone in a highly targeted manner. Whether those waves can be used for imaging is an open question, but it's not an obviously wrong approach.
I think the skin example is good but for other reasons. Does the fact that we happen to be able to see what's on your skin make us somehow at a disadvantage versus if we couldn't see it? Of course not. It's not an information hazard being able to see people's moles. We adjust our protocols (maybe implicitly) to account for the noise.
I've always been baffled by the fact that in response to the fact that doctors make harmful interventions in response to some diagnostic tests -- eg prostrate screening at too young an age -- the medical establishment recommends the tests not be done.
Why not first just try recommending you don't act on certain results of the tests? I could imagine a different world where they tried that but patients were so pushy doctors felt they didn't have a choice. But that doesn't seem to be this world. I don't remember any long period where the recommendations pushed not treating lumps discovered during mammograms or the like.
If it's about how the patient reacts then it's kinda paternalistic since some people will be totally unbothered finding out that info. Is this downstream of legal liability?
---
Not that most of medicine makes sense to me. Still seems bonkers to me that it's a horrible ethical violation to give a patient even a vaccine without their consent but you can withdraw a medication they've been taking their entire adult lives and it's perfectly fine. Indeed it might be judged unethical to allow the patient to maintain their current state.
The scanner problem runs parallel to a much deeper issue: how AI doesn't just scan or filter history, but actively reformats it.
When testing how major bots (ChatGPT, Claude, Copilot, Grok) process the exact same historical data—like a closed murder case—you quickly realize we aren't facing simple 'hallucinations.' It’s a systematic rewrite. The dangerous part isn't the AI generating raw code or images; it's the digital software overriding physical hardware (digital history vs. biological consensus). If the scanners become the architects of what we perceive as 'proven truth,' digital censorship will stop looking like a sign of power and more like a total crisis of information control.
Question from a medicine-person for the physics-folks here:
Ultrasound probes currently are both the transmitter and also the receiver. The limitation of this is that it can only image objects that it has line of sight on. This presents one of the main limitations of transthoracic echocardiography (TTE): your beam has to dodge the ribs in order to get to the good stuff (the heart). This is also what makes transesophageal echo (TEE) so useful; it provides amazing pictures (no ribs in the way) at the somewhat inconvenient cost of needing to shove an ultrasound probe down the esophagus.
Could you have a transmitter and receiver that are separate? Aka, can you have a pitcher throw a sound wave that glances off an object and is caught by a receiver standing at a 90 degree angle away?
Could you have *multiple* transmitters and receivers that are all separate? The analogy for this in my head is: if an individual whale in a pod transmits a sound wave, are the other whales able to reconstruct a 3D model from that sound wave? If the whale pod was a hive mind, what level fidelity image could it get?
That's what's interesting to me about this Midjourney project; unlike traditional ultrasound it has multiple angles of transmitters and receivers. Can you take advantage of this, or is the math too difficult?
Bonus question: could you drop a TEE probe inside the body, and have them received by ultrasound receivers outside the body?
Yes this can be and is being done. The transducers are reciprocal so they can do both, and you can send a signal from one and receive with another or multiple.
I have been reading a lot of the vitriolic debates on X. I can't help but feel that both sides (tech optimists and medical doctors) have major valid points but are talking past each other due to differences in preferences and timelines.
The current state of coordination for collecting high quality large data sets with outcome data that is accessible for decision making for physicians and patient is severely lacking but I don't know how to fix this. I think the tech optimist argument is basically we just need to collect obscene amounts of data and get AI to sort it all out as quickly as possible.
>The Midjourney Scanner probably won’t help with [things DEXA is used for], because some bones are behind other bones or air
The bones most relevant for DEXA's biggest use-case (osteoporosis screening) are not hidden in those ways: tib/fib/femur, and also radius/ulna/humerus. Assuming you can shoot the ultrasound waves from any angle you want. Lumbar spine is sort of hidden, depends how good the scanner is. However the real advantage of DEXA is that it gives you bone *density*, which ultrasound cannot (AFAIK?). But the Midjourney scanner looks like it can get 3D bone geometry, which DEXA cannot (since it's a 2D projection). So, maybe an opportunity here?
I agree that the main on-ramp for revenue is probably longevity bros paying for better (or "better") reports on muscle mass, fat content, etc, though.
More interesting approach would be to improve already existing MRI diagnostic with software solution. I heard of an Swedish startup working on a simulation based methods that would also dramatically cut scanning time, and their results look promising.
This post sounds like it makes the assumption that AI-based ultrasound is limited to detecting what human doctor can detect via ultrasound.
We are currently quite poor at detecting a lot of issues around lower back pain, and AI-assisted ultrasound can do a better job at telling which fascia planes move against each others and which don't.
If you treat medical diagnosis as a software problem instead of treating it as being about training human doctor assessment, I think there's a lot of room for getting better and it seems this MidJourney has a good chance to do that.
Hm, there are a couple of claims you are making that are just inaccurate, particularly in your MRI/Ultrasound comparisons.
(Qualifications: I did fMRI in the Gallant Lab in grad school, where we were able to decode viewed images from brain activity. More recently, I have been working on functional brain imaging with ultrasound, using the same chip architecture that Midjourney is using, from Butterfly)
1. Higher resolution -- not necessarily and it depends on what you're comparing. Ultrasound's resolution (for linear beamforming) is limited by the transmission wavelength, which could be as low as 100 microns. MRI is limited by the slew rate of magnetic gradients, i.e. so imaging at higher resolution takes a lot longer. Ultrasound set ups like this could enable imaging the body moving in real time in a way that ultrafast MRI will never be able to match at the same resolution.
2. Ultrasound can go through bone, and it actually penetrates quite well. The problem is things like the terbicular layer on the inside is the skull, where the microstructure is super nonuniform. This makes it so you can't do standard beamforming. But here's a key: this is fundamentally a computational tomography problem, not a physics problem. Approaches like FWI have demonstrated that you can even reconstruct gray matter through the skull. And it's hard to imagine a company better-positioned than Midjourney to tackle this shape of problem. (There are also amortized inference approaches to FWI that could make it a whole lot faster)
I think we're going to see ultrasound become faster and (a lot) cheaper than MRI, and it will be particularly good at imaging things that are moving. Because you can collect it so much more cheaply, we'll be able train diagnostic and segmentation models much more easily.
Heck, I might have talked myself into applying for a job there 🤣
I also tweeted a reply, repeating it here - I'm a physicist and a former physicist colleague of mine used to work on ultrasound imaging, here's what he said:
> Ultrasound can go through bone. It just reflects a bit at the boundaries, so regular dumb ultrasound wands show behind it as a shadow. But if you can detect small signal, and do processing, yes, the information is there. And it is not crazy small signals, like 10 to 100 times lower maybe. Not a million times lower.
And on others claiming otherwise:
> Yeah, I think this is a case of "I'm an ultrasound person, and I know that ultrasound is blocked by bone". Kind of like, "I'm a NASA engineer, and I know it it absolutely not worth reusing rockets"
Also, my partner and I are expecting a child, and on Monday got an ultrasound done, in which I can clearly see the inside of my daughter's skull (including a measurement of the size of the corpus callosum). They're thin bones in a fetus I'm sure but ultrasound can obviously go through bone to some extent.
> Also, my partner and I are expecting a child, and on Monday got an ultrasound done, in which I can clearly see the inside of my daughter's skull (including a measurement of the size of the corpus callosum). They're thin bones in a fetus I'm sure but ultrasound can obviously go through bone to some extent.
Going just by Wikipedia ( https://en.wikipedia.org/wiki/Medical_ultrasound#Neonatology ), it seems that imaging inside skulls of fetuses or babies is done by imaging through the gaps in the skull (fontanelles) -- remember, infants have multiple distinct cranial bones with gaps between them; they only fuse into a single skull later. Once the last of the gaps closes up at about 1-1.5 years of age, I think then the only way to do an inside-skull ultrasound is the calculation-heavy approach that Scott discusses and that apparently has yet to be demonstrated.
I don't think that's the case here, since what you get isn't like, an image of just a cone-shaped region inside the head due to a focused beam of ultrasound going through a specific gap - it's an image of the whole baby, and you can see the entire outline of the skull and its contents.
Also, I can just see the inside of her head in every ultrasound image that happens to include a cross section of the head, regardless of whether it was an image trying to deliberately look at the brain or not. Nothing deliberate seems to be needed. I am pretty sure it is just a case of ultrasound straightforwardly going through bone, it's just the bone is not very thick in this case so it's not that big a deal.
Here's an example image:
https://x.com/doubleunplussed/status/2067745730794787139
As mentioned in the tweet, I'm pretty sure the gap at the top of her skull is a shadow, not an actual gap (even though I know such gaps really do exist). The thickest region of bone that ultrasound has to go through is where the skull surface is tangential to the ultrasound propagation direction, so just below that is where shadowing would be the strongest. The gap looks like it has a sharp edge instead of the exponential-ish decay you would expect from shadowing, which is weird, but across different images it moves based on angle to the ultrasound direction, not keeping a fixed spot on the skull so I am pretty sure it's a shadow. Anyway that's not where the ultrasound would be entering. The plane of the image doesn't even have the eye sockets in it so that's not how the ultrasound is getting in either.
Hm, interesting. Yeah, no idea, not my area, but what you say sounds plausible.
So bone is losing you 10-20dB? (are we talking about signal power or signal amplitude here?) Some googling suggests a traditional scanner has sidelobe peaks at around -13dB. So an object behind the bone could plausibly get swamped by artefacts from bone elsewhere.
Unfortunately I don't know nearly enough about beamforming to say how this would translate to a pool surrounded by transducers.
Don't know if amplitude or power, but sounds like different thickness of bones are going to change transmission by orders of magnitude, so it's probably not meaningful to ask which without specifying what kind of bone!
Congratulations!
Thank you!
Yes, also as a physicist, there are necessarily frequencies of ultrasound that can penetrate bone.
This reminds me of taking a radiation safety training led by an engineer; one of his quiz questions was "what is the maximum energy of a gamma ray?" His desired answer was 2 MeV (IIRC), because that's the highest energy you would see if working with terrestrial isotopes. The correct answer, as someone who works with particle accelerators and astrophysical surveys, is of course infinity.
