I'm an interventional radiologist with a master's in computer science. People outside radiology don't get why AI hasn't taken over.
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.
This is like saying that self-driving cars won't ever become a thing because someone behind the wheel needs to be to blame. The article cites AI systems that the FDA already has cleared to operate without a physicians' validation.
> This is like saying that self-driving cars won't ever become a thing because someone behind the wheel needs to be to blame.
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.
> Waymo has what? 700 cars on the road ...
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.
Operating in carefully selected urban locations with no severe weather. They are nowhere close to general purpose FSD.
Self driving cars in 2025 in USA is like Solar PV in China in 2010, it will take a while, but give them time to learn, adapt and expand.
And where the solar panel were in China in 2000? Because self-driving cars on public roads in USA have been a WIP for 10 years at least.
Yes and I would swear that 1700 of those 2000 must be in Westwood (near UCLA in Los Angeles). I was stopped for a couple minutes waiting for a friend to come out and I counted 7 Waymos driving past me in 60 seconds. Truth be told they seemed to be driving better than the meatbags around them.
(brainfart)
2000 is only 1 of every 100,000
You also have Mercedes taking responsibility for their traffic-jam-on-highways autopilot. But yeah. It's those two examples so far (not sure what exactly the state of Tesla is. But.. yeah, not going to spend the time to find out either)
Merceds has accepted a degree of liability? https://www.prescouter.com/2024/04/mercedes-benz-level-3-dri...
I'm curious how many people would want a second opinion (from a human) if they're presented with a bad discovery from a radiological exam and are then told it was fully automated.
I have to admit if my life were on the line I might be that Karen.
A bad discovery probably means your exam will be read by someone qualified, like the surgeon/doctor tasked with correcting it.
False negatives are far more problematic.
Ah, you're right. Something else I'm curious about with these systems is how they'll affect difficulty level. If AI handles the majority of easy cases, and radiologists are already at capacity, so they crack if the only cases they evaluate are now moderately to extraordinarily difficult?
Let's look at mammography, since that is one of the easier imaging exams to evaluate. Studies have shown that AI can successfully identify more than 50% of cases as "normal" that do not require a human to view the case. If group started using that, the number of interpreted cases would drop in half although twice as many would be normal. Generalizing to CT of the abdomen and pelvis and other studies, assuming AI can identify a sub population of normal scans that do not have to be seen by a radiologist, the volume of work will decline. However, the percentage of complicated cases will go up. Easy, normal cases will not be supplementing the Radiologist income the way it has in the past. Of course, all this depends upon who owns the AI identifying normal studies. Certainly, hospitals or even packs companies would love to own that and generate that income from interpreting the normal studies. AI software has been slow to be adopted, largely because cases still have to be seen by a radiologist, and the malpractice issue has not been resolved. Expect rapid changes in the field once malpractice solutions exist.
The problem is, you don't know beforehand if it's a hard case or not.
A hard to spot tumor is an easy negative result with high confidence by an AI
From my experience the best person to read these images is the medical imaging expert. The doctor who treats the underlying issue is qualified but it's not their core competence. They'll check of course but I don't think they generally have a strong basis to override the imaging expert.
If it's something serious enough a patient getting bad news will probably want a second opinion no matter who gave them the first one.
But since we don't know where those false negatives are, we want radiologists.
I remember a funny question that my non-technical colleagues asked me during the presentation of some ML predictions. They asked me, “How wrong is this prediction?” And I replied that if I knew, I would have made the prediction correct. Errors are estimated on a test data set, either overall or broken down by groups.
The technological advances have supported medical professionals so far, but not substituted them: they have allowed medical professionals to do more and better.
I willing to bet every one here has a relative or friend who at some point got a false negative from a doctor.. Just like drivers that have made accidents.. Core problem is how to go about centralizing liability.. or not.
Id be more concerned about the false negative. My report says nothing found? Sounds great, do I bother getting a 2nd opinion?
You pay extra for a doctor's opinion. Probably not covered by insurance.
That's horrific. You pay insurance to have ChatGPT make the diagnosis. But you still need to pay out of pocket anyway. Because of that, I am 100% confident this will become reality. It is too good to pass up.
Early intervention is generally significantly cheaper, so insurers have an interest in doing sufficiently good diagnosis to avoid unnecessary late and costly interventions.
People will flock to "AI medical" insurance that costs $50/mo and lets you see whatever AI specialist you want whenever you want.
