Demand for human radiologists is at an all-time high
worksinprogress.news
worksinprogress.news
Can AI read diagnostic images better than a radiologist? Almost certainly the answer is (or will be) yes.
Will radiologists be replaced? Almost certainly the answer is no.
Why not? Medical risk. Unless the law changes, a radiologist will have to sign off on each imaging report. So say you have an AI that reads images primarily and writes pristine reports. The bottleneck will still be the time it takes for the radiologist to look at the images and validate the automated report. Today, radiologist read very quickly, with a private practice rads averaging maybe 60-100 studies per day (XRs, ultrasounds, MRIs, CTs, nuclear medicine studies, mammograms, etc). This is near the limit of what a human being can reasonably do. Yes, there will be slight gains at not having to dictate anything, but still having to validate everything takes nearly as much time.
Now, I'm sure there's a cavalier radiologist out htere who would just click "sign, sign, sign..." but you know there's a malpractice attorney just waiting for that lawsuit.
Which is literally the case so far. No manufacturer has shown any willingness to take on the liability of self driving at any scale to date. Waymo has what? 700 cars on the road with the finances and lawyers of Google backing it.
Let me know when the bean counters sign off on fleets in the millions of vehicles.
They have over 2000 on the road and are growing: https://techcrunch.com/2025/08/31/techcrunch-mobility-a-new-...
Of course there's 200M+ personal vehicles registered in the US.
I have to admit if my life were on the line I might be that Karen.
False negatives are far more problematic.
It's the same way that we can save time and money if we just don't wash our hands when cooking food. Sure it's true. But someone WILL get sick and we WILL get in trouble for it
theres an MBA salivating over that presntation
Planes can land themselves with zero human intervention in all kinds of weather conditions and operating environments. In fact, there was a documentary where the plane landed so precisely that you could hear the tires hitting the center lane marker as it landed and then taxied.
Yet we STILL have pilots as a "last line of defense" in case something goes wrong.
I'm not fully up to speed on the Autonomi / Garmin Autoland implementation found today on Cirrus and other aircraft -- but it's not for "everyday" use for landings.
[0] https://pilotinstitute.com/can-an-airplane-land-itself/
[1] https://askthepilot.com/questionanswers/automation-myths/
Medicine is existential. The job of a doctor is not to look at data, give a diagnosis and leave. A crucial function of practicing doctors is communication and human interaction with their patients.
When your life is on the line (and frankly, even if it isn't), you do not want to talk to an LLM. At minimum you expect that another human can explain to you what is wrong with you and what options there are for you.
What? If I don't trust the machine or the software running it, absolutely I do, if I have to share the road with that car, as its mistakes are quite capable of killing me.
(Yes, I can die in other accidents too. But saying "there's no reason for me to care if the cars around me are filled with people sleeping while FSD tries to solve driving" is not accurate.)
On a basic level, software exists to expedite repetitive human tasks. Diagnostic radiology is an extremely repetitive human task. When I read diagnostics, there’s a voice in the back of my head saying, “I should be writing code to automate this rather than dictating it myself.”
That's it?
I don't know. Doesn't sound like a very big obstacle to me. But I don't think AI will replace radiologists even if there was a law that said like, "blah blah blah, automated reports, can't be sued, blah blah." I personally think the consulting work they do is really valuable and very difficult to automate, we would be in an AGI world where radiologists get replaced, which seems unlikely.
The bigger picture is that we are pretty much obligated to treat people medically, which is a good thing, so there is a lot more interest in automating healthcare than say, law, where spending isn't really compulsory.
A lot of things are one law amendment away from happening and they aren’t happening. This could well become another mask mandate, which while being reasonable in itself, rubs people wrong way just enough to become a sacred issue.
Radiologists will validate the results but either find themselves clicking "approve, approve, approve" all day, or disagree and find they were wrong (since our hypothesis is that the AI is better than a human). Eventually, this will be common knowledge in the field, hospitals will decide to save on costs and just skip the humans altogether, lobby, and get the law changed.
And it’s definitely not a 0.05 percent difference. AI will perform better by a long shot.
Two reasons for this.
1. The AI is trained on better data. If the radiologist makes a mistake that mistake is identified later and then the training data can be flagged.
2. No human indeterminism. AI doesn’t get stressed or tired. This alone even without 1. above will make AI beat humans.
Let’s say 1. was applied but that only applies for consistent mistakes that humans make. Consistent mistakes are eventually flagged and shows up as a pattern in training data and the AI can learn it even though humans themselves never actually notice the pattern. Humans just know that the radiologists opinion was wrong because a different outcome happened, we don’t even have to know why it was wrong and many times we can’t know… just flagging the data is enough for the AI to ingest the pattern.
Inconsistent mistakes comes from number 2. If humans make mistakes that are due to stress the training data reflecting those mistakes will be minuscule in size and also random without pattern. The average majority case of the training data will smooth these issues out and the model will remain consistent. Right? A marker that follows a certain pattern shows up 60 times in the data but one time it’s marked incorrectly because of human error… this will be smoothed out.
Overall it will be a statistical anomaly that defies intuition. Similar to how flying in planes is safer than driving. ML models in radiology and spam will beat humans.
I think we are under this delusion that all humans are better than ML but this is simply not true. You can thank LLMs for spreading this wrong intuition.
You can take a 4k photo of anything, change one pixel to pure white and a human wouldn't be able to find this pixel by looking at the picture with their eyes. A machine on the other hand would be able to do it immediately and effortlessly.
Machine vision is literally superhuman, For example Military camo can easily fool human eyes. But a machine can see through it clear as day. Because they can tell the difference between
Black Hex #000000 RGB 0, 0, 0 CMYK 0, 0, 0, 100
and
Jet Black Hex #343434 RGB 52, 52, 52 CMYK 0, 0, 0, 80
> Can AI read diagnostic images better than a radiologist? Almost certainly the answer is (or will be) yes.