OT: Isn't there some finite energy threshold beyond which a photon would become a black hole?
– not a physicist
Actually no! The reasons for this are fairly complicated, but one relatively (pun somewhat intended) simple way to think about it is that photons are massless. There is an equivalence between mass and energy, but mass is an invariant. In the photon's rest frame, it has no mass, energy, or momentum, regardless of what energy it has when moving relative to us. (This also means that photons do not experience time.)
Except absolute 0 isn't a possible temperature.
The gamma photon is never traveling through perfectly empty nothingness, the space it's traveling through must have some positive temperature. In the case of space, about 3 Kelvin. So the gamma ray can interact with a CMB microwave to become a black hole.
This had been theorized for quite some time, but was recently disproven (for typical scenarios): https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.133.041401
Ok, as a physicist too, I have to say this. The contribution to the curvature tensor comes from the mass-energy tensor, not from the invariant mass of any particles involved. You should totally expect that a sufficiently energetic photon would cause some quantum gravity effects, which why to begin with there was that whole paranoia about "CERN might be creating black holes and will destroy the earth", of which only the second part was dumb.
(For anyone reading, the trivial proof that any such black holes are short lived, is that the atmosphere is regularly bombarded by particles that are 100000x more energetic than anything at CERN, it's just you cannot set up the kind of detectors you have at the LHC there.)
(If you're thinking "wait but that means the notion of whether something causes a black hole is observer dependent", yes, it is. That's the intuition behind the maths of Rindler metric, showing that event horizons can appear by a change of frame.)
Finally, even if you decide that quantum gravity cannot be a thing, quantum field theory itself tells you that something must break at the Landau pole energy, which one way or another would prevent photons as we understand them from existing at that energy. That's even more intrinsic than the QG argument, as it's a problem in the theory describing the photon and its interactions by themselves, nothing else required.
LHC black hole theories have nothing to do with photons, and certainly not with *single* photons.
The fetal skull is cartilage, not bone, and it has gaps in it.
I don't know much about the physics, but I would be surprised if every doctor everywhere had missed the fact that actually ultrasound through bone worked just fine. So if it's theoretically possible, there must be some practical blocker. I said in the post that it was probably compute, and I don't see any reason to change that assessment.
Here's an interesting story about doctors placing a transparent window in someone's skull so they could ultrasound the brain - https://keck.usc.edu/news/usc-researchers-pioneer-new-brain-imaging-technique-through-clear-window-in-patients-skull/ . It's pretty cool, but the lengths they would go to in order to make this work suggest that there's not currently a simpler solution.
Ultrasound can penetrate the skull pretty easily, the bigger problem is signal aberrations produced by the surface wave and the skull having a different sound speed then just water or the brain.
I am doing research specifically on transcranial ultrasound based loosely on these techniques at a national lab, although I regrettably have to be sparse with the details. https://pubmed.ncbi.nlm.nih.gov/40993881/
You are very right in your assessment that computation is a limitation, after we run scans and get data running the back modeling algorithms takes a lot of time even with with only a couple hundred receive elements.
The first thing I noted when I heard of this, is just how much computing power this is going to take.
This is a very interesting technology that is still a fairly young field, so I'm curious to see if midjourneys idea goes anywhere
I would expect that the amount of compute would depend a lot on the algorithms used.
This feels like the sort of thing where there will be some clever algorithm that gets the job done much faster.
If there is an algorithm for that, it will probably exist in the brains of dolphins and whales.
Open question whether it's easier to discover the algorithm or learn how to scan a dolphin's brain.
The last ~10 years or so there has been a lot of development in water surface penetrating laser scanners (green laser), used for doing bathymetric surveys f.ex. I once watched a presentation from a guy who's phd helped pioneer this technology. As I understood it the breakthrough here was do to lots of complicated maths interpreting the laser signal subject to various refraction etc. breaking the water surface. I don't know if it is similar at all, but thought it might be worth mentioning.
Interesting, you're the doctor, but Claude disagrees (about it being cartilage, not about the gaps). I asked:
> Is the skull of a 20w gestational age fetus more accurately described as cartilage than bone?
And Claude said:
> No — that's a common intuition but it's not accurate, at least for the part of the skull people usually have in mind.
>
> The cranial vault (the flat bones: frontal, parietals, squamous occipital and temporal) forms by intramembranous ossification, meaning it ossifies directly from mesenchyme and never passes through a cartilage stage at all. Ossification centres appear around weeks 7–9, so by 20 weeks these are genuinely bone — just thin and incompletely mineralised, with large gaps between the plates. Crucially, the soft spaces between them (sutures and fontanelles) are fibrous connective tissue, not cartilage. So the "softness" of a fetal/neonatal skull, and the ability of the plates to overlap during birth (molding), comes from thin membranous bone plus fibrous gaps, not from a cartilaginous skull.
I mean I definitely don't think doctors are wrong that it can't go through bone in the sense of "gets reflected heavily with poor transmission", I'm not saying they missed some blatantly obvious fact.
The fetal ultrasound is just demonstrating that when the bone is "thin and incompletely mineralised" you do get enough transmission for regular ultrasound wands to work. This to me just shows that's it's obviously a quantitative problem and not absolute - bone five times as thick will have to-the-power-of-five as much transmission which would be extremely little, and the reflections complicate things for sure (though I am sceptical that the processing will be as intractable for AI labs as for academia given the large gap in brainpower between them - we will see).
So it's just that the statement "it can't go through bone" is too absolute - as you said, the truth of the matter is more quantitative - how much signal to noise do you have, and can you process it. Clearly adult skulls are too thick for current approaches.
Fair, I guess I got the gestational age wrong.
I don't think the claim is doctors missed it in the sense that 'ohh look I can see' they missed it in the sense that there is a lot of shit you can even see just by shining visible light through your body if you do enough smart things to pull a weak signal from the noise and integrate over long time periods.
One obstacle is that at a certain energy, the ultrasound beams do actually start to heat up your tissue (or in the case of lithotripsy, break up kidney stones). Diagnostic ultrasounds are currently rated *well* below the amount of power output that would produce a physiologic effect. But if we find that we can ultrasound through bone but that it requires more output, this could present a barrier.
A very very common procedure in cardiology is TEE (transesophageal echo). You sedate the patient and place a cardiac ultrasound probe in the esophagus. This is most often used for posterior structures that TTE can't see well (ie the left atrial appendage) or when there's limited windows due to body habitus or ribs. The fact we go to such lengths on a *routine* basis also suggests there's not currently a simpler solution.
Sure you could use some statistical techniques to pull signal from behind the bone, but what do you realistically expect is the maximum amount of signal you could pull? Any large lesions or tumors are going to need an MRI anyway, anything smaller cannot be reliably reported to the patient as anything other than noise.
Ultrasound is already the cheapest it can get, so the only benefit to this technology is some unreliable behind the bone scanning, which will almost always need further scanning anyway.
Just make MRI machines cheaper. China is already working on ultra low cost MRI machines and scanning.
Very good write up. medical doctors are probably in the top 10% of risk adversity, I would guess. So its natural for them to be skeptical of new tech even if it is potentially useful especially once trade offs are considered.
Where do you get "top 10% of risk [aversion]," or that doctors are generally skeptical of new tech? Doctors picked up on technologies like mRNA vaccines, CGMs and insulin pumps, and wearables pretty quickly.
Overall I think this is a solid analysis, and I think the Scanner deserves between the level of hype Twitter is giving it and the amount you are giving it.
I do think it’s possible that this ends up being on net saving a large number of lives, and so relative to expected alternatives I’m very happy that an AI image company has gone into not-yet-AI imaging!
One thing that’s a material difference understated in your analysis between yearly and weekly scans. If it’s important, you can get a yearly or couple times yearly MRI.
Midjourney claims that their per-scan costs will be negligible. This implies that your spa membership will likely include ~unlimited scans.
I think that the data you can get from doing a weekly time series of your body could be substantially better for diagnosis than getting a once a year snapshot, especially for reducing false positives!
Not all false positives will be eliminated by this, but if you see something transient and worrying on an MRI it’s a pain and a real cost to somebody to get another one.
If you see something concerning on a Midjourney Scanner, you can monitor it as frequently as you would like! This makes it easier to refrain from doing potentially harmful interventions, versus having to wait 6 months for another scan, stressing about if it’s continuing to grow.
I think that alone can make it worthwhile; and if they *can* better map air and bone with their setup as they claim, then it seems clear the hype isn’t overblown. It’s not clear they can, but I believe they believe they can, and so I think it’s great that they are spending their private dollars on this bet!
"I think that the data you can get from doing a weekly time series of your body could be substantially better for diagnosis than getting a once a year snapshot, especially for reducing false positives!"
I think this is misunderstanding the threat model. "False positives" are real consisting things in the body, they just don't matter, and they can't always be distinguished from things that matter by their growth rate.
With weekly scans it wouldn't just be growth rate, but permanence. A "zit" inside the body that grows quickly but disappears 3 weeks later will be noticed and filtered out by weekly scans. Annual scans either see the zit or don't see it, but that doesn't help very much, does it?
Unless you're lumping in some fast-growing benign tumours into the "false positive" category. If you've got a benign tumour that takes 2 years to grow to the size of a watermelon, wouldn't it also be better to catch it earlier, even if it isn't cancerous?
Would this be useful for those of us who can't have MRI scans? I have a pacemaker embedded under my skin and this means I'm not allowed anywhere near an MRI scanner.
Potentially yes.
>Using this intuition pump, I don’t think Paul’s argument stands. Most pimples come after birth! Even if you got a skin scan every month and could compare the current skin scan to the previous one, you’d still be at high risk of fretting over pimples and demanding biopsies for them.
What?? This is the opposite conclusion you should be drawing from this intuition pump. We DO get regular skin scans, multiple times per day in fact, because we CAN see our skin with our eyes and therefore notice the irregularities and changes. And yet people aren't constantly running off to the doctor to get their pimples and moles biopsied. Seeing it with high frequency allows a better understanding of what is and is not concerning, which desensitizes people to weird but benign stuff happening. There are false positives and there are false negatives, but it's not some constant crisis for normal people seeing stuff change all the time, because you just get used to it.