Think a problem here is the sycophantic nature. If I’m a hypochondriac, and I have some new onset symptoms, and I prompt some LLM about what I’m feeling and what I suspect, I worry it’ll likely positively reinforce a diagnosis I’m seeking.
I mean, we already have deductibles and out-of-pocket maximums. If anything, this kind of policy could align with that because it's prophylactic. We can ensure we maximize the amount we retrieve from you before care kicks in this way. Yeah, it tracks.
It sounds fairly reasonable to me to have to pay to get a second opinion for a negative finding on a screening. (That's off-axis from whether an AI should be able to provide the initial negative finding.)
If we don't allow this, I think we're more likely to find that the initial screening will be denied as not medically indicated than we are to find insurance companies covering two screenings when the first is negative. And I think we're better off with the increased routine screenings for a lot of conditions.
Self-care is being Karen since when?
It's not. I was trying to evoke a world where it's become so common place that you're a nuisance if you're one of those people who questions it.
Need to work on the comedic delivery in written form because you just came off as leaning on a stereotype
"Cancer? Me? I'd like to speak to your manager!"
In reality it's always a good decision to seek a second independent assessment in case of diagnosis of severe illness.
People makes mistakes all the time, you don't want to be the one affected by their mistake.
The FDA can clear whatever they want. A malpractice lawyer WILL sue and WILL win whenever an AI mistake slips through and no human was in the loop to fix the issue.
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
What's the difference in the lawsuit scenario if a doctor messes up? If the AI is the same or better error rate than a human, then insurance for it should be cheaper. If there's no regulatory blocks, then I don't see how it doesn't ultimately just become a cost comparison.
> What's the difference in the lawsuit scenario if a doctor messes up? If the AI is the same or better error rate than a human, then insurance for it should be cheaper
The doctor's malpractice insurance kicks in, but realistically you become uninsurable after that.
Same problem for any company.
> What's the difference in the lawsuit scenario if a doctor messes up?
Scale. Doctors and taxi drivers represent several points of limited liability, whereas an AI would be treating (and thus liable for) all patients. If a hospital treats one hundred patients with ten doctors, and one doctor is negligent, then his patients might sue him; some patients seeing other doctors might sue the hospital if they see his hiring as indicative of broader institutional neglect, but they’d have to prove this in a lawsuit. If this happened with a software-based classifier being used at every major hospital, you’re talking about a class action lawsuit including every possible person who was ever misdiagnosed by the software; it’s a much more obvious candidate for a class action because the software company has more money and it was the same thing happening every time, whereas a doctor’s neglect or incompetence is not necessarily indicative of broader neglect or incompetence at an institutional level.
> If there's no regulatory blocks, then I don't see how it doesn't ultimately just become a cost comparison.
To make a fair comparison you’d have to look at how many more people are getting successful interventions due to the AI decreasing the cost of diagnosis.
yeah but at some point the technology will be sufficient and it will be cheaper to pay the rare $2 million malpractice suit then a team of $500,000/yr radiologists
theres an MBA salivating over that presntation
It won't be because your insurance rate will go up.
Also $2M is absurdly absurdly low for a successful medical malpractice suit
This is essentially what's happened with airliners.
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.
No - planes cannot "land themselves with zero human intervention" (...). A CAT III autoland on commercial airliners requires a ton of manual setting of systems and certificated aircraft and runways in order to "land themselves" [0][1].
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/
Not only that but they are even less capable of taking off on their own (see the work done by Airbus' ATTOL project [0] on what some of the more recent successes are).
So I'm not sure what "planes can land on their own" gets us anyway even if autopilot on modern airliners can do an awful lot on their own (including following flight plans in ways that are more advanced than before).
The Garmin Autoland basically announces "my pilot is incapacitated and the plane is going to land itself at <insert a nearby runway>" without asking for landing clearance (which is very cool in and of itself but nowhere near what anyone would consider autonomous).
[0] https://www.youtube.com/watch?v=9TIBeso4abU (among other things, but this video is arguably the most fun one)
Edit: and yes maybe the "pilots are basically superfluous now" misconception is a pet peeve for me (and I'm guessing parent as well)
Taking off on their own is one thing. Being able to properly handle a high-speed abort is another, given that is one of the most dangerous emergency procedures in aviation.
Agreed. I had to actually reject a takeoff in a C172 on a somewhat short runway and that was already enough stress.
