This is similar only in using an array of off-axis ultrasound tomography receivers but otherwise unrelated in that it's a serious publication with detailed information demonstrating potential medical utility. Near as I could tell, Midjourney Medical is an idea for a trendy spa treatment dressed up to look sci-fi cool. It's based on a repackaging of 40 of the exact same chip in Butterfly's handheld, full contact USB pocket-sized scanner which plugs into a mobile phone and is already available.
The CalTech team who wrote the Nature paper appears to be using an array of Olympus transducers equipped with their own custom lenses and a rotating emitter. Notably, the CalTech paper is focused on evaluating potential clinical benefits.
This is cool! I've got a personal project in the works to build an ultrasound machine just to learn how it works. There's a lot of electronics design, PCB design, FPGA work etc that in learning.
The probes themselves are used medical grade ones from ebay.
The hard part is that there can be over hundreds of piezo elements each needing to receive on their own channel. That means lots of expensive ADCs. Not to mention either having many pulser circuits or muxing a pulser to many elements.
I did a tech deep-dive into the Midjourney tank, and this is basically the origin of that (the first author David Garrett worked at Midjourney for a while) so I have some thoughts on the tech.
A ring of devices at 60cm (70cm for Midjourney) diameter means much longer distances than even the deepest tissue imaging ultrasound is typically used for. Geometric 1/r^2 attenuation through water and exponential attenuation through tissue. Sensitivity is always the key question for any sensor. This is why they use 1MHz ultrasound (lower frequencies attenuate less) rather than say 5-10Mhz, but that also means lower resolution. Sound wavelength in water/tissue at 1MHz is roughly 1.5mm. So in pursuit of sensitivity, good resolution becomes a challenge.
What sensors could we line up around the tank? This work has gone to great lengths to build custom transducers for receiving the ultrasound signal, but I couldn't easily find a direct characterization of sensitivity (pardon me if I missed it; I'd love to know more) but I don't think they are claiming a breakthrough in sensitivity -- only a new (tank) system architecture. AFAIK, all available technology, be it Butterfly/Olympus/whatever have a fundamental tradeoff in cost and sensitivity, because they all pick different operating points on the Pareto curve set by the underlying technology. Butterfly is not as sensitive as the top-grade devices, but the top-grade hospital ultrasound machines are way more expensive and bulky (roughly 150k USD & 150Kg). It would be a real challenge to fit dozens of them around the tank! Let alone pay for them, if one wants to deploy at scale.
If cost and bulk were not a constraint, we could definitely use many hospital-grade sensors and get high-fidelity images. The problem is what happens if the sensitivity is not quite enough. Look at the online discourse from doctors about how low-sensitive full-body scans are asking for trouble from over-diagnosis (incidentalomas) and iatrogenic side-effects. Low sensitivity full-body scan is basically the diagnostic version of p-hacking. Just like the US air force found that there is no "fully average" pilot, every body scanned will always have some anomaly or the other. And most of them never need to be clinically intervened on. That's not quite the fault of the device, but that's still a massive gap in the deployment strategy and it will require retraining doctors and restructuring large parts of the existing healthcare system.
It's easier to just invent technology and build higher-fidelity sensors. This device is mostly just an assemblage of available sensors, we need innovation that produces sensors which are smaller + inexpensive + less power hungry. It's not just a question of assembling it into different form factors or scaling production, or nudging people into using these by frequenting spas. AFAICT that's barely getting any discussion in the chatter around the Midjourney launch.
Lastly, it's really important that this doesn't require a trained sonographer. That's a real scaling bottleneck if we want to enable exponentially more usage of scanning. (Not taking a dig at the paper authors, but only at the online discourse) If we really wanted to scale this to population level use... the tank form-factor is good, but guess what's easier: having smaller devices one could just drop into a bathtub at home :-)
You need a full ring of sensors to be receiving at all times. The Caltech design has a rotating element that emits once at each location on the ring. The Midjourney design instead uses the same receiver elements to also emit one at a time, in turns. But wherever the emitter fires from, all the receivers need to be listening.
If you move a single sensor around then you need to multiply the scanning time by the same amount, and repeat the same "experiment" for each measurement position.
Even if it's not the same resolution there's still benefits, a bit one that it will likely be WAY less expensive, both upfront and in maintenance, even after building the water tank. It might also replace a fair amount of routine ultrasound usage because of the reduction of operator skill.
Also, there's plenty of people who can't be in MRI machines, and likely a lot safer for everyone else.
MRI is done using X rays which have very very small wavelength compared to sound. And resolution of images i.e. less blurry varies inversely proportional to the wavelength. Yes some algo can make slightly better but I am certain that nothing can beat x rays in clarity for humans
MRI uses RF wavelengths, which are measured in meters. Ultrasound waves are measured in fractions of a millimeter. Wavelength is not the relevant factor here.
21 comments:
Be interested to know if this is related to the same process/technology that Midjourney announced last week[0]
0. https://www.midjourney.com/medical/blogpost
Yes. IIRC David Garrett (the first author of the Nature paper) also worked at Midjourney in the intervening period.
Unpaywalled link to the paper: https://www.semanticscholar.org/reader/ef7ae3bff634710f87124...
This is similar only in using an array of off-axis ultrasound tomography receivers but otherwise unrelated in that it's a serious publication with detailed information demonstrating potential medical utility. Near as I could tell, Midjourney Medical is an idea for a trendy spa treatment dressed up to look sci-fi cool. It's based on a repackaging of 40 of the exact same chip in Butterfly's handheld, full contact USB pocket-sized scanner which plugs into a mobile phone and is already available.
The CalTech team who wrote the Nature paper appears to be using an array of Olympus transducers equipped with their own custom lenses and a rotating emitter. Notably, the CalTech paper is focused on evaluating potential clinical benefits.