I think the reason we don't freak out about skin irregularities isn't that we see our own skin regularly, it's that we see other people's, and have a robust cultural awareness of skin irregularities.
Yeah, maybe an awkward analogy because it combines two things:
- You can see that pimples exist
- You can look at the pimples closely with the naked eye (and eg see that they don't look like melanoma).
I'm imagining that a scanner could see that a pimple existed, but couldn't look at it and see what it was without a biopsy. I think this is closer to the real situation.
What exactly would stop us from doing all the preventative screening in the world, and then have doctors simply not inform the patient of the things that seem harmless. the patient is no worse off than they were before, and it seems to eliminate the EV negative side.
If the fear is that the threat of lawsuits will make doctors always err on the side of disclosing too much, then this seems like much less of a problem for less litigious cultures than the US, which is essentially every culture on Earth
What do you mean by "not inform"? We're talking about scans here. Presumably people will have access to the actual images produced. Certainly there'd be rioting in the streets if you tried to stop people from viewing their own scans. And many of the actually-not-cancerous growth are visible to the naked eye for a layperson.
As I said, I don't fully understand the tradeoffs - but at least one answer to your question is that I think patients have a right to their medical data by law.
I've heard about this problem with lawyers before (but haven't experienced it myself). Someone got cancer in a specific place? Look through the back-catalog for *anything* weird at that place, the slightest shadow, and sue the doctors who didn't pursue it.
There are things the doctor concludes are harmless. There are things the doctor concludes are harmful. And then there's the middle ground. Lots of room there.
> If you could guarantee that the diagnostician would be perfectly rational (because it’s an AI) and perfectly immune to patient pressure (because it’s an AI)
You can't, though. Even the most casual interaction with LLMs will tell you that they're neither perfectly rational nor immune to pressure. In fact, ChatGPT's susceptibility to bullying is famous. Some people see flattery as the primary use case for these machines, and what is flattery but submission to social pressure?
You could use a narrower non-LLM model which just spits out raw percentage chances of whatever conditions your interested in, no?
I don't think the auto-diagnostician would look like a chatbot. It might be LLM-based, but the prompts would be highly standardized. You would send it the image, it would apply the standard factual prompt, and you would never have a chance to flatter it or anything.
That doesn't guarantee that what's inside the mostly-black box is perfectly rational. It does shut down most sources of social pressure, but perfect rationality is simply not what this branch of technology offers.
I think you can train an LLM to follow rational rules and do pretty well.
Consider for example that most Claudes will do well on lots of probability problems that would stump most doctors.
You can, but due to all of its hallucinations, on average it is far less than "perfectly rational". But to be fair, the ultrasonic scanner would probably employ some kind of classifier rather than an LLM, so it might be less prone to hallucinations. No matter what though, neural networks are stochastic, and are thus not "perfectly rational". To be fair, the same is true of humans.
Chatbots are (normally) stochastic. Neural networks are not. LLMs add the stochasticity at the end by sampling from the logit distribution.
That certainly doesn't make neural nets "perfectly rational," but I think the irrationality is almost entirely unrelated to the stochasticity.
Randomness is also introduced during training, as it is performed via stochastic gradient descent. But you are right, by and large the irrationality of NNs is not merely due to randomness; I was merely pointing out that *perfect* rationality is unachievable due to randomness.
Hallucinations are way down this year. Can't remember the last time I saw one. This does not tell you good things about my recall ability.
FWIW, I see hallucinations pretty regularly. I do a fair number of highly specific Google searches, and when you put in words with no obvious connection to one another the search-support AI can get a bit desperate with its interpretations.
Which is one reason I try to ignore the thing, but my eyes read automatically.
I saw one yesterday, admittedly from free chatgpt.
Cost me 20 minutes of chasing down a phantom before I concludes it was wrong.
Being good at math - or good at applying logical rules in general - is not the same as being perfectly logical. As legions of irrational mathematicians and philosophers can demonstrate.
I'm very good at math and I know a fair bit about probability but I don't necessarily make good decisions.
I'm not trying to say that these machines are useless. But we have to stop assuming that our various irrationalities are somehow inextricably tied to our meat. Artificial emotion is as real as artificial intelligence. Artificial stupidity, artificial insanity, artificial bigotry...we see it all, because silicon-based minds can do anything carbon-based minds can do.
I think the point in this case is you can tune the classifier until you get your preferred ratio of false positives to false negatives, then freeze it and not worry about whether the patient is paranoid or the doctor is scared of lawsuits.
I think this is an interesting development from Midjourney. But the discussion misses the point a little bit: the premise of AI is that collecting massive amounts of data allows you to sift out emergent patterns. It's nothing like the scientific method, there are no hypotheses and no false positives in the collection and the analysis. I can imagine a world where you take a billion scans, run a vast amount of future-learning algorithms on the data and then infer lots of currently non-obvious results. Imaging brains and lungs from complex refraction and reflection patterns is one possibility. Classifying polyps into dangerous or not while they're still tiny seems perfectly doable, or recommending physio to people before they fall and break a hip, or detecting plaque buildup before a stroke or heart attack and treating it then. Why should we wait until there are symptoms to treat, rather than acting pre-emptively? The only reason currently is the lack of data - maybe this supplies it. More likely, the result will be something no-one even thought of yet, because we're still at Henry Ford's 'faster horses' stage of developing the technology. And most likely, it'll be of no use at all.
On a different dimension of questions about this technology: How different will their "using AI to reconstruct images from ultrasound data" be from the currently-existing using AI to do the equivalent of a CSI TV-show "enhance" to significantly upscale images?
The AI will presumably be trained on a bunch of samples to produce images that look like the samples and are reasonably consistent with the incoming data. However, one of the problems with using current AI imaging models to do this sort of thing (in, for example, the upscaling case) is that it does not distinguish what is solidly supported by the data from what is a merely plausible guess that is not directly contradicted by the data. As a result, it's very easy to use the current technology to produce things that are almost entirely guesswork with a very loose connection to reality.
That may be fine for artwork in the hands of someone who is expecting a mostly-fictional guess, but that is not at all what one wants in a medical scanner.
Do the people who are going to be creating this device have the necessary background and understanding to distinguish what's data-supported and what's guesswork in the created images? And are they going to design the UI and output of the machine in such a way that medical technicians can also distinguish the two?
If Midjourney is serious about building this technology instead of running a full scam, then the answer to your "How different" question is "very".
The difference is that "what is solidly supported by the data" vs "what is a merely plausible guess that is not directly contradicted by the data" is a much steeper gradient when you have a large enough amount of sensor data. You verify if the image/physical model matches the sensor data by solving some differential equation (the forward-modeling). And in the other direction, traditionally you either are lucky enough to process the data with some numerically cheap way, in which case you won't need much of AI's help, or have to guess an initial model, solve the equation, compare the answer to the data, and do this many times intelligently to minimize the misfit (the inversion). Most of these problems are ill-posed, so even traditional methods only give you a "plausible guess" in some sense, but that's typically good enough, and there are ways to determine how uncertain the inversion result is.
AI methods can step in at different points. The training data would be data you generate from physics-based forward modeling, and you can train the AI model either in the forward direction so that it can be a fast surrogate forward model (typically solving the differential equation is the slow part of the process, so if AI can do this, it'll already be a huge speedup), or in the other direction so that it can produce an inversion directly.
The structure of the problem is so different that I'm struggling to analogize the space to generating artwork from prompts, but how it should work is not a mystery. To add a bit of context, I work in an adjacent field where some of my colleagues are doing precisely this kind of work, training AI models to replace traditional physics-based inversion. And to be sure, most of these people are physicists and applied mathematicians, so I kind of share your concern (as well as the concern of a few other commentors) from a different angle, that is, if a lot of the expertise required is physics, what is Midjourney specifically bringing to the table? I don't have an answer, but the optimistic stab-in-the-dark guess is that they got especially clever with designing neural nets that process images somewhere along the way.
For the intuition pump about skin conditions that you had to scan for: I think if you run a filter where you only bother about something if it persists and maybe grows over the course of several months, I think you cut the false positive rate down far enough that it might be worth investigating any remaining hits so long as you could do so not-too-invasively and not-too-expensively. Don't know if this analogy would hold for internal organs.
> One possible answer is convenience. Midjourney claims these ultrasounds could be much faster and more comfortable than an MRI (which involves ~60 minutes in a giant scary metal tube that sometimes kills you if you forget to take off your jewelery).
Is it possible to do what they're trying to do with a CT scanner? I feel like you're underselling just how inconvenient an MRI is. It makes loud noises (that are loud even *with* the hearing protection they give you) and you gotta remain still for a long time. CT scans are much easier to undergo.
CT scans still aren't much fun and importantly they give you a non-negligible dose of radiation.
I need to do yearly MRI. Are they fun? Not at all but the whining about them is largely overstated. The noise is real but you learn to zone it out.
For people with claustrophobia, they can either get benzos or one of the newer, smaller machines...
I mean I'm just saying this as someone who had to have a bunch of scans last year and I gotta say I found the CTs way easier to deal with.
That still sounds like a huge difference in practice, like the difference between an injected vaccine and a pill.
CTs give too much radiation to be a good choice for elective scans.
Oh, dang, didn't realize. Guess that makes sense, they do have you put on the shield...
I find MRI scans comfy and relaxing. Anyone else like them?
Yep! I've fallen asleep in 2 of my last 3. They really don't bother me at all, I just don't have claustrophobia
> This is an interesting take. If you could guarantee that the diagnostician would be perfectly rational (because it’s an AI) and perfectly immune to patient pressure (because it’s an AI), would that switch the sign of primary imaging-based screening from negative to positive?
Perhaps you can make a perfectly rational diagnostician, which is immune to patient pressure, but you can't make a *system* of diagnosticians that is immune to patient pressure because the patient can go elsewhere.
Huh. I have no idea about the medicinal background, but I always thought the advice against population-wide scanning was simple Bayes, i.e. if the incidence of a condition is low enough, even the test with almost perfect specifity (on an order of magnitude that makes any sense) would have a large false discovery rate, just mathematically. Am I off-base with that?