I'm sorry, but I disagree, and I think you are making a wild assumption here. I am up to date on the latest AI products in radiology, use several of the, and none of them are even in the ballpark on this. That vast majority are non-contributory.
It is my strong belief that there is an almost infinite variation in both human anatomy and pathology. Given this variation, I believe that in order for your above assumption to be correct, the development of "AGI" will need to happen.
When I interpret a study I am not just matching patterns of pixels on the screen with my memory. I am thinking, puzzling, gathering and synthesizing new information. Every day I see something I have never seen before, and maybe no one has ever seen before. Things that can't and don't exist in a training data set.
I'm on the back end of my career now and I am financially secure. I mention that because people will assume I'm a greedy and ignorant Luddite doctor trying to protect my way of life. On the contrary, if someone developed a good replacement for what I don, I would gladly lay down my microphone and move on.
But I don't think we are there yet, in fact I don't think we're even close.
I can easily imagine that humans are better at really digging deeply and reasoning carefully about anomalies that they notice.
I doubt they're nearly as good as computers at detecting subtle changes on screens where 99% of images have nothing worrisome and the priors are "nothing is suspicious".
I don't want to equate radiologists with TSA screeners, but the false negative rate for TSA screening of carryon bags is incredibly high. I think there's an analog here about the ability of humans to maintain sustained focus on tedious tasks.
Seems like an over simplification, but let's say it's just true. Wouldn't you rather spend your time on novel problems that you haven't seen before? Some ML system identifies easy/common ones that it has high confidence in, leaving the interesting ones for you?
Think about how you learned anatomy. You probably looked at Netter drawings or Grey's long before you ever saw a CT or MRI. You probably knew the English word "laceration" before you saw a liver lac. You probably knew what a ground glass bathroom window looked like before the term was used to describe lung findings.
LLMs/LVMs ingest a huge amount of training data, more than humans can appreciate, and learn connections between that data. I can ask these models to render an elephant in outer space with a hematoma on its snout in the style of a CT scan. Surely, there is no such image in the training set, yet the model knows what I want from the enormous number of associations in its network.
Also, the word "finite" has a very specific definition in mathematics. It's a natural human fallacy to equate very large with infinite. And the variation in images is finite. Given a 16-bit, 512 x 512 x 100 slice CT scan, you're looking at 2^16 * 26214400 possible images. Very large, but still finite.
Of course, the reality is way, way smaller. As a human, you can't even look at the entire grayscale spectrum. We just say, < -500 Hounsfield units (HU), that's air, -200 < fat < 0, bone/metal > 100, etc. A gifted radiologist can maybe distinguish 100 different tissue types based on the HU. So, instead of 2^16 pixel values, you have...100. That's 100 * 26214400 = 262,440,000 possible CT scans. That's a realistic upper-limit on how many different CT scans there could possibly be. So, let's pre-draft 260 million reports and just pick the one that fits best at inference time. The amount you'd have to change would be miniscule.
That being said, there are no radiologists available to hire at any price: https://x.com/ScottTruhlar/status/1951370887577706915
What was left out was that these "cutting edge" AI imaging models were old school CNNs from the mid 2010's, running on local computers. It seems only right now is the idea of using transformers (what LLMs are) is being explored.
In that sense, we still do not know what a purpose build "ChatGPT of radiology" would be capable of, but if we use the data point of comparing AI from 2015 to AI of 2025, the step up in ability is enormous.
And in AI tech, even "5 years ago" is a different era.
In year 2025, we have those massive multimodal reasoning LLMs that can crossreference data from different images, text and more. If the kind of effort and expertise that went into general purpose GPT-5 went into a more specialized medical AI, where would its capabilities top out?
Do you have any typical examples of this you could try to explain to us laymen, so we get a feel for what this looks like? I feel like it's hard for laymen to imagine how you could be seeing new things outside a pattern every day (or week}.
I am nowhere near as good as our worst radiologist (who is, frankly... not great). It's not even close.
AI will probably never taking over, what we really need is AI working in tandem with radiologist and complementing their work to help with their busy schedule (or limited number of radiologist).
The OP title can also be changed to "Demand for human cardiologist is at an all-time high", and is still be true.
For example in CVDs detection cardiologist need to diagnose the patient properly, and if the patient not happy with the diagnostic he can get a second opinion from another cardiologist, but cardiologist number is very limited even more limited than radiologist.
For most of the countries in the world, only several hundreds to several thousands registered cardiologist per country, making the ratio about 1:100,000 cardiologist to population ratio.
People expecting cardiologist to go through their ECG readings but do you know that reading ECG is very cumbersome. Let's say you have 5 minutes ECG signals for the minimum requirement for AFib detection as per guideline. The standard ECG is 12-lead resulting in 12 x 5 x 60 = 3600 beats even for the minimum 5 minutes durations requirements (assuming 1 minute ECG equals to 60 beats). Then of course we have Holter ECG with typical 24-hour readings that increase the duration considerably and that's why almost all Holter reading now is automated. But current ECG automated detection has very low accuracy because their accuracy of their detection methods (statistics/AI/ML) are bounded by the beat detection algorithm for example the venerable Pan-Tompkins for the fiducial time-domain approach [1].
The cardiologist will rather spent their time for more interesting activities like teaching future cardiologists, performing expensive procedures like ICD or pacemaker, or having their once in a blue moon holidays instead of reading monotonous patients' ECGs.
I think this is why ECG reading automation with AI/ML is necessary to complement the cardiologist but the trick is to increase the sensitivity part of the accuracy to very high value preferably 100% so the missing potential patients is minimized for the expert and cardiologist in the loop exercise.
[1] Pan–Tompkins algorithm:
https://en.wikipedia.org/wiki/Pan%E2%80%93Tompkins_algorithm
Paul Kedrosky had an interesting analogy when the automobile entered the scene. Teamsters (the men who drove teams of horses) benefited from rising salaries, even as new people declined to enter the "dead end" profession. We may well be seeing a similar phenomenon with Radiologists.
