The Midjourney Scanner
Scott mentioned my take, so I better elaborate
For the two people who read me but not Scott Alexander, he recently mentioned my take on X.com about the Midjourney Scanner. So I figured I’d put my opinions in one place. In one sentence, I think it’s very likely the final product will be a good product for a health spa, possibly giving information on body composition, and that’ll give it enough legs to see if it can replace an MRI.
The Scanner
For anyone who doesn’t know the AI company Midjourney, which specializes in making great AI art, recently announced they’re pivoting into health spas, because why not?12 The spas will have all the things you’re used to in such places, plus a full body ultrasound scanner. The scanner is what everyone is talking about. So what is it? Basically you get dunked in a pool of water up to your neck and bombarded by sound waves from all directions. The whole process will take about a minute, plus time to change and dry off. It’s faster than most medical scanning technology I’ve experienced, so it has that going for it.
How Does it Work
Check out their page on the subject for all the information anyone has on this technology. Lots of people are claiming more info, but there’s almost no other info out there.
How Useful is it?
NO ONE KNOWS! This is the part that social media has trouble understanding. The machine isn’t built yet. According to the CEO they have only scanned around 12 people with their prototypes so far, and they haven’t released the full scans of those people. We also don’t have the AI analysis of those images. And we haven’t compared the information we don’t yet have, with established scans of the same people to see how good they are. So no one knows, not even Midjourney.
Will it Replace MRIs?
I basically answered this in the previous question, we don’t know but probably not. There’s a bunch of theoretical reasons why it can’t do exactly what an MRI does, but clever engineering and analysis has gotten around similar problems in the past. It seems like something that’s impossible until it happens, but that doesn’t mean this scanner will work as promised.
All that said, I’m rooting for them to succeed.
If it Can’t Replace MRIs will it be useful at all?
Yes. Wellness treatments and sports medicine often scan people to determine body composition. What is your body fat percentage? Lots of people have scales in their homes that give them a BF% reading3 and lots of people go to fancy gyms, clinics, or mobile DEXA scan vans to get a measurement from a DEXA scan. DEXAScans are popular, fairly well validated in the scientific literature, and almost too noisy to be useful for individuals. It’s very possible that the Midjourney scanner will be able to be more precise than a DEXA scan.
What good are better DEXA Scans?
Getting a precise body fat measurement weekly would actually be game changing for a lot of people. It’s not a super important medical need, but for anyone working on weight loss it changes the game. Most people who want to lose weight want to lose body fat, but it’s easy to lose weight by losing muscle, losing retained water, losing glycogen4, or just changing the amount of food in your digestive tract. What I’m saying is scale measurements are noisy.
The best tool for this is a DEXA scan. The problem is they cost $50-200 per scan and can be off by as much as 5 percentage points in either direction. This makes them the gold standard for experiments where you’re measuring groups of people5 but of only limited value for individuals.
If you could cheaply and frequently get a precise number related to your “weight” loss progress6 every week it would be much easier to know that everything is working or if your plan isn’t working. This means you can adjust your plan as needed much faster and have the peace of mind that things are on track, which is a big blocker for many people.
Also, if you’re very overfat and lifting weights while trying to lose weight, you can gain muscle while losing fat. It is very discouraging to be over 300lbs, losing 1-2lbs of fat per week but putting on .75-1.5lbs of muscle per week. The scale barely moves and it can be hard to see progress until the total amount of fat tissue gets low enough, but you’re transforming your body every day. An objective measure that this is working would be invaluable.
So What Now
All data collection tools are useless until validated. If all this can do is provide body fat percentage for personal use with a 1% margin of error7, that’ll be enough to make it a market success. That should let them get enough scans, and enough repeat scans to validate any other uses the machine is capable of, plus iterate on the device and software side of things.
Body fat readings are popular enough to be growing businesses in many urban centers and the current tech is bad enough that I’m optimistic that something like Midjourney can do better.
Isn’t getting a full body scan every few weeks or months bad?
Lots of people on social media are talking about the unnecessary care that often arises after people get optional MRIs. I don’t think that’s going to happen. These things are launching in a health spa. What does that mean? It means it’s basically a Fitbit.