That’s the usual textbook logic yeah, but there are other knobs you can play with to make the usual logic maybe not apply. Maybe if your test gives a number between 0 and 100, instead of freaking out wherever the number is above 50, you wait until it’s above 90, or even 95, which reduces the false discovery rate to maybe something manageable. Or maybe there’s a cheap follow up test that’s easy enough that you’re okay with the screening test having a crazy false discovery rate.
I am satisfied to read this post right after finishing Zvi's weekly AI roundup, which dropped earlier today, and similarly highlighted the MidScanner as an interesting development. The future moves fast, and getting an Actual Doctor's Opinion was exactly the follow-up I was hoping to go look for, so now I don't need to.
Got introduced to DEXA scans years ago by someone with pretentions of high-powered athleticism (who predictably ignored all the fine-grained recommendations, that's a true stereotype), and although I don't have any "real reason" to use them, it's still an interesting experience? Filling out more data on one's character sheet is intrinsically satisfying, and routine medical care doesn't give most of that information. There's certainly some kind of, eh, moral medical hazard to beware of...but that seems like more of an issue with scrupulosity and poor intuition than a fault of the information-giving source? Personally I just found it really cool to get hard proof that, yes, regularly exercising was paying good muscle mass dividends, even if I'll never have a washboard stomach or similar highly-visible signifiers. Which is the reverse intuition from pimple biopsies: it's Good, Acktually to know that good things are happening behind the scenes, when undeniable proof like six-pack abs is already so far along you no longer need the confirmation. Flywheel effects are quite motivating!
The screening MRI cost/risk conclusion here in Taiwan is the opposite of the AMA consensus. Not by a little, it’s wholesale, and not literally but practically mandated by labor law.
Basically everyone in this hood with a full time professional job at a company with >10 people gets a non-MRI full body set of exams every two years. And doing this is a legal obligation that falls on individual citizens here to get done and to provide the results to the HR department if you join a company.
There are thousands of companies here that have enough profit or funding to spend the extra $1000/year, so that people can get full body MRI with initial follow up CT scans if anything from the MRI warrants a higher res look. And basically every company of any type has a majority of the deal with the health exam companies so that if an employee wants to go for the more extensive testing, They can pay an out-of-pocket fee that’s about half of what the retail rate is for the incremental, more sophisticated tests.
Health regs for this: https://laws.mol.gov.tw/flaw/FLAWDAT0201.aspx?id=FL015034
One of the exam packages, this one is recommended every 2 years and has 105-107 areas included in the various tests, is what we’ve got:
https://www.eonway.com/wp-content/uploads/2024/05/3-1-Tumor-Screening-Package-100000-2-day_Examination-Items_20241206-1448x2048.jpg
I was told something similar about Japan.
Is there any evidence that these scans deliver health benefits, or that these benefits exceed the costs, including the financial and stress costs of "false positives" that Scott's post highlighted? I'd like to think the answers are yes. In the UK we delude ourselves that we have the best healthcare system in the world, so we obviously don't have anything like this!
They definitely deliver benefits in terms of some probability of catching cancer early. Not super huge benefits
How this trades of with stress and finance isn’t clear, utility functions vary
Very interesting, but are there any studies showing this is effective in those countries?
I asked all the llms and the median conclusion was basically “if you a clear minded and wealthy private individual of middle or late age, it is defensible but by no means obvious or necessary”
Can’t be done for the whole population for financial means, so it isn’t surprising that we have a good explanation to give doctors for why it isn’t beneficial for genpop
Typo:
> is no doubt be very
Should be “is no doubt very” or “will no doubt be very”
Thanks, fixed.
I don't think this is a scam (though I do think Midjourney are hyping it up too much, but that's normal).
But I am surprised at the lack of scepticism from the AI community, including from quite sensible people, about something which seems quite improbable on the face of it (an image model company pivoting into medtech and immediately inventing a revolutionary new technology). This makes me think we are probably going to see actual scams emerge out of the AI space soon (and in fact I wouldn't be surprised if at least one neolab is dodgy)
This reminds me of Demis Hassabis' story about how he was explaining AI to Mark Zuckerberg, and Zuckerberg seemed appropriately impressed and excited, and then they started talking about random silly technologies that wouldn't work (eg the metaverse?) and Zuck seemed equally impressed and excited about all of them.
I'm increasingly worried that many people who are excited about AI aren't actually smart enough to realize it's a good technology, they just have a "technology always works" prior (admittedly a refreshing change from the usual "technology is always bad" prior) and this will become increasingly concerning as their successful AI ventures elevate them higher and higher.
The LJ-99 room temperature superconductor thing a couple of years ago seems a good example of this. Lot of people got very excited by something that the scientific priors were heavily against
I think we're just beginning to see the scams arriving, as ever criminals have been in advance of the rest of us in adopting new tech in order to better victimise people.
I agree the most likely result is that this will be part of a spa/boutique private health clinic treatment package, not widely adopted in general hospitals (imagine trying to fill and empty baths for 15-30 patients a day) - a new spin on "taking the waters".
At least this AI pivot shouldn't do as much harm as radium:
https://cosmeticsandskin.com/aba/glowing-complexion.php
"Medical and other uses of radium
Medically, radium was usually injected or taken in pills. It was used to treat a wide range of ailments including hair loss, impotence, atherosclerosis, high blood pressure, rheumatism, gout, sciatica, nephritis and anaemia. This led to a craze for radium-based products, and radioactivity in general, during the 1920s and 1930s. It was added to a wide range of commercial products including: wool for babies, water dispensers, chocolate, soda water, male supports, foundation garments, condoms, toothpaste, suppositories, cigarettes, cleaning products, boot polish, fertilisers, luminous paints and cosmetics. All of these commercial applications had a common theme; that the rays given off by radium had a ‘vitalising’ effect on the human body.
Cosmetics
A number of companies put radioactive materials into cosmetics or made their cosmetics radioactive using the process of induction; i.e., making them temporarily radioactive by exposing them to a radioactive source. As the original work on radium was carried out by the French scientists, Marie and Pierre Curie, it is not surprising that this was more common in France than elsewhere. England also had a number of cosmetic companies that used radium but the same cannot be said of the United States."
It's the latest, cutting-edge science, it must be safe and effective!
This attitude toward radium shows up clearly if you read the Oz or Conan books written at the time.
I wouldn't be so quick to dismiss this. No doubt your clinician-followers are smart and competent - but they might also be a tad cynical and burnt out.
The Midjourney Scanner page specifically mentions dolphin echolocation as an analogy. Dolphins use lower frequency sound waves that can penetrate more deeply than medical ultrasound, and image bone. Yes, longer wavelengths sacrifice higher-resolution, but with with a few tricks borrowed from super-resolution microscopy (which sees objects smaller than the diffraction limit of light) it might be possible to "break" the limit of longer wavelength sonar.
A new high-resolution, full-body, non-ionizing imaging modality could compliment MRI - and with some fancy AI analysis could reduce false-positives and actually prove useful. Who knows?
Or, this could be a vaporware scam. :-)
I don't think dolphins can see through bone (and it doesn't seem relevant to their use case of eg finding fish). Where did you hear that they could?
Ultrasound tomography through bone seems like it should be an easier problem than seismic tomography through the entirety of planet Earth!
A quick DDG search makes the application of seismic tomography techniques to medical ultrasound seem very promising: https://www.imperial.ac.uk/earth-science/research/research-groups/transmission-and-reflection-ultrasound-tomography-trust/research
(I especially like that there's a fully open-source seismic tomography software stack, using GPU compute, originally developed by (and for) geophysicists, now being used for medical imaging: https://www.stride.codes )
I’ve done both, ultrasound through bone is way worse.
Fascinating! Worse in what sense? Just a less well-developed field? Or more difficult/limited in some fundamental physical way? (Stricter constraints on non-destructive frequencies and amplitudes, perhaps..?)
Even if this were an obviously good product, I would wonder whether Midjourney was positioned to pull it off. They will need to bring in a lot of expertise in ultrasound sensing/medical devices/manufacturing and build a functioning team quite quickly. This stuff is all slower and more painful than developing a software product.
If I were them, I would be looking for an existing innovative medical devices company to partner with/acquire/buy a dept from.
Doctors are famously bad at probabilistic reasoning. Any new testing technology that has a different statistical threshold for clinical action than what they are used to is going to hopelessly confuse them. It will take decades for new generations of doctors to act sane when exposed to new kinds of tests with high rate of false positives.
In this case, usually insurance companies can save you by acting like designated party poopers and refusing to cover the treatments and additional tests that the doctors will ask for just in case. But it won't save the people who pay out of pocket, and it will be very poor consolation for the occasional patient who turns out to be a false-false-positive or the CEO who gets murdered in retaliation.
> If you could guarantee that the diagnostician would be perfectly rational (because it’s an AI) and perfectly immune to patient pressure (because it’s an AI)
Is this some kind of hypothetical AI we don't yet have? Because AFAICT LLMs aren't perfectly rational—they may be less biased than many humans, but they do have their own systemic biases—and they are also famously, uh, mune to client pressures.
I don't think client pressures are a problem - I don't imagine this auto-diagnostician as a chatbot you can talk to, just as a pipeline you submit the scan photos to that has a factory prompt and returns its opinion.
I also think you can quantify/standardize/avoid AI biases better. For example, I bet there are a lot of tough probability questions that Claude would get right 100% of the time, and randomly selected doctors much less often.
> they are also famously, uh, mune
Etymology watch:
𝘮𝘶𝘯𝘶𝘴 (n.): "public service", tax in whatever form it's assessed (money, unskilled labor, military service, putting on a festival...), or a donation to the public coffers.
𝘪𝘮𝘮𝘶𝘯𝘪𝘴 (adj.): describing someone who doesn't have to pay taxes.
Would this be useful for babies? I had a child in the nicu, who got quite a few ultrasounds of heart, brain, and abdomen. Also a CT, which required general anesthesia and thus intubation, and still didn't hold still well enough to get the extra clarity over bedside ultrasound that the cardiologist wanted.
It would work for babies, but I'm not sure it would be better than existing tech. I think whether it could replace CTs would depend on the exist CT indication. I also don't know how still the baby would have to be in the water bath.
As far as using AI to normalize interpretation of other types of scans, just a couple months ago I underwent what Mass General described as an AI-assisted MRCP (magnetic resonance cholangiopancreatography--that is, an MRI focusing on the bile ducts) as part of the treatment for my primary sclerosing cholangitis.