Having flown military jets . . . I'm thankful I only ever had to high-speed abort in the simulator. It's sporty, even with a tailhook and long-field arresting gear. The nightmare scenario was a dual high-speed abort during a formation takeoff. First one to the arresting gear loses, and has to pass it up for the one behind.
There's no other regime of flight where you're asking the aircraft to go from "I want to do this" to "I want to do the exact opposite of that" in a matter of seconds, and the physics is not in your favor.
How's that not autonomous? The landing is fully automated. The clearance/talking isn't, but we know that's about the easiest part to automate it's just that the incentives aren't quite there.
It's not autonomous because it is rote automation.
It does not have logic to deal with unforeseen situations (with some exceptions of handling collision avoidance advisories). Automating ATC, clearance, etc, is also not currently realistic (let alone "the easiest part") because ATC doesn't know what an airliner's constraints may be in terms of fuel capacity, company procedures for the aircraft, etc, so it can't just remotely instruct it to say "fly this route / hold for this long / etc".
Heck, even the current autolands need the pilot to control the aircraft when the speed drops low enough that the rudder is no longer effective because the nose gear is usually not autopilot-controllable (which is a TIL for me). So that means the aircraft can't vacate the runway, let alone taxi to the gate.
I think airliners and modern autopilot and flight computers are amazing systems but they are just not "autonomous" by any stretch.
Edit: oh, sorry, maybe you were only asking about the Garmin Autoland not being autonomous, not airliner autoland. Most of this still applies, though.
There's still a human in the loop with Garmin Autoland -- someone has to press the button. If you're flying solo and become incapacitated, the plane isn't going to land itself.
Right. None of this works without humans. :)
One difference there would be that the cost of the pilots is tiny vs the rest that goes into a flight. But I would bet that the cost of the doctor is a bigger % of the process of getting an x-ray.
At the end of day, there's a decision needs to be made and decisions have consequences. And in our current society, there are only one way we know about how to make sure that the decision is taken with sufficient humanity: by putting a human to be responsible for making that decision.
Tesla still hasn't accepted liability for crashes caused by FSD. They in fact fight any such claims in court very vigorously.
They have settled out of court in every single case. None has gone to trial. This suggests that the company is afraid not only of the amount of damages that could be awarded by a jury, but also legal precedent that holds them or other manufacturers liable for injuries caused by FSD failures.
https://www.nbcnews.com/news/us-news/tesla-autopilot-crash-t...
Tesla hit with $243 million in damages after jury finds its Autopilot feature contributed to fatal crash
The verdict follows a three-week trial that threw a spotlight on how Tesla and CEO Elon Musk have marketed their driver-assistance software.
Tesla isn't the north star here
Very questionable reasoning: using a traffic analogy to argue against medical reality.
Medicine does not work like traffic. There is no reason for a human to care whether the other car is being driven by a machine.
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.
You often don't speak to the radiologist anyway. Lots of radiologist work remotely and don't meet and speak with every patient.
There's some sort of category error here. Not every doctor is that type of doctor. A radiologist could be a remote interpretation service staffed by humans or by AI, just as sending off blood for a blood test is done in a laboratory.
> There is no reason for a human to care whether the other car is being driven by a machine.
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.)
So, you need a moral support human? Like a big plushie, but more alive?
You know, for most humans, empathy is a thing; all the more so when facing known or suspected health situations. Good on those who have transcended that need. I guess.
What is the point of the snark? If you are going to find out that you are dying within a year, do you want to get that as an E-Mail?
The point is: I don't see "emotional support" as a vital part of the job of a radiologist.
[dead]
I'm moderately amused that as an interventional radiologist, you didn't bother to mention that IRs do actual procedures and don't just WFH. When I was doing my DxMP residency there was a joke among the radiology residents that IRs had slotted into the cushiest field of medicine and then flopped the landing by choosing the only subfield that requires physical work.
Well I do enjoy procedures. As for diagnostics, it’s very different when you come from a CS background.
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.”
> Unless the law changes...
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.
> I don't know. Doesn't sound like a very big obstacle to me
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.
I think you're right, and the source of the change of legislation may well be related to Medicare and Medicaid cost cutting. After all, it's difficult to see the federal government.
Very few things the general public wants stays just one law amendment away for long. And almost all of those things are for the benefit of powerful people.