Finally, I'd like to point out that rising salaries mean there are greater incentives to find alternative solutions to this rising cost. Given the erratic political situation, I will not be surprised to see a relatively sudden transition to AI interpretation for at least a minority of cases.
I feel that human processes have inertia and for lack of a better word, gatekeepers feel that new, novel approaches should be adopted slowly and which is why we are not seeing the impact, yet. Once a country with the right incentive structure (e.g. China ) can show that it can outperform and help improve the overall experience I am sure things will change.
While 10 years progress is a lot in ML, AI , in more traditional fields it probably is a blip to change this institutional inertia which will change generation by generation. All that is needed is an external actor to take the risk and show a step change improvement. Having experienced how healthcare in US I feel people are only scared to take on bold challenges
From the article
My late wife had to have a stent placed in a vein in her brain to relieve cranial pressure. We had to travel to to New York for an interventional radiologist and team to fish a 7 inch stent and balloon from her thigh up.
At the time, we had to travel to NYC, and the doctor was one of a half dozen who could do the procedure in the US. Who’s going to train the future physician the skills needed to develop the procedure?
For stuff like this, I feel like AI is potentially going to erase certain human knowledge.
But if everyone involved has a profit motive, you end up cutting at those cost reductions. "We'll save you 100 bucks, so give us 50", done at the AI model level, the AI model repackager, the software suite that the hospital is using, the system integrators that manage the software suite installation for the hospital, the reseller of the integrator's services through some consultancy firm, etc etc.
There are so many layers involved, and each layer is so used to trying to take a slice, and we're talking about a good level of individualization in places that aren't fully public a la NHS, that the "vultures" (so to speak) are all there ready to take their cut.
Maybe anathema to say on this site, but de-agglomeration really seems to have killed just trying to make things better for the love of the game.
Just as one example a chest CT would’ve cost $450 if done cash. It costed an insurer over $1200 done via insurance. And that was after multiple appeals and reviews involving time from people at the insurance company and the providers office including the doctor himself. The low hanging fruit in American healthcare costs is the stuff like that.
> All that is needed is an external actor to take the risk and show a step change improvement
Who's going to benefit? Doctors might prioritize the security of their livelihood over access to care. Capital will certainly prioritize the bottom line over life and death[0].
The cynical take is that for the time being, doctors will hold back progress, until capital finds a way to pay them off. Then capital will control AI and control diagnosis, letting them decide who is sick and what kind of care they need.
The optimistic take is that doctors maintain control but embrace AI and use it to increase the standard of care, but like you point out, the pace of that might be generational instead of keeping pace with technological progress.
[0] https://www.nbcnews.com/news/us-news/death-rates-rose-hospit...
But it doesn't lead to increased throughput because there needs to be human validation when people's lives are on the line.
Planes fly themselves these days, it doesn't increase the "throughout" or eliminate the need for a qualified pilot (and even a copilot!)
That's more than a problem of inertia
That's true for AI-slop-in-the-media (most of the internet was already lowest effort garbage, which just got that tiny bit cheaper) and probably also in medicine (a slight increase in false negatives will be much, much more expensive than speeding up doctors by 50% for image interpretation). Once you get to the point where some other doctor is willing (and able) to take on the responsibility of that radiologist, then you can eliminate that kind of doctor (but still not her work. Just the additional human-human communication)
If statistically their error rate is better or around what a human does then their insurance is a factor of how many radiologists they intend to replace.
The system is designed a nanny-state fashion: there's no way to release practitioners from liability in exchange for less expensive treatments. I doubt this will change until healthcare pricing hits an extremely expensive breaking point.
You're probably right.
But we don't ask a second opinion to an "algorithm", we want a person, in front of us, telling us what is going on.
AI is and will be used in the foreseeable future as a tool by radiologists, but radiologists, at least for some more years, will keep their jobs.
When an X-Ray is ordered, there is usually a suspected diagnosis in the order like "suspect sprain, pls exclude fracture", "suspect lung cancer". Patients will complain about symptoms or give the impression of a certain illness. Things like that already place a bias on the evaluation a radiologist does, but they are trained to look past that and be objective. No idea how often they succeed.
[1] I don’t know the US system so it’s just a guess
4 years med school
2 years computer science
6 years of residency (intern year, 4 years of DR, 1 year of IR)
16 years...
4 years undergrad - major and minor not important, met the pre-med requirements 2 year grad school (got a master's degree, not required, but I was having fun) 4 years medical school 5 years radiology residency
Famous last words.
I think we all have become hyper-optimistic on technology. We want this tech to work and we want it to change the world in some fundamental way, but either things are moving very slowly or not at all.
I always wonder if honking at a Waymo does anything. A Waymo stopped for a (very slow) pickup on a very busy one lane street near me, and it could have pulled out of traffic if it had gone about 100 feet further. The 50-ish year old lady behind it laid on her horn for about 30 seconds. Surreal experience, and I'm still not sure if her honking made a difference.
I like Waymos though. Uber is in trouble.
Just to clarify, have you ridden in a Waymo? It didn't seem entirely clear if you just experienced living with Waymo or have ridden in it.
I tried it a few times in LA. What an amazing magical experience. I do agree with most of your assertions. It is just a super careful driver but it does not have the full common sense that a driver in a hectic city like LA has. Sometimes you gotta be more 'human' and that means having the intuition to discard the rules in the heat of the moment (ex. being conscious of how cyclists think instead of just blindly following the rules carefully, this is cultural and computers dont do 'culture').
As do most of the ridesharing drivers I interact with nowadays, sadly.
The difference is that Waymo has a trajectory that is getting better while human rideshare drivers have a trajectory that is getting worse.
I spend a lot of time as a pedestrian in Austin, and they are far safer than your usual Austin driver, and they also follow the law more often.
I always accept them when I call an Uber as well, and it's been a similar experience as a passenger.
I kinda hate what the Tesla stuff has done, because it makes it easier to dismiss those who are moving more slowly and focusing on safety and trust.
However, like railroad safety is expensive heavily regulated, self driving car companies have the same issue.