As far as accuracy goes, its conclusion that the liver was heavily scarred and the bile ducts constricted was at least supported by the doctor who later attempted an ERCP (endoscopy into the bile ducts to try to clean them out manually), who couldn't even access the left side and who told me there wasn't any point in trying because the scarring was so bad.
I was a little confused when I didn't see it mentioned in any of the comments in the post. Maybe it's just a new enough technology that not many places have it? Proximity to a facility like Mass General was part of the reason I moved to Rhode Island in the first place. I did find this study from 2025, so it's not unheard of as a concept:
https://pmc.ncbi.nlm.nih.gov/articles/PMC11809031/
I would also like to register my annoyance at the lack of high-quality audio recordings of MRI procedures. There is one channel on YouTube which has a few recordings, and even names the different scanning processes by sound, but none like what I've heard in the MRIs I've had (and between my Ulcerative Colitis and my PSC, I've had a few, though I don't go out of my way to get them).
Seriously, there are parts of even a regular abdominal MRI which wouldn't need any editing beyond cutting the number of pulses to a multiple of four in order to make some truly great EDM. About two thirds of the way through the MRCP, I started grinning like an idiot because the rapid clicking, without various high and low pitches cycling on and off at regular intervals, was so perfectly musical.
Worth noting I always bring a set of foam earplugs to go on under the headphones (I always have a bag of them on me when I go out), but since moving out of Arizona I've never needed to use my own because the facilities have always offered them. This turns MRIs into what I imagine guided meditation must be like: wearing loose, comfortable clothing; lying on a soft, padded table; and listening to breathing instructions for some 20 minutes.
I don't think the noises you here in the MRI are entirely sounds. I think some are the changing magnetic fields directly stimulating your ears
I appreciate applying the principle of charity and arguing about the merits of the proposal as presented by Midjourney. But I just can't ignore that it's Midjourney specifically proposing this. A company, whose main area of expertise is creating very plausible and realistic looking images that do not correspond to an actual state of the physical world, is claiming that they've pivoted to solving a very different set of technical challenges to build a system that (presumably) still outputs a realistic looking image.
This feels like if the guy famous for magic tricks involving humans hidden in ingeniously constructed boxes announced his pivot to building mechanical chess robots. It's not like there's zero overlap but it's a bit too much of a coincidence.
I love your analogy and I’m surprised not to hear many others talking about this.
This was the first thing I thought of. If your entire business is making up fake images, becoming a tomography service is a bizarre thing to do.
Also, I don't really see the overlap, in this case _or_ in the hypothetical case of the guy who's famous for stage magic pivoting to chess robots. The problems aren't the same. The approaches aren't the same. There's no useful transfer from doing one thing to doing the other thing.
I'm not really worried about this; this would be past my https://www.astralcodexten.com/p/bounded-distrust level.
This is a new technique that will yield results different from those of other techniques.
It'll take time to relate the results to those of other techniques and to the bodily problems of interest.
However useful (or not) it may be initially it can only get more useful with further development.
BTW: "The bowels are full of air" that ain't air!
"even when entrepreneurs set up their own full-body MRI clinics that don’t cost the system anything, the medical system recommends against using them"
the system can hardly say "this is dangerous when ordinary people do it but it's harmless to billionaires paying out of pocket"
besides, demand is demand. building your own MRI still raises demand for everything that goes into building an MRI, hiring a radiologist still raises demand for radiologists, etc.
1) This is all entirely hypothetical. We have no idea what this tech actually does or what the use cases might be.
2) Detecting cancer is relatively easy. Proving that that detection has a positive effect on people's lives is hard. Prostate cancer is common yet most cases cause no harm while others are fatal. Figuring out how to deal with that is still an ongoing issue.
There is vast literature on failed screening attempts. People might want to look at same arguments about the NHS-Galleri trial, a major cancer screening test which failed its most critical trial.
I didn’t know you were a psychiatrist. Psychiatry seems to be a unique branch of medicine that has so far avoided the march of instrumentalism in medicine. There are no probes or blood tests or any physical objective correlates of mental states such as depression.
Are you also seeking an instrument to observe the mind?
I think they do have some physical objective correlates that work in mice, but they also require an autopsy to measure.
Szasz was right. Psychiatry should not be considered a branch of medicine. Not sure what it is.
With regards to the whole body MRI and things that are “obviously bad” it’s helpful to think in terms of exactly what you are trying to catch
H. Gilbert Welch divides cancers into birds, rabbits, and turtles.
Birds have already flown away, widely metastatic cancer that either respond to immunotherapy anyway or kills you. It doesn’t matter how early you catch them
Turtles were never going to cause an issue anyway. They grow too slowly.
So what you need to catch are the rabbits — cancers that can be surgically removed before they escape and become incurable. What cancers can an MRI even plausibly detect? We have better, more reliable screenings than MRI for breast, skin, colon, cervical, and lung cancer.
MRI is bad / mediocre at catching GI cancers and uterine cancer early
So really what you are specifically searching for is early pancreatic cancer that’s surgically resectable, ovarian cancer, renal cancer, brain cancer, throat/neck cancer, thyroid cancer and bone cancer
Many, many screening strategies have failed for pancreatic, thyroid and ovarian cancer. They’re too aggressive and hard to see. However this is legitimately a place where MRI could be useful. Renal cancer and prostate cancer are usually turtles. Brain cancer is almost always either a turtle or a bird. Bone cancer is extremely rare. Neck and throat cancer is rare aside from smokers
What about non cancerous diseases? Spinal bone findings are famously worthless in asymptomatic patients. Simple blood work is better for fatty liver and CKD. We still don’t know how to treat most things a brain MRI can see. MRI is a bad test to catch blockages in arteries.
So overall the utility of whole body MRI in non smokers is limited to possible early screening for pancreatic, ovarian, thyroid and maybe prostate /renal cancer, all of which have long and failed histories of screening tests in clinical trials. That’s the benefit you’re trying to achieve. And it’s an impossible benefit to get without accepting a LOT of false positives
There is a screening program for pancreatic cancer in Czechia, but for people with history of pancreatitis only. Allegedly this makes the program efficient.
Would definitely make sense. Pre test probability matters quite a bit for these tests. It’s why only smokers are screened for lung cancer
Most head and neck cancer these days are HPV related. Most thyroid cancer is a turtle. Screening programs detect the slow growing and miss the very rare fatal cases.
I think doctors are working on a principle of don’t bother with a test if the result is irrelevant to what you do. If the rate of false positives is high enough that you’re just going to assume any positive result is a false positive, the test is not worth doing.
(The implication would seem to be — only run the test if you have evidence on other grounds that the patient is likely to have the thing you’re testing for)
> Using this intuition pump, I don’t think Paul’s argument stands. Most pimples come after birth! Even if you got a skin scan every month and could compare the current skin scan to the previous one, you’d still be at high risk of fretting over pimples and demanding biopsies for them.
Not getting this. As I see it, the skin comparison support's PG's argument.
The reason we don't freak out about observed anomalies on the skin is because they are routine, familiar, and understood-harmless.
The reason we do freak out about observed anomalies in deep imaging scans is because they are surprising, foreign (to experience), and not understood at all (again - according to patient experience). But the surprise is because we rarely look - not because the anomalies are rare.
If we had frequent high res scans that frequently showed blips coming in and out, and if our siblings and neighbors and parents had the same, then our reactions would settle pretty quickly to treating these more like pimples than potential cancers to be cut out.
Incidentilomas are real and a real problem. But it is also very difficult to access medical care and run the gauntlet in order to get any testing. Something like a connective tissue disorder, or autoimmune, or rare musculoskeletal diseases are impossible in many locations to get any medical help. Everything requires a code, with the doctors working for the insurance companies. If getting the code is difficult for a doctor or requires arguing, or extra work, the aren't going to often do it. Their time and the hypocratic creed gets ground down by insurance bureaucracy and $20 an hour sparsely staffed call centers, designed for the grind. In other words I feel there is much real estate between incidentilomas and having some ability for patients to self help, with ai tools, and hopefully medical oversight.
Isn't the world better off with more investment (of time and dollars) towards innovative medical scanning technology than another - increasingly commoditized - image generation model?
Does weighing in with a bucket of cold water (as an influential voice, particularly in the target group of people who might decide to make those sort of pivots & investments) promote or deter others from following that same path?
(There is also a point around - if you believe AGI is coming in <10 years - and you understand the "bitter lesson" - is a technology that brings 1,000x the resolution & data that exists today likely to be useful or useless?)
Somewhat paradoxically, if "superintelligent" AI is coming in 10 years, then anything you do today -- be it technology, art, politics, etc. -- is useless, unless it directly impacts said AI. Either the AI is going to be malicious, in which human civilization immediately ends and nothing you've ever done will matter; or it will be benevolent, in which case paradise on Earth is immediately ushered in, human civilization as we knew it immediately ends (to be replaced by something else), and nothing you've done up to that instant will matter.
First of all, I'm not really a pragmatist/instrumentalist. I think it's generally useful to say true things without calculating the second-order effects too hard. See https://www.astralcodexten.com/p/less-utilitarian-than-thou
But also, "don't criticize technology if it's in a good/important field" is a great way to have all of your good/important fields full of constant scams, so much so that nobody ever tries to do good work in them because all of the investment money flows to scammers.
I'm not calling this tech a scam! I think there are some scenarios where it might find niche uses. But I'm pretty concerned about people in my orbit talking about how now all other medical scans are obsolete, or how this proves that the medical cartel was keeping the miracle technologies from us before.
But also, people are constantly freaking out about the possibility of there being "hype" in AI. I think in order to keep actually important things credible, it's important to push back against some of the actually-hyped things so that people have some trust that if Silicon Valley is saying something will be very important, it might actually be very important.
I rarely agree with Scott on anything, but this time I agree 100%. Whether this is good or bad, I cannot say :-/
Alright, I like this for Midjourney, I think it's a cool experiment, I doubt this will come to a hospital near you any time soon.