I'm not going to comment on whether AI is better than human radiologists or not, but if it is, what will happen is this:
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.
What about the patient that doesn't want to pay $6,000 to go from 99.9% accuracy to 99.95% accuracy?
This is exactly the tradeoff that works in healthcare of poor countries, mostly because the alternative is no healthcare
I don't think the legal framework even allows the patient to make that trade off. Can a patient choose 99.9% accuracy instead of 99.95% accuracy and also waive the right to a malpractice lawsuit?
You know the crazy thing about this? For this application I think it’s similar to spam. AI can easily be trained to be better than a human.
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.
I think its the other way around AI would certainly have better accuracy than a human, AI can see things pixel by pixel.
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
I used to think that too. AI can already do better at screening mammography than a radiologist with a lower miss rate. Given that, insurance rates to cover AI should be even lower than for a radiologist, lawsuits will happen, but with a smaller number of missed cases the number should go down.
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'm a diagnostic radiologist with 20 years clinical experience, and I have been programming computers since 1979. I need to challenge one of your core assumptions.
> 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.
Can a human reliably carefully study for hours on end imaging from screening tests (think of a future world where whole-body MRI scanning for asymptomatic people becomes affordable and routine thanks to AI processing) and not miss subtle anomalies?
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.
> Can a human reliably carefully study for hours on end imaging from screening tests
This is actually very common in radiology where some positions have shifts of 8-12 hours, where one isn't done until all the studies on the list have been read.
> think of a future world where whole-body MRI scanning for asymptomatic people becomes affordable and routine thanks to AI processing) and not miss subtle anomalies?
The bottleneck in MRI is not reading but instead the very long acquisition times paired with the unavailability of the expensive machinery.
If we charitably assume that you're thinking of CT scans, some studies on indiscriminate imaging indicate that most findings will be false positives:
https://pmc.ncbi.nlm.nih.gov/articles/PMC6850647/
Do any of these models know how to say "I don't know"? This is one of my biggest worries about these models.
> When I interpret a study I am not just matching patterns of pixels on the screen with my memory.
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?
Yes, that would be ideal, if we could build such a system. I think we cannot with current tech.
Your belief is held by many, many radiologists. One thing I like to highlight is that LLMs and LVMs are much more advanced than any model in the past. In particular, they do not require specific training data to contain a diagnosis. They don't even require specific modality data to make inferences.
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.
Maybe I’m misunderstanding what you’re calculating, but this math seems wildly off. Sincerely don’t understand an alternate numerical point being made.
> Given a 16-bit, 512 x 512 x 100 slice CT scan, you're looking at 2^16 * 26214400
65536^(512*512) or 65536 multiplied by itself 262144 times for each image. An enormous number. Whether or not assume replacement (duplicates) is moot.
> That's 100 * 26214400 = 262,440,000
There are 100^(512*512) 512x512 100-level grayscale images alone or 100 to the 262144 power - 100 multiplied 262144 times. Again how you paring down a massive combinatoric space to a reasonable 262 mil?
Hi aabajian, thanks for replying!
I might quibble with your math a little. Most CTs have more than 100 images, in fact as you know stroke protocols have thousands. And many scans are reconstructed with different kernels, i.e. soft tissue, bone, lung. So maybe your number is a little low.
Still your point is a good one, that there is probably a finite number of imaging presentations possible. Let's pre-dictate them all! That's a lot of RVUs, where do I sign up ;-)
Now, consider this point. Two identical scans can have different "correct" interpretations.
How is that possible? To simplify things, consider an x-ray of a pediatric wrist. Is it fractured? Well, that depends. Where does it hurt? How old are they? What happened? What does the other wrist look like? Where did they grow up?
This may seems like an artificial example but I promise you it is not. There can be identical x-rays, and one is fractured and one is not.
So add this example to the training data set. Now do this for hundreds or thousands of other "corner cases". Does that head CT show acute blood, or is that just a small focus of gyriform dystrophic calcification? Etc.
I guess my point it, you may end up being right. But I don't think we are particularly close, and LLMs might not get us there.
Haha, I’m also an IR with AI research experience.
My view is much more in line with yours and this interpretation.
Another point - I think many people (including other clinicians) have a sense that radiology is a practice of clear cut findings and descriptions, when in practice it’s anything but.
At another level beyond the imaging appearance and clinical interpretation is the fact that our reports are also interpreted at a professional and “political” level.