Decentralized driving decentralizes risk.
so when I have my _own_ robot to do it, it'll be easy and cheap.
Is there any saying that exists about overestimating stuff in the near term and long term but underestimating stuff in the midterm? Ie flying car dreams in the 50s etc.
We still don't have flying cars 70 years later, and they don't look any more imminent than they did then. I think the lesson there is more "not every dream eventually gets made a reality".
"We always overestimate the change that will occur in the next two years and underestimate the change that will occur in the next ten."
I think about this quote a lot these days, especially while reading Hacker News. On one hand, AI doesn't seem to be having the productivity and economic impacts that were predicted, but on the other, LLMs are getting gold medals at the Math Olympiads. It's like the ground is shifting beneath our feet, but it's still too slow to be perceptible.
Rain, Snow etc. are still challenges but needs a bold bet in a place that wants to show how futuristic it is. The components are in place (Waymo cars), what is needed is high enough labor cost to justify the adoption.
This is exactly what came to my mind also.
https://www.reuters.com/technology/tesla-video-promoting-sel...
for me i have been riding in waymos the last year and have been very pleased with the results. i think we WANT this technology to move faster but the some of the challenges at the edges take a lot of time and resources to solve, but not fundamentally unsolvable.
they are likely semi autonomous, which is still cool, but I wish they'd be honest about it
The interesting thing is that there are problems for which this rule applies recursively. Of the remaining 20%, most of it is easier than the remaining 20% of what is left.
Most software ships without dealing with that remaining 20%, and largely that is OK; it is not OK for safety critical systems though.
What I find really crazy is that most trains are still driven by humans.
The THSR I rode solved the wildlife problems with a big windshield wiper. Not sure what else there is to do. It’s a train.
When there are relatively few dangerous edge cases, technology often works better than we expect. TikTok's recommendation algorithm and Shazam are in this category.
Gee, seems like about the worst fucking thing in the world for diagnostics if you ask me, but what do I know, my degree is in sandwiches and pudding.
We've seen this with all of the players, with many dropping out due to the challenges.
Having said that, there are several that are fielded right now, with varying degrees of autonomy. Obviously Waymo has been operating in small-ish geofences for a while, but they are managed >200% annual growth readily. Zoox just started offering fully autonomous drives in Vegas.
And even Tesla is offering a service, albeit with safety monitors/drivers. Tesla Semi isn't autonomous at all, but appears ready to go into volume production next year too.
Your prediction will look a lot better by 2030.
Easy to forget the rest of the world does not and never has ticked this way.
Don't get me wrong, optimism and thinking of the future are great qualities we direly need in this world on the one hand.
On the other, you can't outsmart physics.
We've conquered the purely digital realm in the past 20 years.
We're already in the early years of the next phase were the digital will become ever more multi-modal and make more inroads into the physical world.
So many people bring an old mindset to a new context, where maring of errors, cost of mistakes or optimizing the last 20% of a process is just so vastly different than a bit of HTML, JS and backend infra.
The truck part seems closer than the car part.
There are several driverless semis running between Dallas, Houston, and San Antonio every day. Fully driverless. No human in the cab at all.
Though, trucking is an easier to solve problem since the routes are known, the roads are wide, and in the event of a closure, someone can navigate the detour remotely.
These are all stepping stones, and eventually the technology is mature enough to productise. You would be shocked by how good Tesla FSD is right now. It can easily take you on a cross country trip with almost zero human interactions.
Not even close.
The vast majority of people have a small number of local routes completely memorized and do station keeping in between on the big freeways.
You can see this when signage changes on some local route and absolute chaos ensues until all the locals re-memorize the route.
Once Waymo has memorized all those local routes (admittedly a big task), it's done.
Yikes.
I recommend you take some introductory courses on AI and theory of computation.
You can do 99% of it without AGI, but you do need it for the last 1%.
Unfortunately, the same is true for AGI.
So no, they don't have AGI and there is a lot to reach "working under every condition everywhere" levels of self-driving.
This was why he went into industrial robotics instead, where it was clear that the finances could work out today.
Who is "we"? The people who hype "AI"?
For some reason, enthusiasts always think this time is different.
Still true as work conditions are harsh, schedule as well, responsibilities and fines are high but payment is not.
In the case of Musk it has worked out. His lies have earned him a fortune and now he asks Tesla to pay him out with a casual 1 trillion paycheck.
But hey, costs are lower that way.
As I recall, one group had fairly good success, but eventually someone figured out that their data set had images from a low-COVID hospital and a high-COVID hospital, and the lettering on the images used different fonts. The ML model was detecting the font, not the COVID.
[a bit of googling later...]
Here's a link to what I think was the debunking study: https://www.nature.com/articles/s42256-021-00338-7
If you're not at a university, try searching for "AI for radiographic COVID-19 detection selects shortcuts over signal" and you'll probably be able to find an open-access copy.
I couldn't for the life of me understand how this was supposed to work. If the coughing of COVID patients (as opposed to patients with other respiratory illnesses) actually sounds meaningfully different in a statistically meaningful way (and why did they suppose that it would? Phlegm is phlegm, surely), surely a human listener would have been able to figure it out easily.
[1] https://academic.oup.com/pmj/article/98/1157/212/6958858?log...
Sorry, but in the absence of general limiting principles that rule out such a scenario, that's how it's going to shake out. Visual models are too good at exactly this type of work.
Like medicine, self-driving is more of a seemingly-unsolvable political problem than a seemingly-unsolvable technical one. It's not entirely clear how we'll get there from here, but it will be solved. Would you put money on humans still driving themselves around 25-50 years from now? I wouldn't.
These stories about AI failures are similar to calling for banning radiation therapy machines because of the Therac-25. We can point and laugh at things like the labeling screwup that pjdesno mentioned -- and we should! -- but such cases are not a sound basis for policymaking.
They do take tests, don't they?
They don't all score 100% every time, do they?