Let's start with the medical response. I don't want to question doctor's medical expertise. They know their field, they're highly trained, and most importantly they're the ones directly talking with the patient and bear ultimate responsibility. But...these people cannot manage tech systems. If there's doctors in the audience, sorry, but you would still be using paper documents if we let you, Congress had to pass the HITECH act in 2009 to mandate you to use electronic records systems on pain of losing Medicare/Medicaid reimbursement and the first thing you did was walk straight into the arms of the Epic EMR monopoly which...charges a lot for not the greatest quality software and stifles innovation because they can just copy any new innovation and integrate it into your workflows which they monopolize. Like, I'm sorry, I can just walk into any hospital or health system in the country, see the software you use, and know that this is not your area of competence.
The only company I've ever seen crack this is Abridge with the AI scribes. Like, I know doctors are down on this, doctors are down on every new technology that could impact their patients and they have a long track record of failure, both in adoption and what they do adopt when legally forced to.
As for Midjourney, I think the tech is definitely there and I think the goofy spa thing is probably the right way to go. AI imaging is definitely not my area of expertise but I think the following dataset probably doesn't exist anywhere and is extremely valuable:
#1 Monthly or quarterly full body scans for a couple thousand people, preferably more than ten thousand.
#2 A full list of all their medical outcomes.
Like, I'm confident we have datasets where we suspect there's cancer, we send the patient to get an MRI, and then we determine whether there's cancer. That's a proven technology. But there's very specific and that's not really what we want. We want a full-body scan of everything, on a semi-regular basis, and then all outcomes over the subsequent years. We don't care whether it's cancer or arthritis or some muscularskeletal degeneracy thing or a fractured bone or whatever. The dream is you get a full body scan and then it diagnoses everything. And the best way to get that is....well, scan a bunch of people semi-regularly and then track their outcomes over time.
And how would I sell that, how would I build this dataset? Well, I can't go to hospitals, they can't figure out binary sepsis diagnosis, they'll never do this. But, like, what if I put a goofy golden spa where:
"Our spa will have hot tubs, saunas, cold plunges, and cozy rooms with pools of golden light which softly scan your body.
It should be a place you love going, whether it's by yourself, or with friends. It should be available 24/7. The scans are a side-effect. You barely think of them when going to the spa. But suddenly, you have a huge library of data about your health."
Like, what if we just sell a ritzy spa experience to rich technophiles in San Francisco? There's got to be tens of thousands of those. Midjourney can probably do this at a loss for awhile, undercost it and burn some VC cash, because if they've got weekly whole body scans for a substantial population and then they can get access to those patient's medical records and outcomes (which this patient population would probably sign off on)...woof. That's an insanely valuable dataset. That's a dream.
And there's a big, big feedback loop. Start by selling a spa experience to rich techies, cool, maybe all it is for now is a better Dexa scan. But once your population starts having health issues, once you've gotten weekly scans of, say, 70 people who got breast cancer and you start really being able to flag it in the early days, you can publicize some wins and start to get users outside the rich techie crowd...woof.
Because I think the fundamental is right. I don't think anyone seriously doubts the following: if you had a weekly/monthly dataset of full-body scans for hundreds of thousands of people and then were able to track their health outcomes, you'd be able to predict a wild number of bad health events.
TLDR; I don't think MidJourney is selling a mature product right now, I think they have a MVP that could really scale if they get users/get data. I'm not surprised that doctors and medical professionals are down on this; they have a really bad track record on technology adoption.
Blarg.
Thank you for basically writing what I was thinking.
I would add to this that I think people tend to have a strong bias towards first principles reasoning in cases like these, which gets in the way of being appropriately empirical. An argument that "all X is good for is Y" is usually not a great reason to avoid experimenting with other uses for X, and this is doubly true when you're talking about data generation. Ultrasound might have nothing to offer here, but we need way more, not less, of this type of experimentation. Whether that justifies any of the annoying punditry or hype Scott encounters in his milleu is a different matter.
> All of your organs are constantly growing little pimple-like things that don’t matter. Depending on the location and type, we call these cysts, polyps, fibroids, lipomas, adenomas, hamartomas, etc. We are bad at distinguishing these from dangerous tumors - and so far the studies show that when we try to do it in the entire population on a mass scale, we cause more problems than we solve. That could change in the future! But it’s a separate bet from whether ultrasound tomography will be good or not, and the ultrasound tomography won’t be very useful unless we solve this problem too.
How would this change in the future, if NOT by scanning more regularly?
I (not a doctor) am absolutely willing to buy the argument that omni-scanning is _impractical_. Maybe if we start today, it eventually saves 1 billion QALYs a year in early detection and treatment, but only after the upfront payment of 100 billion QALYs in medical anxiety and expense.
But practicality (and the Midjourney specific stuff) aside, it just seems that... diagnostic data should be commutative?
Say modern medical has a rule, "If you're coughing up blood, and the MRI shows a lump in your lungs, you have lung cancer."
Patient A presents with blood-coughing. They're ordered an MRI, which shows a lump in their lungs. The doctor says, "It's probably lung cancer, here's your costly and uncomfortable but life-saving treatment plan."
Patient B shows up for their appointment. Due to a clerical mistake, before seeing the doctor, they're sent in for an MRI, which shows a lump in their lungs. The doctor says, "Unfortunately, our medical system is bad at handling uncertainty, so we're gonna have to recommend you have the same costly and uncomfortable treatment as Patient A, even though you're probably fine." Patient B says, "Well, I have been coughing up blood lately..."
Surely the doctor updates on learning this information, even if their diagnosis isn't allowed to reflect it.
> if you tell your hospital orderlies “please transport this frail 90-year old to my giant water-filled tank, and lower them in slowly” they will stab you with your own scalpel.
My laughter forced a bit of coffee out my nose.
There's another SF company doing ultrasonic phased-array stuff. Their goal is stimulating lucid dreaming though, here: https://www.prophetic.com/. Might find it interesting. I wonder if this means that these ultrasonic phased arrays have just recently gotten cheap enough to use, or that LLMs like to recommend them to founders, or something.
I would love a deep(er) dive from Scott about mainstream medicine's arguments against full-body scans and the counterarguments. I too have been struggling to wrap my mind fully around the issue and have noticed the frequent conflation of purely medical considerations with economic ones from many anti-screening people. At the same time, I understand the medical concerns aren't trivial.
In addition to raising the thresholds for action or developing new technologies, one could imagine AI-driven improvements to the analysis of existing imaging types. And perhaps there are ways to reduce the downsides of false positives (e.g., additional, more focused scans before any biopsies). Anyway, the point is I'd vote for a "Much More Than You Wanted to Know" and further exploration/untangling/speculation about all of this.
Well, it seems AI has now reached the point of all new technologies when it's peddled as a miracle cure/quack nostrum. The same as radium (stunning new discovery can only be good for what ails you!) and electricity:
https://www.nuclearmuseum.org/virtual/vex1/toc.htm
https://fineartamerica.com/featured/radium-beauty-treatment-advert-library-of-congress-serial-and-government-publications-divisionscience-photo-library.html
https://exhibits.library.duke.edu/exhibits/show/good-vibrations/advertisement-and-consumerism
The AI ultrasound bath reminds me of the Faradic current baths, which it seems are still going, and traditional hydrotherapy:
https://www.fixhealth.com/blogs/how-faradic-foot-bath-helps-people-with-flat-foot
Though I had not previously known about the boob bath instead of having your breasts smushed on the X-ray plate, definitely making progress there!
Here is a question right up the alley of the "rationalist community".
There is a famous named law about the observation that most journalistic output about your own area of expertise will be inaccurate so you should assume that most journalistic output is inaccurate (period) and that you just do not notice this outside your own area of expertise. Crucially the journalistic output sounds plausible to the layman. You have heard this one.
So now suppose you learned that a relevant AI player was trying to move into your area of expertise and you judge the product market fit to be ... well let's say less than perfect. Should this, by analogy to the journalism example, make you, in the parlance of this community, "update" your beliefs about product market fit in other potential applications of AI?
An acquaintance was late to board games this morning because his flight back from SF coming back from the release party was delayed and he needed some extra sleep. He doesn't work for Midjourney but for a supplier. Things that came up in conversation is that system apparently involves racks of Nvidia GPUs and they're not trying to store all the data they collect but process it as they go.
I agree that the Ultra-Ultrasound sounds like it might be a bad idea. But the general fact that a big (or medium sized) AI company is explicitly pivoting to medical scanning & diagnosis is very exciting for me. I have chronic "non specific" neck & lower back pain, which can be pretty debilitating when it's bad and affects many aspects of my life even when it isn't. I've seen so many different specialties and gone through so many different front & second line treatments. I've even been to the Spine Center at a pretty famous medical system, where the first doctor I met opened with, "There's a lot we don't know about the spine."
So besides being a seemingly very good fit for the things that AI does well (synthesizing a lot of information and recognizing patterns), medical diagnosis is also pretty much the only near to mid term AI capability that I personally care about. Claude is fun & saves time in some of my work, but is very much not worth the risk of AI-launched nukes IMO. An AI that can tell me how to not be in pain every day for the rest of my life, though... I'd take a little nuke risk for that!
Quoting Dr. Harris from the article: "To the lay person, it seems obvious that finding a cancer on imaging earlier is always better, but in many cases, it’s not only not better it can be worse! Especially on a population basis when you account for medical misadventures that may ensue from unnecessary or earlier or futile treatment."
I dislike this kind of framing, especially with the "lay person" condescension. From an epistemological standpoint, it is always better (basic information theory). It can only be worse because it is scaffolded within an inferior decision making system. The medical profession seems particularly confident in taking their current consensus best practices as fact, and one gets the feeling everyone within the profession is constantly getting a memo of what they're supposed to say. When talking with a medical professional (not all, but an annoyingly high percentage of the time), you always feel feel like you're talking to a generic representative of the entire profession, not a single human being.
As a trivial aside, I learned at some point that birthmarks are not—as the common wisdom suggests—perfectly benign. Pigmented birthmarks carry approximately the same cancer risk as ordinarily acquired skin moles, which is to say very little risk. However, that risk is definitely relatively much higher than any equivalently sized region of 'unblemished' skin.
All that to say: If you bother to monitor ordinary moles for malignancy, then you should also monitor your birthmarks for malignancy.