This is cool! I've got a personal project in the works to build an ultrasound machine just to learn how it works. There's a lot of electronics design, PCB design, FPGA work etc that in learning. The probes themselves are used medical grade ones from ebay.
The hard part is that there can be over hundreds of piezo elements each needing to receive on their own channel. That means lots of expensive ADCs. Not to mention either having many pulser circuits or muxing a pulser to many elements.
Have a look at this open source ultrasound project for some inspiration! https://un0rick.cc/
Also my personal project which was inspired by un0rick but a bit simpler: https://williammeng.com/rtl-ultrasound.html
Relevant article:
Scanning the Body with Sound
https://www.caltech.edu/about/news/scanning-the-body-with-so...
Pictures are surely what everyone wants to see, thanks. Animation that looks similar below.
I’m not sure its that similar to MRI, but there are clear advantages to using US and it’s a new technique which hopefully get better.
https://medicalxpress.com/news/2026-04-body-ultrasound-captu...
Heck yeah! Thanks for posting.
I did a tech deep-dive into the Midjourney tank, and this is basically the origin of that (the first author David Garrett worked at Midjourney for a while) so I have some thoughts on the tech.
A ring of devices at 60cm (70cm for Midjourney) diameter means much longer distances than even the deepest tissue imaging ultrasound is typically used for. Geometric 1/r^2 attenuation through water and exponential attenuation through tissue. Sensitivity is always the key question for any sensor. This is why they use 1MHz ultrasound (lower frequencies attenuate less) rather than say 5-10Mhz, but that also means lower resolution. Sound wavelength in water/tissue at 1MHz is roughly 1.5mm. So in pursuit of sensitivity, good resolution becomes a challenge.
What sensors could we line up around the tank? This work has gone to great lengths to build custom transducers for receiving the ultrasound signal, but I couldn't easily find a direct characterization of sensitivity (pardon me if I missed it; I'd love to know more) but I don't think they are claiming a breakthrough in sensitivity -- only a new (tank) system architecture. AFAIK, all available technology, be it Butterfly/Olympus/whatever have a fundamental tradeoff in cost and sensitivity, because they all pick different operating points on the Pareto curve set by the underlying technology. Butterfly is not as sensitive as the top-grade devices, but the top-grade hospital ultrasound machines are way more expensive and bulky (roughly 150k USD & 150Kg). It would be a real challenge to fit dozens of them around the tank! Let alone pay for them, if one wants to deploy at scale.
If cost and bulk were not a constraint, we could definitely use many hospital-grade sensors and get high-fidelity images. The problem is what happens if the sensitivity is not quite enough. Look at the online discourse from doctors about how low-sensitive full-body scans are asking for trouble from over-diagnosis (incidentalomas) and iatrogenic side-effects. Low sensitivity full-body scan is basically the diagnostic version of p-hacking. Just like the US air force found that there is no "fully average" pilot, every body scanned will always have some anomaly or the other. And most of them never need to be clinically intervened on. That's not quite the fault of the device, but that's still a massive gap in the deployment strategy and it will require retraining doctors and restructuring large parts of the existing healthcare system.
It's easier to just invent technology and build higher-fidelity sensors. This device is mostly just an assemblage of available sensors, we need innovation that produces sensors which are smaller + inexpensive + less power hungry. It's not just a question of assembling it into different form factors or scaling production, or nudging people into using these by frequenting spas. AFAICT that's barely getting any discussion in the chatter around the Midjourney launch.
Lastly, it's really important that this doesn't require a trained sonographer. That's a real scaling bottleneck if we want to enable exponentially more usage of scanning. (Not taking a dig at the paper authors, but only at the online discourse) If we really wanted to scale this to population level use... the tank form-factor is good, but guess what's easier: having smaller devices one could just drop into a bathtub at home :-)
--
PS: I also wrote more broadly about the Midjourney scanner, looking at both the tech and the road to deployment: https://woventhought.substack.com/p/visions-and-blindspots-t...
Do you need to fit that many sensors around the edge or can you rotate the ring of fewer sensors?
You need a full ring of sensors to be receiving at all times. The Caltech design has a rotating element that emits once at each location on the ring. The Midjourney design instead uses the same receiver elements to also emit one at a time, in turns. But wherever the emitter fires from, all the receivers need to be listening.
If you move a single sensor around then you need to multiply the scanning time by the same amount, and repeat the same "experiment" for each measurement position.
I really wanted to make a career in tomography. But had only one elective from another department and didn't know how to make it into this stream.
It’s difficult. Medical devices are not easy for a small business to break into and the big manufacturers don’t typically hire entry-level.
Everyone I know in the field has come into it laterally either from a prior software or hardware job in an unrelated field.
Tomography: Come today, join tomorrow.
You still can! It looks like it is an amazing mix of physics, math, engineering, acoustics, biology, basically everything.
The results really look rather blurry compared to the MRI...
Presumably that's just a matter of better system modelling to fix?
Even if it's not the same resolution there's still benefits, a bit one that it will likely be WAY less expensive, both upfront and in maintenance, even after building the water tank. It might also replace a fair amount of routine ultrasound usage because of the reduction of operator skill.
Also, there's plenty of people who can't be in MRI machines, and likely a lot safer for everyone else.
MRI is done using X rays which have very very small wavelength compared to sound. And resolution of images i.e. less blurry varies inversely proportional to the wavelength. Yes some algo can make slightly better but I am certain that nothing can beat x rays in clarity for humans
MRI uses RF wavelengths, which are measured in meters. Ultrasound waves are measured in fractions of a millimeter. Wavelength is not the relevant factor here.
You are thinking of CT scans. MRIs use strong magnetic fields and MHz-range radio waves.
No X rays are involved in MRI.