Everything else besides the above in TFA is extraneous. Machine learning models could have absolute perfect performance at zero cost, and the above would make it so that radiologists are not going to be "replaced" by ML models anytime soon.
>Human radiologists spend a minority of their time on diagnostics and the majority on other activities, like talking to patients and fellow clinicians.
The vast majority of radiologists do nothing other than: come in (or increasingly, stay at home), sit down at a computer, consume a series of medical images while dictating their findings, and then go home.
If there existed some oracle AI that can always accurately diagnose findings from medical images, this job literally doesn't need to exist. It's the equivalent of a person staring at CCTV footage to keep count of how many people are in a room.
I think it may be selection bias.
Generalizing this to all radiologists is just as wrong as the original article saying that radiologists don't spend the majority of their time reading images. Yes, some diagnostic radiologists can purely read and interpret images and file their results electronically (often remotely through PACS systems). But the vast majority of radiology clinics where I live have a radiologist on-site, and as one example, results for suspicious mammograms where I live in Texas are always given by a radiologist.
And as the other comment said, many radiologists who spend the majority of their time reading images also perform a number of procedures (e.g. stereotactic biopsies).
I could have just gone to med school and never deal with layoffs, RTO, etc.
I also recently had surgery and the surgeon talked to the radiologist to discuss my MRI before operating.
It's sort of like saying "sometimes a cab driver talks to passengers and suggests a nice restaurant nearby, so you can't automate it away with a self-driving cab."
Does that sounds like an assistance's job?
In the same fashion as construction worker just shows up, "performs a series of construction tasks", then go home. We just need to make a machine that performs "construction tasks" and we can build cities, railways and road networks for nothing but the cost of the materials!
Perhaps this minor degree of oversimplification is why the demise of radiologists have been so frequently predicted?
Do you have some kind of source? This seems unlikely.
The current "workflow" is primary care physician (or specialist) -> radiology tech that actually does the measurement thing -> radiologist for interpretation/diagnosis -> primary care physician (or specialist) for treatment.
If you have perfect diagnosis, it could be primary care physician (or specialist) -> radiology tech -> ML model for interpretation -> primary care physician (or specialist.
PCPs don't have the training and aren't paid enough for that exposure.
To understand why, you would really need to take a good read of the average PCP's malpractice policy.
The policy for a specialist would be even more strict.
You would need to change insurance policies before your workflow was even possible from a liability perspective.
Basically, the insurer wants, "a throat to choke", so to speak. Handing up a model to them isn't going to cut it anymore than handing up Hitachi's awesome new whiz-bang proton therapy machine would. They want their pound of flesh.
If you're getting a blood test, the pipeline might be primary care physician -> lab with a nurse to draw blood and machines to measure blood stuff -> primary care physician to interpret the test results. There is no blood-test-ologist (hematologist?) step, unlike radiology.
Anyway, "there's going to be radiologists around for insurance reasons only but they don't bring anything else to patient care" is a very different proposition from "there's going to be radiologists around for insurance reasons _and_ because the job is mostly talking to patients and fellow clinicians".
HackerNews is often too quick to reply with a “well actually” that they miss the overall point.
How often do they talk to patients? Every time I have ever had an x-ray, I have never talked to a radiologist. Fellow clinicians? Train the xray tech up a bit more.
If the mote is 'talking to people' that is a mote that doesn't need an MD, or at least not a full specialization MD. ML could kill radiologist MD, radiologist could become the job title of a nurse or x-ray tech specialized in talking to people about the output.
That's fine. But then the xray tech becomes the radiologist, and that becomes the point in the workflow that the insurer digs out the malpractice premiums.
In essence, your xray techs would become remarkably expensive. Someone is talking to the clinicians about the results. That person, whatever you call them, is going to be paying the premiums.
Is this uncommon in the rest of the US?
If we had followed every AI evengelist sugestion the world would have collapsed.
But he said it in the context of a Q&A session that happened to be recorded. Unless you're a skilled politician who can give answers without actually saying anything, you're going to say silly things once in a while in unscripted settings.
Besides that, I'd hardly call Geoffrey Hinton an AI evangelist. He's more on the AI doomer side of the fence.
With interventional radiologists and radio-oncologists it's different but were talking about radiologists here...
At the time? I would say he was a AI evangelist.
People can't tell what they'll eat next sunday but they'll predict AGI and singualrity in 25 years. It's comfy because 25 years seems like a lot of time, it isn't.
https://en.wikipedia.org/wiki/List_of_predictions_for_autono...
> I'd wager in 25 years we'd get there so I think his opinion still has a large percentage of being correct.
What percent, and which maths and facts let you calculate it ? The only percent you can be sure about is that it's 100% wishful thinking
Maybe don't?
And a man to a radiologist for a lumbar perineural injection.
And a person to a radiologist for a subacromial bursa injection.
And a month ago I sent a woman to a radiologist to have adenomyosis embolised.
Also talked to a patient today who I will probably send to a radiologist to have a postnephrectomy urinary leak embolised.
Is an LLM going to do that ?
There is the another issue.
If AI commoditises a skill, competent people with options will just shift to another skill while offloading the commoditised skill to someone else.
Due to automated ECG interpretation built into every machine, reimbursement has plummeted. So I have let my ECG interpretation skills rust while focusing on my neurology and movement disorder skills.They are fun ... I also did part of a master's in AI decades ago ( Prolog, Lisp , good times, machine vision, good times...)
So now if someone needs a ECG, I am probably going to send them to a cardiologist who will do a ECG, Holter, Echo, Stress Echo etc. Income for the nice friendly cardiologist, extra cost and time for the patient and the health system.
I can imagine like food deserts, entire AI deserts in medicine that nobody want to work in. A bit like geriatrics, rural medicine and psychiatry these days.
Automating as much of that as possible and making healthcare more accessible should be pursued. Just like automated ECG interpretation made basic ECG more accessible.
Oof - I hope the tools you're using as a physician are better than in the field as a paramedic.
I have never met a Lifepak (or Zoll) that doesn't interpret anything but the most textbook sinus rhythm in pristine conditions as "ABNORMAL ECG - EVALUATION NEEDED".