I had quite a large one on my right forearm and it went on me a few years ago very quickly. Two different melanomas, two different genetic signatures. Apparently very unusual; one of them was gotten rid of. It was a flat discoloration. The other one formed like a mole and I didn't notice it soon enough
Primum non cognōscere would've been a great title
It feels unlikely to me that this particular device is going to be useful in any meaningful way, but the medical community's fear of overmeasurement feels very counterproductive to the future of medical research. More preventative scans of healthy people seems obviously good in the long run; the trick is getting past the harm of acting on incidentalomas in the short run. Not all radiologists are opposed to this sort of thing - I used to work at one of these full body MRI companies and we worked with plenty of enthusiastic ones - but presenting this data needs to be done very intentionally.
It may be a faux pas to link to my own post from years ago, but https://ostro.ws/post-healthcare has some more thoughts about how we can massage the messaging to make more measurement less dangerous.
Isn’t this literally the plot of the movie Pursuit of Happiness?
A colleague sent me this.
This is vaporware, it cannot work.
When a solid body is immersed into water bubbles form on the surface. They are a problem. Even if you degass the water - which you have to do for good acoustical measurements - the skin will bring trapped air and fresh bubbles.
Supposedly the images are from a demo unit. It may be that there are enough gaps between such bubbles to allow for the propagation of the waves between them in a similar way as with infant skulls.
I have no idea how the images were generated by a company whose expertise is in generating images of things that don’t exist.
The problem with bubbles is that they resonate and appear much larger than their physical size. You don’t need them to be placed closely to completely dominate the reflection.
This is why, for example, cavitation-free propellers are a must for stealth submarines. Bubbles light up on sonar like a giant disco ball.
And it’s a spa. Meaning flowing heated water. Full of bubbles.
Forget about it, this will never work.
Lol on the image generation part. I’m on the “it can be real but very disappointing side” of the fence here. Certainly possible for them to be faking everything but that seems unlikely. We’ll see
> This is vaporware, it cannot work.
> When a solid body is immersed into water bubbles form on the surface. They are a problem.
There is a big difference between a problem, and a showstopper. There are all sorts of potential solutions to this. From chemicals that disrupt bubble formation, to underwater brushes or Jacuzzi jets to physically dislodge the bubbles to fancy algorithms that remove bubbles from the data.
You don’t work in ultrasound I take it. It always looks like there are “all sorts of potential solutions” in a domain you are not familiar with. For example, brushes make bubbles. Jets make bubbles. Do I need to go on?
But water bath ultrasound CT exists and is used all the time. How does SoftVue deal with bubbles?
They mention a “consumable” that provides an interface. A brief search didn’t find any technical details. So I don’t know.
But look at their device. This is no casual spa full-body scan. Dedicated bath for a small part of the body. Took years (decades, really) to make work, and still only as a supplemental to normal mammography. I stand by my assessment: Midjourney is full of it.
According to their pamphlet, the disposable Sequr Breast Interface only touches the end of the breast, by the nipple and uses gentle suction to "center, shape, and steady [the breast] during the image acquisition process". The majority of the breast is just in water.
Made a comment on a reply, but I had another thing I want to add here.
For one transcranial ultrasound is a thing that's already done and being pursued by different labs and research groups. I'm doing work specifically on transcranial ultrasound tomography, very similar to this at a national lab: https://pubmed.ncbi.nlm.nih.gov/40993881/
The thing I wanted to say that I haven't heard talked about enough, is the biggest advantage of this technology is that you can determine the sound speed of different artifacts, allowing much more quantitative analysis over the current qualitative "It looks like x", which is how it is done now.
This is not the first time a company has tried to bring ultrasound tomography to the market, there was a notable company who actually did bring a machine that would use ultrasound tomography to do the same job mammograms do, and no one bought it. For a couple reasons, mainly that most doctors feel that mammograms are the final form of that technology, but also just because that market is so saturated.
After doing some research and talking to colleagues more versed in the field. I do not think they are currently doing tomography as stated. I think they are doing compound B mode imaging. Tomography can penetrate and image through bone. Compound B mode imaging functionally cannot. I think this is a significant difference that they haven’t talked about yet. They are using butterfly Ultrasound-on-a-chip transducers which do not give you the raw rf data required to do the inversion to do tomography. They have built in beam forming and processing. I think it’s just helical B mode imaging. Which is backed up by the fact that the guy they hired to do this did his PhD in helical B mode imagining.
That difference really hampers the usability of this technology for me. Hopefully they will address this soon.
Additionally, this group published a paper doing this on a rat which they later redacted because it anatomically made no sense. Alluding to them feeding information into an AI which output garbage.
I’m curious to see where this goes.
Siemens Healthineers to take over production of Birds sneakers
Compared to the SoftVue ultrasound tomography scanner, the Midjourney scanner has several orders of magnitude more transducers/sensors, orders of magnitude more equivalent computational power (presumably using AI to greatly increase the efficiency), and they may be attempting full waveform inversion. If it works, it's potentially a much better image than your standard ultrasound, and could potentially see through bone. With the 360° scanning, they can potentially scan around air pockets even if they can't see inside the air pockets.
I want to make two broad points which tie into each other:
1. The over-diagnosis problem is rooted in trying to do medicine with population level data. What we really need is more individual tracking / continuous monitoring with high quality sensors of physiology.
2. The Midjourney tank is a step in the wrong direction. It's still a huge device, does a full body scan at low sensitivity, and they are aiming for 50k devices worldwide. What would help is almost the opposite -- an inexpensive and highly sensitive scanner that we can manufacture in the millions to billions, and hand to everyone for home use.
----
Anybody who's learned to ride a bicycle knows that over-steering is dangerous, and the surest way to take the system off-track. The correct method when controlling any complex system is always to acquire high-fidelity sensory input, at very low latency, but act on it sparingly.
If the average person can learn how to ride a bike, then I'm confident they can also internalize the same in the context of medicine; I propose the hurdle is that the epistemology of medicine is commonly misunderstood.
Contrary to the "Anna Karenina principle": every healthy human body is different, and well-regulated in its own way. It is not a machine that is supposed to be a perfect replica of some template human.
Here is where both technologists and doctors commonly go wrong: they implicitly adopt the latter perspective, and think that leveraging more data means comparing over the population to average out deviations. The natural consequence of this is to start "fixing" deviations, and this is the territory of dangerous over-steering. Perhaps because statistics attunes us to average over individual differences, any mention of data unfortunately makes people instinctively jump to this "cross-sectional" corner. Tests which are point-in-time and compare one human to others are problematic in this respect. This is especially acute when the body is measured along many many axes: like a full body scan.
The USAF study of pilot anatomy proves this point: no pilot is average in every attribute.
https://noblestatman.com/uploads/6/6/7/3/66731677/cockpit.flaw.averages.pdf
If we start from first principles and ask how to improve the quality of medicine, we end up in a very different corner of design space: the "longitudinal" corner.
Good doctors understand that biological systems operate on the opposite of the Anna Karenina principle. What they are actually doing, when helping a patient, is bringing all their knowledge and experience to bear on understanding the specific individual in front of them, and then steering them with the gentlest nudge necessary. The most useful data for this is actually high-res longitudinal tracking of this individual, so the doctor can help troubleshoot how *this physiology* might be going off homeostasis, and nudge it back. This is also why good doctors always focus on specific symptoms and troubleshoot very narrowly and in a hypothesis-driven manner; the goal is to fix the problem at hand and get back on track, rather than "solving" every deviation from population average.
Correspondingly, it is a huge mistake to think of the role of preventive scanning as comparing an individual's physiology to the human template. Every human is largely functional as they are, and there is likely zero immediate clinical signal in deviations from the population average. Doctors get up in arms about increased scanning because they think this is what they are supposed to look for (after all, textbooks explain how the typical human body behaves, and medical practice is always emphasizing population level-guidelines, and the statistics education we beat everyone with always emphasizes averaging).
So what is the role of preventive scanning and big data? Recall that you can't ride a bicycle by looking at the road immediately in front of you and reacting to every little bump. To steer well you've got to look farther ahead and react only to those perturbations which could compound to take you off-course if left unattended. The true role of preventive scanning is to have early warning of things that might upset homeostasis further down the line: high-resolution and low latency. And the role of big data is to help develop good models for how likely the *combination of observed deviations* is to upset homeostasis in the longer term i.e. to triage which deviations are likely benign and what situations need escalated attention… much like what a good doctor does. Given that doctors cannot attend to everyone everywhere all at once, we need all the diagnostic tracking possible to improve how we make this decision. Today this process is ad hoc and completely implicit, which means a large occurrence of both false positives and false negatives.
Therefore, the insights necessary to help any individual are very unlikely to come from a point-in-time observation; what is needed is an extended duration monitoring of the full physiology. Eg. "Yeah, we observe a few small blobs in that tissue over there, but as such they don't seem to be growing over the last year. But oh you are experiencing weight loss; hmmm. Have you additionally observed X/Y/Z? The possible cause could be related to this other indication on this other tissue, if A were causing B were causing C which then manifests in symptoms X, Y, Z." Notice how that sounds far more like a consultation with a good doctor, rather than the trope of let's find and fix (a la Procrustes) every deviation so you can fit the population average.
This clarifies that the most important axis of resolution for diagnostic information is time, and not population. What would help is highly-sensitive medical imaging that can track each the personal physiology of each human over extended periods.
The US Air Force concluded that designing the cockpit for the average pilot resulted in more accidents and fatalities, and their solution was to design a cockpit that every individual pilot could adapt to their anatomy.
----
Unfortunately the Midjourney tank is not enough of a step forward in the technology to enable the kind of imaging abundance we need. It's still a huge device, does a full body scan at low sensitivity, and they are aiming for 50k devices worldwide. That's not very different from the total number of MRI machines in the world today, and how many MRI scans has the typical person had?
We need an innovation in both the sensitivity and the form factor: an inexpensive scanner that will be more sensitive than the ultrasound technology which powers Midjourney, in a device which we can manufacture in the millions to billions, and hand everyone for ubiquitous use.