I can imagine a busy neurosurgeon running a good practice calling the hospital CEO to discuss unforgiving interpretations of post op scans from the AI bot……
> I can imagine a busy neurosurgeon running a good practice calling the hospital CEO to discuss unforgiving interpretations of post op scans from the AI bot……
I have fielded these phone calls, lol, and would absolutely love to see ChatGPT handle this.
Johns Hopkins has an in house AI unit where they train their own AI's to do imaging analysis. In fact this center made the rounds a few months ago in an NYT story about AI in radiology.
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.
AI can detect a Black person vs a White person via their chest x-rays. Radiologists say there is no difference. Turns out they're wrong. https://www.nibib.nih.gov/news-events/newsroom/study-finds-a...
That being said, there are no radiologists available to hire at any price: https://x.com/ScottTruhlar/status/1951370887577706915
THERE ARE NO RADIOLOGISTS AVAILABLE TO HIRE AT ANY PRICE!!!
True, and very frustrating. Imaging volume is going parabolic and we cannot keep up! I am offering full partnership on day one with no buy-in for new hires. My group is in the top 1% of radiology income. I can't find anyone to hire, I can only steal people from other groups.
"Latest products" and "state of the art" are two very, very different classes of systems. If anything medical has reached the state of a "product", you can safely assume that it's somewhere between 5 and 50 years behind what's being attempted in the labs.
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?
> Every day I see something I have never seen before, and maybe no one has ever seen before.
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}.
Doesn't most of the stuff a radiologist does get double checked anyways by the doctor that orders the scan in the first place? I guess not a more typical screening scan like a mammogram. However, for anything else like a CT, MRI, Xray, etc. I expect the doctor/NP that ordered it in the first place will want to take a look at the image itself and not just the report on the image.
A primary physician (or NP) isn't in a position to validate the judgement of a specialist. Even if they had the training and skill (doubtful), responsibility goes up, not down. It's all a question of who is liable when things go wrong.
As an ER doc I look at a lot of my own studies, because I'm often using my interpretation to guide real-time management (making decisions that can't wait for a radiologist). I've gotten much better over time, and I would speculate that I'm one of the better doctors in my small hospital at reading my own X-rays, CTs, and ultrasounds.
I am nowhere near as good as our worst radiologist (who is, frankly... not great). It's not even close.
As a working diagnostic radiologist in a busy private practice serving several hospitals, this has been my experience as well.
We have some excellent ER physicians, and several who are very good at looking at their own xrays. They also have the benefit of directly examining the patient, "it hurts HERE", while I am in my basement. Several times a year they catch something I miss!
But when it comes to the hard stuff, and particularly cross-sectional imaging, they are simply not trained for it.
I’m fascinated. What makes a great radiologist so much better than the average?
Calling the edge cases correctly, I would think.
I hurt my arm a while back and the ER guy didn't spot the radial head fracture, but the specialist did. No big deal since the treatment was the same either way.
Im not the OP and I’m an MR tech.
I rate techs against non-radiology trained physicians in terms of identifying pathology. However techs aren’t anywhere near the ability of a radiologist.
Persuading junior techs not to scan each other and decide the diagnosis is a reoccurring problem, and it comes up too often.
These techs are trained and are good. I have too many stories about things techs have missed which a radiologist has immediately spotted.
You're specifically trained to look at the scans, and not to do 75 other things as well, only to use scans to aid your whatever you're doing.
Not meaningfully. Beyond basics like a large tumor, a bone break, etc, there’s alot too it.
my pcp doesn’t even have the tools to view an mri even though part of a hospital system.
That’s an issue with that practice. I had the tools to view MRIs in my laptop.
that’s heartening. do you believe the average pcp is competent to review an mri and act on it given the specialist’s report?
also note the hospital system is extremely paranoid about data management and probably wouldn’t allow a pcp to have mri data on a laptop. even specialists seem to only review mri on hospital desktops.
>People outside radiology don't get why AI hasn't taken over
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
This seems like a task an A.I. would be really good at or even just a standard algorithm.
As in.. for small durations of "never" ? ..
But this indicates lack of incentives to reduce healthcare costs by optimisation. If AI can do something well enough , and AI + humans surpass humans leading to costs reductions/ increased throughput this should be reflected in the workflows.