AI is going to augment radiologists first, and eventually, it will start to replace them. And existing radiologists will transition into stuff like interventional radiology or whatever new areas will come into the picture in the future.
I am a medical school drop-out — in my limited capacity, I concur, Doctor.
My dentist's AI has already designed a new mouth for me, implants &all ("I'm only doing 1% of the finish-work: whatever the patient says doesn't feel just quite right, yet"—myDMD). He then CNCs in-house on his $xxx,xxx 4-axis.
IMHO: Many classes of physicians are going to be reduced to nothing more than malpractice-insurance-paying business owners, MD/DO. The liability-holders, good doctor.
In alignment with last week's (H)(1)(b) discussion, it's interesting to note that ~30% of US physician resident "slots" (<$60kUSD salary) are filled by these foreigner visa-holders (so: +$100k cost per applicant, amortized over a few years of training, each).
It has similar insights, and good comments from doctors and from Hinton:
“It can augment, assist and quantify, but I am not in a place where I give up interpretive conclusions to the technology.”
“Five years from now, it will be malpractice not to use A.I.,” he said. “But it will be humans and A.I. working together.”
Dr. Hinton agrees. In retrospect, he believes he spoke too broadly in 2016, he said in an email. He didn’t make clear that he was speaking purely about image analysis, and was wrong on timing but not the direction, he added.
It's clearly satire with the little jabs like this.
Similar to how a model that can do "PhD-level research" is of little use to me if I don't have my own PhD in the topic area it's researching for me, because how am I supposed to analyze a 20 page research report and figure out if it's credible or not?
There’s wildly varying levels of quality among these options, even though they could all reasonably be called “PhD-level research.”
AI art is getting better but still it's very easy for me to quickly distinguish AI result from everything else, because I can visually inspect the artifacts and it's usually not very subtle.
I'm not a radiologist, but I would imagine AI is doing the same thing here, making up things that are cancer, missing things that aren't cancer, and it takes an expert to distinguish the false positives from true. So we're back at square one, except the expertise has shifted from interpreting the image to interpreting the image and also interpreting the AI.
I suppose first of all, is that generally agreed? People aren't expecting a LLM to give a radiology opinion, the same as way that you can feed in a PDF or an image into ChatGPT and ask it something about it, are they?
I'm interested whether most people here have a higher opinion of ML than of the generative AIs, in terms of giving a reliably useful output. Or do a lot of you think that these also just create so much checking it would be easier to just have a human do the original work?
I think it's probably worth excluding self-driving from my above question, since that is a particularly difficult area to agree anything on.
I actually disagree in that it's not easy for me at all to quickly distinguish AI images from everything else. But I think we might differ what we mean by "quickly". I can quickly distinguish AI if I am looking. But if I'm mindlessly doomscrolling I cannot always distinguish 'random art of an attractive busty woman in generic fantasy armor that a streamer I follow shared' as AI. I cannot always distinguish 'reply-guy profile picture that's like a couple dozen pixels in dimensions' as AI. I also cannot always tell if someone is using a filter if I'm looking for maybe 5 seconds tops while I scroll.
As a related aside, I've started seeing businesses clearly using ChatGPT for their logos. You can tell from the style and how much random detail there is contrasted with the fact that it's a small boba tea shop with two employees. I am still trying to organize my thoughts on that one.
Edit:
Example: https://cloudfront-us-east-1.images.arcpublishing.com/brookf...
> In 2025, American diagnostic radiology residency programs offered a record 1,208 positions across all radiology specialties, a four percent increase from 2024, and the field’s vacancy rates are at all-time highs.
One reason I hear (anecdotally) that vacancy rates are so high is that fewer top quality people are going into radiology. That is, when med students choose a specialty, they're not just choosing for now, but they need to choose a specialty that will be around in 35-40 years. Many med students see the writing on the wall and are reluctant to invest a huge amount of blood, sweat, tears and money into a residency when tech may potentially short circuit their career eventually.
So what you see is that even though AI is not there yet (I'd really highlight this from the article: "First, while models beat humans on benchmarks, the standardized tests designed to measure AI performance, they struggle to replicate this performance in hospital conditions." For the programmers in the room, it's like AI that can solve all the leetcode problems but then falls over when you get into a moderately complicated situation) but there is a shortage of radiologists now because med students are worried what will happen in 10/15/20 years.
"Human radiologists spend a minority of their time on diagnostics and the majority on other activities, like talking to patients and fellow clinicians"
I've had the misfortune of dealing with a radiologist or two this year. They spent 10-20 minutes talking about the imaging and the results with me. What they said was very superficial and they didn't have answers to several of the questions I asked.
I went over the images and pathology reports with ChatGPT and it was much better informed, did have answers for my questions, and had additional questions I should have been asking. I've used ChatGPT's information on the rare occasions when doctors deign to speak with me and it's always been right. Me, repeating conclusions and observations ChatGPT made, to my doctors, has twice changed the course of my treatment this year, and the doctors have never said anything I've learned from ChatGPT is wrong. By contrast, my doctors are often wrong, forgetful, or mistaken. I trust ChatGPT way more than them.
Good image recognition models probably are much better than human radiologists already and certainly could be vastly better. One obstacle this post mentions - AI models "struggle to replicate this performance in hospital conditions", is purely a choice. If HMOs trained models on real data then this would no longer be the case, if it is now, which I doubt.
I think it's pretty clearly doctors, and their various bureaucratic and legal allies, defending their legal monopoly so they can provide worse and slower healthcare at higher prices, so they continue to make money, at the small cost of the sick getting worse and dying.
This article shows its more likely tech/software is probably the first major job to be disrupted significantly before other industries. There is an assumption from tech workers that you need a tech person to employ the AI to do the automation meaning they are the last to go. I think that assumption is questionable if AI gets good enough. Especially if I can get "spec writers/QA's/BA's/etc" to do the automation of industries once the regulation/liability side is worked out per industry. Hearing and seeing more than just rumors of AI tooling that mirrors whole software developer workflows being trialed in large tech firms now; SWE is the lowest juiciest hanging fruit.