I believe we are coming up on the technology to make it possible, but it requires thinking in a different framework: https://woventhought.substack.com/p/visions-and-blindspots-the-midjourney
----
PS: If I were to venture into a corollary example of over-pathologization by fitting to the population average, I believe mental health is manifesting a crisis due to exactly this mistake. But that is a tangent to the discussion at hand, and Scott has a better understanding of the ground reality than I do, so that would be an interesting branch for another time
>Correspondingly, it is a huge mistake to think of the role of preventive scanning as comparing an individual's physiology to the human template. Every human is largely functional as they are, and there is likely zero immediate clinical signal in deviations from the population average.
This is absolutely not true. There are many measures of human anatomy and physiology for which deviance from average in one or both directions predicts shortened life. Heart size is one of many. So is walking speed.
Yes, the current system uses deviations from population average as a diagnostic signal, but have you considered that might not be the best thing to do? That's the point I'm making.
(edited for clarity)
>Yes, we do that right now
We do what right now?
I didn't say a word about anything we do, I pointed out about a pattern in the data regarding many measures of anatomy and physiology.
Hi: I need help from people who love science! Can data center cooling loops recycle electricity from condensation as MIT's groundbreaking 2013 experiments suggest?
As data centers face an unprecedented energy and cooling crisis driven by AI infrastructure, billions are being spent on managing thermal phase changes—specifically evaporation and condensation.
In my latest work at Decoding Science, I’ve outlined a simple backyard citizen-scientist experiment that suggests we fundamentally misunderstand electron behavior during these phase changes. Standard solid-state models predict a neutral net current during condensation on these surfaces, but geometric manipulation reveals a measurable electrical anomaly.
If this anomaly scales, it means the massive humidity and condensation loops already running inside data center cooling infrastructure could be harnessed to harvest electricity directly from waste heat. Perhaps equally significant is how understanding evaporation will facilitate the cooling process.
I’ve laid out the simple, reproducible experiment to prove this effect. Would you be interested in a brief look at how rethinking electron kinetics during condensation could impact the future of data center efficiency?
Best regards,
Thomas Alan White
Funny anecdote.
Did a pancreatic MRI last month (GLP1 caused enzymes to shoot up, looks harmless). And of course didn't want to wait 3 days for the results, so I took the CD, bought a $10 CD reader and loosed Claude Code on the data.
It ... worked? Installed a bunch of software to analyze it, did a bunch of things to try to compensate for not being a human, and in the end found about as much as the final report - which isn't as exciting as it sounds, because it rounded up to "nothing".
Note that I already have a project with all my clinical history and genetic data. As one does. So the value here is in correlations - you can have Fable level LLMs pour over all your data in the background, for what will soon cost $20 per month. Correlate genetics, lab results, MRIs, occasional Midjourney scans, your daily data (Oura, weight scale, Apple Watch) and latest research.
Currently this takes a certain personality to not obsess over health (or, in my case, to enjoy obsessing over health). But raw Opus is already pretty discerning. Fable with a proper harness can definitely do this in a practical manner.
Wow, AI or I guess possibly the Virgin Mary just hid my post. But I believe the intrepid can still see it by clicking on "hidden."
Possibly it ran afoul of an anti-spam content filter, given some of the words involved.
So will comments like this present one, which mention kink and porn without being either, be filtered out?
Apparently not. Could be the algorithm has some grammatical nuance, or "hot" was what put it over the threshold.
I'm not a physicist, just a medical student going into neurosurgery with a PhD in neuroimaging, so I don't know much about ultrasound, but we regularly do high-intensity focused ultrasound (HiFu) procedures, so at least some ultrasound waves obviously can penetrate bone in a highly targeted manner. Whether those waves can be used for imaging is an open question, but it's not an obviously wrong approach.
I think the skin example is good but for other reasons. Does the fact that we happen to be able to see what's on your skin make us somehow at a disadvantage versus if we couldn't see it? Of course not. It's not an information hazard being able to see people's moles. We adjust our protocols (maybe implicitly) to account for the noise.
I've always been baffled by the fact that in response to the fact that doctors make harmful interventions in response to some diagnostic tests -- eg prostrate screening at too young an age -- the medical establishment recommends the tests not be done.
Why not first just try recommending you don't act on certain results of the tests? I could imagine a different world where they tried that but patients were so pushy doctors felt they didn't have a choice. But that doesn't seem to be this world. I don't remember any long period where the recommendations pushed not treating lumps discovered during mammograms or the like.
If it's about how the patient reacts then it's kinda paternalistic since some people will be totally unbothered finding out that info. Is this downstream of legal liability?
---
Not that most of medicine makes sense to me. Still seems bonkers to me that it's a horrible ethical violation to give a patient even a vaccine without their consent but you can withdraw a medication they've been taking their entire adult lives and it's perfectly fine. Indeed it might be judged unethical to allow the patient to maintain their current state.
The scanner problem runs parallel to a much deeper issue: how AI doesn't just scan or filter history, but actively reformats it.
When testing how major bots (ChatGPT, Claude, Copilot, Grok) process the exact same historical data—like a closed murder case—you quickly realize we aren't facing simple 'hallucinations.' It’s a systematic rewrite. The dangerous part isn't the AI generating raw code or images; it's the digital software overriding physical hardware (digital history vs. biological consensus). If the scanners become the architects of what we perceive as 'proven truth,' digital censorship will stop looking like a sign of power and more like a total crisis of information control.
Question from a medicine-person for the physics-folks here:
Ultrasound probes currently are both the transmitter and also the receiver. The limitation of this is that it can only image objects that it has line of sight on. This presents one of the main limitations of transthoracic echocardiography (TTE): your beam has to dodge the ribs in order to get to the good stuff (the heart). This is also what makes transesophageal echo (TEE) so useful; it provides amazing pictures (no ribs in the way) at the somewhat inconvenient cost of needing to shove an ultrasound probe down the esophagus.
Could you have a transmitter and receiver that are separate? Aka, can you have a pitcher throw a sound wave that glances off an object and is caught by a receiver standing at a 90 degree angle away?
Could you have *multiple* transmitters and receivers that are all separate? The analogy for this in my head is: if an individual whale in a pod transmits a sound wave, are the other whales able to reconstruct a 3D model from that sound wave? If the whale pod was a hive mind, what level fidelity image could it get?
That's what's interesting to me about this Midjourney project; unlike traditional ultrasound it has multiple angles of transmitters and receivers. Can you take advantage of this, or is the math too difficult?
Bonus question: could you drop a TEE probe inside the body, and have them received by ultrasound receivers outside the body?
Yes this can be and is being done. The transducers are reciprocal so they can do both, and you can send a signal from one and receive with another or multiple.
Would this be useful for imaging muscles, tendons, ligaments, and fascia? If so, I could see it being used for PT and sports medicine.
I have been reading a lot of the vitriolic debates on X. I can't help but feel that both sides (tech optimists and medical doctors) have major valid points but are talking past each other due to differences in preferences and timelines.
The current state of coordination for collecting high quality large data sets with outcome data that is accessible for decision making for physicians and patient is severely lacking but I don't know how to fix this. I think the tech optimist argument is basically we just need to collect obscene amounts of data and get AI to sort it all out as quickly as possible.
>The Midjourney Scanner probably won’t help with [things DEXA is used for], because some bones are behind other bones or air
The bones most relevant for DEXA's biggest use-case (osteoporosis screening) are not hidden in those ways: tib/fib/femur, and also radius/ulna/humerus. Assuming you can shoot the ultrasound waves from any angle you want. Lumbar spine is sort of hidden, depends how good the scanner is. However the real advantage of DEXA is that it gives you bone *density*, which ultrasound cannot (AFAIK?). But the Midjourney scanner looks like it can get 3D bone geometry, which DEXA cannot (since it's a 2D projection). So, maybe an opportunity here?
I agree that the main on-ramp for revenue is probably longevity bros paying for better (or "better") reports on muscle mass, fat content, etc, though.
More interesting approach would be to improve already existing MRI diagnostic with software solution. I heard of an Swedish startup working on a simulation based methods that would also dramatically cut scanning time, and their results look promising.
Any good writeups of what (if anything) this pivot means for the AI economy? Early sign of the bubble deflating / bursting? Or not really?
Re: imaging through bone:
Looks like they're having some more success, this just came out. https://alephneuro.com/blog/ultrasound-brain
Big surprise: Dunning-Kruger in full effect.
"I can generate movies, therefore I'm a radiologist."
This post sounds like it makes the assumption that AI-based ultrasound is limited to detecting what human doctor can detect via ultrasound.
We are currently quite poor at detecting a lot of issues around lower back pain, and AI-assisted ultrasound can do a better job at telling which fascia planes move against each others and which don't.
If you treat medical diagnosis as a software problem instead of treating it as being about training human doctor assessment, I think there's a lot of room for getting better and it seems this MidJourney has a good chance to do that.
Hm, there are a couple of claims you are making that are just inaccurate, particularly in your MRI/Ultrasound comparisons.
(Qualifications: I did fMRI in the Gallant Lab in grad school, where we were able to decode viewed images from brain activity. More recently, I have been working on functional brain imaging with ultrasound, using the same chip architecture that Midjourney is using, from Butterfly)
1. Higher resolution -- not necessarily and it depends on what you're comparing. Ultrasound's resolution (for linear beamforming) is limited by the transmission wavelength, which could be as low as 100 microns. MRI is limited by the slew rate of magnetic gradients, i.e. so imaging at higher resolution takes a lot longer. Ultrasound set ups like this could enable imaging the body moving in real time in a way that ultrafast MRI will never be able to match at the same resolution.
2. Ultrasound can go through bone, and it actually penetrates quite well. The problem is things like the terbicular layer on the inside is the skull, where the microstructure is super nonuniform. This makes it so you can't do standard beamforming. But here's a key: this is fundamentally a computational tomography problem, not a physics problem. Approaches like FWI have demonstrated that you can even reconstruct gray matter through the skull. And it's hard to imagine a company better-positioned than Midjourney to tackle this shape of problem. (There are also amortized inference approaches to FWI that could make it a whole lot faster)
I think we're going to see ultrasound become faster and (a lot) cheaper than MRI, and it will be particularly good at imaging things that are moving. Because you can collect it so much more cheaply, we'll be able train diagnostic and segmentation models much more easily.
Heck, I might have talked myself into applying for a job there 🤣
> perfectly immune to patient pressure (because it’s an AI)
Um. AI sycophancy is solved then?