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
Three things explain this. First, while models beat humans on benchmarks, the standardized tests designed to measure AI performance, they struggle to replicate this performance in hospital conditions. Most tools can only diagnose abnormalities that are common in training data, and models often don’t work as well outside of their test conditions. Second, attempts to give models more tasks have run into legal hurdles: regulators and medical insurers so far are reluctant to approve or cover fully autonomous radiology models. Third, even when they do diagnose accurately, models replace only a small share of a radiologist’s job. Human radiologists spend a minority of their time on diagnostics and the majority on other activities, like talking to patients and fellow clinicians
From the article
Another key extract from the article
> The performance of a tool can drop as much as 20 percentage points when it is tested out of sample, on data from other hospitals. In one study, a pneumonia detection model trained on chest X-rays from a single hospital performed substantially worse when tested at a different hospital.
That screams of over fitting to the training data.
Because that is literally happening. I did a bit of work developing some radiological models and sample size for healthy vs malignant is usually 4 to 1. Then you modify the error function so that it makes malignants more significant (you are quite often working with datasets as low as 500 images, so 80/20 training validation split means you are left with 80 examples of malignant) which means that as soon as you take a realistic sample where one specific condition maybe appears in 1/100 or 1/1000 the false positives make your model practically useless.
Of course SOTA models are much better, but getting medical data is quite difficult and expensive so there is not a lot of them.
Remember, the AI doesn’t create anything, so you add risk potentially to the patient outcome and perhaps make advancement more difficult.
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.
> Who’s going to train the future physician the skills needed to develop the procedure?
i would presume that AI taking over won't erase the physical work, which would mean existing training regimes will continue to exist.
Until one day, an AI robot is capable of performing such a procedure, which would then mean the human job becomes obsolete. Like a horse-drawn coach driver - that "job" is gone today, but nobody misses it.
Performing the procedure requires a high level of skill in interpreting scans (angiograms) in real time.
Yeah there’s no more drivers out there, bro. Lol.
The assumption is that more productive AI + humans leads to cost reductions.
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.
Nobody has a profit motive since doctors get their bills paid per procedure and health insurers have a profit cap.
Consider that the profit cap is a percentage, so increased costs in fact increase the amount of profits to be scooped up. So health insurers that would like to see more cash are incentivized to have costs increase!
I also think that the profit cap percentage is not something that applies across the board to every single player in the healthcare space.
Wait, explain. The insurer thing, I get: they're capped. The doctors seem definitely to have a profit motive!
I think the idea there is that generally speaking doctors get rebated on some sort of defined cost schedule. They'll get $40 from the insurance company for some basic kind of visit, and that is "fixed"...
I don't live in the US and when I did wasn't paying doctors very often... but my impression was that even if the rebate schedule is fixed they could "just" ask for more/less, and the rebate schedule is defined by the insurance company (so the insurance company can increase their costs through this schedule, leading to ways to make profit elsewhere!)
I could be totally offbase, I've always thought the Obamacare profit percentage cap to be a fake constraint.
The doctor doesn’t have a motive to replace himself with an AI? That’s what I meant.
From real world experience as a patient that has had a lot go wrong over the last decade. The problem isn’t lack of automation, it’s structural issues affecting cost.
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.
Calling that "low hanging fruit" isn't accurate, because entrenched and powerful interests benefit from it being kept that way. That extra $750 is valuable to the capitalist that gets it. The jobs to process those appeals and reviews are valuable to the employees who do them. Deleting all of this overnight will fuck these people over to varying degrees, and it could even have macroeconomic implications.
With that said, although it will not be easy, this shit needs to change. Health care in the United States is unacceptably expensive and of poorer quality than it needs to be.
Risks in traditional medicine are standardized by standardized training and credentialing. We haven't established ways to evaluate the risks of transferring diagnostic responsibility to AIs.
> 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...
Having paid $300 for a 10 minute doctor visit, in which I was confidently diagnosed incorrectly, it will not take much for me to minimize my doctor visits and take care into my own hands whenever possible.
I will benefit from medical AI. There will soon come a point where I will pay a premium for my medical care to be reviewed by an AI, not the other way around.
If you’d trust generative AI over a physician, go in wide-eyed knowing that you’re still placing your trust in some group of people. You just don’t have an individual to blame if something goes wrong, but rather the entire supply chain that brings the model and its inference. Every link in that chain can shrug their shoulders and point to someone else.
This may be acceptable to you as an individual, but it’s not to me.