I still assert that the next industry to feel the most pain is the SWE's and tech workers themselves. Skills and expertise in an AI world are no longer moats to your job security and ability to provide for yourself -> regulation, lack of data, physical world interaction, liability, locality. Most professions have some of the above.
Anecdotally in my local social circle as an SWE I'm now seen as the person with the least desirable job from a social status and security perspective, a massive change from 5 years ago. People would rather be a "truck driver" or in this case a "radiologist". I hope I'm wrong of course for my own personal sake.
This said, interpreting images is not an image problem - it’s a human body reasoning problem. If you can’t have AI that replaces any engineer, I’d assume replacing a doctor will be just as unlikely. The healthcare bar is much higher - works in 80% of the coding scenarios may be good enough for software, it’s not good for life critical decisions.
So likely we’re not seeing any impact on jobs from AI in relevant health sectors. Now if you your friends think that the rest of paper pushing won’t be affected, or that their jobs entail some unique people skills, they’re in for a big surprise.
Most engineers, even accountants, any profession with a title really that required some study usually have the moat of liability and/or locality. SWE's don't really have this in general - a unique job that while requiring a degree for many high tech orgs, will be the first to go. As you said 80% is enough for many domains here. Any other engineering profession (e.g. electrical, civil) has other moats that mean they won't be as disrupted.
Most of the people I talk to w.r.t this issue studied in general professions or trades, physical jobs. i.e. SWE is especially affected especially at the higher end where study was required because for the same "effort" of a CS/Engineering degree you could of been in any other profession where there was more protection from AI (bootcamps aside). AI may have the CS/SWE university pathway be redundant - ironic if most college/uni jobs are still safe except for the industry that birthed the AI in the first place.
I think the part that says models will reduce time to complete tasks and allow providers to focus on other tasks is on point in particular. For one CV task, we’re only saving on average <30min of work per study, so it isn’t massive savings from a provider’s perspective. But scaled across the whole hospital, it’s huge savings
Or, far more likely, to cut costs and increase profits.
If you are rich and care about your health (especially as move past age 40), you probably have use for a physiotherapist, a nutritionist, therapist, regular blood analysis, comprehensive cardio screening, comprehensive cancer screening etc. Arguably, there is no limit to the amount of medical services that people could use if they were cheap and accessible enough.
Even if AI tools add 1-2% on the diagnostic side every year, it will take a very, very long time to catch up to demand.
Not saying nothing will come of it, but there is a definite pattern to AI hype cycles, and radiology seems to be one of the recurring points.
At the end of the day if the radiologist makes an error the radiologist gets sued.
If AI replaces the radiologist then it is OpenAI or some other AI company that will get sued each and every time the AI model makes a mistake. No AI company wants to be on the hook for that.
So what will happen? Simple. AI will always remain just a tool to assist doctors. But there will always be a disclaimer attached to the output saying that ultimately the radiologist should use his or her judgement. And then the liability would remain with the human not the AI company.
Maybe AI will "replace" radiologists in very poor countries where people may not have had access to radiologists in the first place. In some places in the world it is cheap to get an xray but still can be expensive to pay someone to interpret it. But in the United States the fear of malpractice will mean radiologists never go away.
EDIT: I know the article mentions liability but it mentions it as just one reason among many. My contention is that liability will be the fundamental reason radiologists are never replaced regardless of how good the AI systems get. This applies to other specialities too.
Are you sure? Who would want to be a radiologist then when a single false negative could bankrupt you? I think it's more likely that as long as they make a best effort at trying to classify correctly then they would be fine.
Not just radiologists but surgeons too for example.
Depends on if the patient sues and how that goes.
It won't lead to bankruptcy as malpractice insurance is there but your premiums would go higher and it could still be costly.
I don’t understand this. All data in existence is fodder for training. Barring the privacy issue which the article implies as orthogonal, training data and actual data are the same set of things.
Test conditions should be identical to real conditions. What in the world is the article saying. What actual differences are there?
The only issue I can think of is when there is LLM like language logic in looking at the results of a scan. Like for example the radiologist looks at 30 sections of the image and they all have different relationships with each other and those relationships end up influencing the outcome. But I doubt it’s like that, radiology should be much simpler than learning a foreign language.
Updated for 2025: https://i.imgflip.com/a5ywre.jpg
I would be very interested if you could provide specific examples.
>It's enhacing the capabilties of radiologists.
So it is not replacing radiologists?
It seems that with AI in particular, many operate with 0/1 thinking in that it can only be useless or take over the world with nothing in between.
Their workday consists of conversations, questions and reading. Something LLMs more than excel at doing, tirelessly and in huge volumes.
And if radiologists are still the top bet due to image recognition being so much hotter, then why not add dermatologists to the extinction roster? They only ever look at regular light images, it should be a lower hanging fruit.
(I'm aware of the nuances that make automation of these work roles hard, I'm just trying to shine some light on the mystery of radiologists being perceived as the perennial easy target)
https://www.outofpocket.health/p/why-radiology-ai-didnt-work...
However, the timelines are far too optimistic. It takes time to refine the technology for day to day use, adapt mindsets and processes, and update regulations as needed. But it will come, it just needs more time.
It's the same story as self-driving vehicles or programming. AI will have an impact. It just takes time.
Our current economic system does not support improved productivity leading to less working (with equal wealth) for the working class.
Because we know how well the jobs address a need, and we also know how many times throughout history we have been promised magic wands that never quite showed up.
And guess who is best equipped to measure the actual level of “magic”? Experts like radiologists. We need them the most along the way, not the least.
If a magic wand actually shows up, it will be obvious to everyone and we’ll all adopt it voluntarily. Just like thousands of innovations in history.
> Some products can reorder radiologist worklists to prioritize critical cases, suggest next steps for care teams, or generate structured draft reports that fit into hospital record systems.