In my experience doctors mostly roll their eyes at any result from consumer medical products like smart watches or Fitbits, even if it’s literally the same make and model they use in their office to measure your Blood O2. Even a credulous doctor will want to order their own tests and make their own evaluation instead of trusting the local health spa.
I expect most doctors will basically ignore the results, the same way they ignore my Fitbit results. Wellness enthusiasts, gym bros, and Bryan Johnson will get lots of scans and treat themselves. Hopefully after several months/years there will be enough data to push for official medical use for a similar device. At that point there will be a different standard of care, maybe something involving a decreased sensitivity to anomalies on a single scan, but attention paid to persistent or worsening anomalies. Who knows? We’ll have months of data by then to guide us to something smart, we don’t have to figure it out a priori.
We’ve been getting lots of consumer medical-like data in our hands over the past 10-15 years, the medical system has managed to avoid overreacting to it8 . I expect they’ll continue to figure out how to ignore Midjourney Medical until David Holz forces them to pay attention with overwhelming data.
Takeaways
In an era where half of Twitter is complaining about tech bro money, the Midjourney guys are spending their own money on building a new medical device. They hope to make a cheap and easy MRI machine, and everyone should be hoping they succeed. But even if they fail there’s a lot of useful work to be done like cheap and reliable body composition measurement. Making a machine that measures body composition within 1% of the true value9 would make the world a better place, and fund further R&D to make the scan even better.
And who knows, it might have other uses beyond finding tumors or bodyfat percentage, the world is big.
It’s not a bubble!
Really, it’s not. They’re doing this with profits, not investor dollars.
They don’t work
it’s not exactly the same as water weight
The errors cancel out
Actually fat loss progress
plus a ton of speculative data that boosters will be excited about but you shouldn’t take seriously at first
Health influencers on the other hand….
This is just a very hopeful guess at what this new tech might be able to do. There’s no real evidence on this question.
When Andrew says re. home body composition scales, "They don't work", that needs some unpacking.
There are 3 types of consumer body composition scale:
1. Ones under $100, which say they measure body composition, but do no /measuring/ of it at all. They have you tell them your weight and a long list of body measurements you make with measuring tape, and then use an equation to /predict/ your percent body fat from that. Sometimes they even have 4 metal pads to pretend they do bioelectrical impedance analysis (BIA), but they don't. One way to know they don't, is to buy one, and use it for a long time, and notice that they always give the same body fat number, to 3 or 4 significant digits, for the same weight, unless you change your body measurements. Another is to weigh yourself, pick up something weighing about 3 ounces, and weigh yourself again. If the scale first shows your previous measurement plus 3 oz, then immediately switches to your previous measurement, it is trying to fool you into thinking it's more-precise than it is with software that says, "If a measurement is within a few ounces of the previous measurement, show the previous measurement instead of the current measurement, because the user is just stepping off the scale and on it again to see how consistent your measurements are." In my limited experience (3 scales which all do both), scales that use one of these dishonest tricks also use the other.
A dishonest scale may be very accurate! I have a Fitindex scale which cost $30 and has a standard deviation of about 0.1 lbs, yet nonetheless cheats to pretend to be accurate within... 0.1 lbs. (Measurements are to the nearest 0.2 lbs.)
2. Ones over $100 which use BIA, but have no handgrips, so the BIA is from one leg to the other. Claude Sonnet 4.6 tells me the Withings Body Smart ($130) and Body Comp ($230) are legit. Because they only have one path to send electricity through, it isn't reasonable to expect them to distinguish subcutaneous fat from visceral fat, though they will claim to do so.
3. Ones over $300 which use BIA and have handgrips. These can be accurate at measuring body fat and visceral fat. These include the InBody H30 ($380), the InBody H40 ($500), and the Withings Body Scan ($500). These 3 are all attested to be accurate by users who've compared them to DEXA scans. The H40 seems to be the best-validated for body composition, and has a big advantage over the H30 in that its software can be updated. I say "big advantage" because these scales report results by Bluetooth, and I expect the Bluetooth protocol will need to be rewritten within a year due to Claude Mythos, and so the H30 will be bricked other than the limited info they can provide on their built-in screen.