Are there systematic reasons why radiologists in hospitals are inaccurately assessing the AI's output? If the AI models are better than humans in testing novel data then, well, the thing that has changed in a hospital situation compared to the AI-Human testing environment is not the AI, it is the human, under less controlled constraints, additional pressures, workloads, etc. Perhaps the AI's aren't performing as poorly as thought. Perhaps this is why they performed better to begin with. Otherwise, production ML systems are generally not as highly regarded as these models when they perform as significantly below test data sets in production. Some is expected, but "struggle to replicate" implies more.
>"Most tools can only diagnose abnormalities that are common in training data"
Well yes, training on novel examples is one thing. Training on something categorically different is another thing all together. Also there are thresholds of detection. Detecting nothing, or with a a lower confidence, or unknown anomaly, false positive, etc. How much of the inaccuracy isn't wrong, but simply something that is amended or expanded upon when reviewed? Some details here would be useful.
I'm highly skeptical when generalized statements exclude directly relevant information to which an is referring. The few sources provided don't at all cover model accuracy, and the primary factor cited as problematic with AI review, lack of diversity in study composition for women, ethnic variation, children, links to a a meta study that was not at all related to the composition of models and their training data sets.
The article begins as what appears to be a criticism of AI accuracy with the thinness outlined above but then quickly moves on to a "but that's not what radiologists do anyway", and provides a categorical % breakdown of time spent where Personal/Meetings/Meals and some mixture of the others combine to form at least a third that could be categorized as "Time where the human isn't necessary if graphs are being interpreted by models."
I'm not saying there aren't points here, but overall, it simply sounds like the hand-wavy meandering of someone trying to gatekeep a profession whose services could be massively more utilized with more automation, and sure-- perhaps at even higher quality with more radiologists to boot-- but perfect is the enemy of the good etc. on that score, with enormous costs and delays in service in the meantime.
Clinical AI also has to balance accuracy with workflow efficiency. It may be technically most accurate for a model to report every potential abnormality with associated level of certainty, but this may inundate the radiologist with spurious findings that must be reviewed and rejected, slowing her down without adding clinical value. More data is not always better.
In order for the model to have high enough certainty to get the right balance of sensitivity and specificity to be useful, many many examples are needed for training, and with some rarer entities, that is difficult. It also inherently reduces the value of the model it is only expected to identify its target disease 3 times/year.
That’s not to say advances in AI won’t overcome these problems, just that they haven’t, yet.
In particular, doctors appear to defer excessively to assistive AI tools in clinical settings in a way that they do not in lab settings. They did this even with much more primitive tools than we have today... The gap was largest when computer aids failed to recognize the malignancy itself; many doctors seemed to treat an absence of prompts as reassurance that a film was clean
Programming is the first job AI will replace. The rest come later.
Spray-and-Pray Algorithms
After AlexNet, dozens of companies rushed into medical imaging. They grabbed whatever data they could find, trained a model, then pushed it through the FDA’s broken clearance process. Most of these products failed in practice because they were junk. In mammography, only 2–3 companies actually built clinically useful products.
Products actually have to be useful.
There were two products in the space: CAD, and Triage. CAD is basically overlay on the screen as you read the case. Rads hated this because it was distracting and because the feature-engineering based CAD from the 80s-90s was demonstrated to be a failure. Users basically ignored "CADs."
Triage is when you prioritize cases (cancers to the top of the stack). This has little to no value because when you have a stack of 50 cases you have to do today, then why do you care about the order? There were some niche use cases but it was largely pointless. It could actually detrimental. The algotithm would put easy cancer cases on the top, so now the user would spend less time on the rest of the stack (where the harder cases would end up).
*Side note:* did you know that using CAD was a billable extra to insurance. Even through it was proven to not work, for years it remained reimbursable up until a few years ago.
Poor Validation Standards
Models collapsed in the real world because the FDA process is designed for drugs/hardware, not adaptive software. Validation typically = ~300 “golden” cases, labeled by 3 radiologists with majority vote arbitration. If 3 rads say it’s cancer, it’s cancer. If they disagree, it's not a good case for the study. This filtering ignores the hard cases (where readers disagree), which is exactly what models need to handle in the real world. Instead of 500K noisy real-world studies, you validate on a sanitized dataset. Companies learned how to “cheat” by over fitting to these toy datasets. You can explain this to regulators endlessly, but the bureaucracy only accepts the previously blessed process. Note: The previous process was defined by CAD, a product that was cleared in the 80s and shown to fail miserably in clinical use. This validation standard that demonstrated grand historical regulatory failure is the current standard that you MUST use for any devices that look like a CAD in mammography.
Politics Over Outcomes
We ran the largest multi-site prospective (15) trial in the space. Results: ~50% reduction in radiologist workload. Increased cancer detection rate. 10x lower cost per study. We even caught cancers missed in the standard workflow. Clinics still resisted adoption—because admitting missed cancers looked bad for their reputation. Bureaucratic EU healthcare systems preferred to avoid the embarrassment even through it was entirely internal.
I'll leave you with one particularly salient story. I was speaking to the head a large US hospital IT/Ops organization. We had a 30 minute conversation about how to avoid putting our software decision in the EMR/PACS so that they could avoid litigation risk. Not once did we ever talk about patient impact. Not Once...
Despite all that, our system caught cancers that would have been missed. Last I checked at least 104 women had their cancers detected by our software and are still walking around. That’s the real win, even if politics buried the broader impact.
The radiologist said No Fracture! So treat for concussion and release. Two days later, back with unendurable head pain. Reexamine: oh, sorry, yes, fractures. Morphine.
Wtf? Are the radiologists so overworked they can only glance at a test? Or was an AI at fault? I'll never know.
Was it CT or X-rays?
X-rays are garbage compared to ct for facial fractures in my (radiography) experience.
Having taken a lot of both: if it’s abnormal on X-rays you do a CT. If the pain or swelling is huge but the X-ray normal, you do a CT. So why bother with X-rays?