You might pay for a great AI diagnosis, but what matters is the diagnosis recognized by whoever pays for care. If you depend on insurance to pay for care, you're at the mercy of whatever AI they recognize. If you depend on a socialized medical care plan, you're at the mercy of whatever AI is approved by them.
Paying for AI diagnosis on your own will only be helpful if you can shoulder the costs of treatment on your own.
At least you can dodge false diagnosis which is important especially when it can cause irreversible damage to your body
Under the assumption that AI has perfect accuracy. Perhaps you dodged the correct diagnosis and get to die 6 months later due to the lack of treatment. Might as well flip a coin.
Doesn't have to be "perfect accuracy". It just has to beat the accuracy of the doctor you would have gone to otherwise.
Which is often a very, very low bar.
What do you call a doctor who was last in his class in medical school? A doctor.
> Doesn't have to be "perfect accuracy". It just has to beat the accuracy of the doctor you would have gone to otherwise.
They made an absolute statement claiming that AI will "at least" let them dodge false diagnosis, that implies a diagnostic false positive rate of ~0%. Otherwise how can you possibly be so confident that you "dodged" anything? You still need a second opinion (or third).
If a doctor diagnosed you with cancer and AI said that you're healthy, would you conclude that the diagnosis was false and skip treatment? It's easy to make frivolous statements like these when your life isn't on the line.
> What do you call a doctor who was last in his class in medical school? A doctor.
How original, they must've passed medical school, certification, and years of specialization by pure luck.
Do you ask to see every doctor's report card before deciding to go with the AI or do you just assume they're all idiots?
And what's the bar for people making machine learning algos? What do you call a random person off the street? A programmer.
Part of the challenge is that machines are significantly different. The radiologist’s statement that an object measured from two different machines is the same and has not changed in size is in large part judgement. Building a model which can replicate this judgement likely involves building a model which can solve all common computer vision tasks, has the full medical knowledge of an expert radiologist, and has been painstakingly calibrated against thousands of real radiologists in hospital conditions.
> If AI can do something well enough , and AI + humans surpass humans leading to costs reductions/ increased throughput this should be reflected in the workflows.
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!)
The article points out that the AI + humans approach gives poorer results. Humans end up deferring to or just accepting the AI output without double checking. So corner cases, and situations where the AI doesn't do well just end up going through the system.
This is what I worry about - when someone gets a little lazy and leans too heavily on the tool. Perhaps their skills diminish over time. It seems AI could be used to review results after an analysis. That would be ok to me, but not before.
If we were serious about reducing healthcare cost by optimization then we would be banning private equity from acquiring hospitals.
What is there to indicate "we" or anyone is serious about reducing healthcare costs? The only thing that will reduce costs is competitive pressure. The last major healthcare reform in the US was incredibly anti-competitive and designed with a goal of significantly raising costs but transferring those costs to the government. How could healthcare costs ever go down when the ONLY way for insurers to increase profits is for costs to go up as their profit is capped at a percentage of expenses.
>...The only thing that will reduce costs is competitive pressure.
Unfortunately, just yesterday there were a surprising amount of people who seemed to argue that increased competition would at best have no effect, and at worst, would actually increase prices:
https://news.ycombinator.com/item?id=45372442
> What is there to indicate "we" or anyone is serious about reducing healthcare costs?
I agree, we clearly aren’t. That’s my point.
Or, maybe artifacts justify prices less so than amounts of souls bothered will. Robotic medical diagnosis could save costs, but it could suppress customers' appetite too, in which case, like you said, commercial healthcare providers would not be incentivized to offer it.
"AI" literally could not care if you live or die.
That's more than a problem of inertia
I think the one thing we will find out with the AI/Chatbot/LLM boom is: Most economic activity is already reasonably close to a local optimum. Either you find a way to change the whole process (and thereby eliminate steps completely) or you won't gain much.
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)
I mean the company providing the AI is free to assume malpractice insurance. If that happens then there is definitely a chance.
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.
Did the need raise through the use of silicon X ray detectors that improved the handling of images and reduced the time needed to get done imaging meaning that it made it faster, cheaper and less cumbersome, increasing the number of requests for X ray imaging?
At some point medical equipment is certified in some way for use. Could the same happen for imaging AIs?
The article mentions a system for diabetic retinopathy diagnosis that is certified and has liability coverage. It sounds like it's the only one where that occurs. For everything else, malpractice insurance explicitly excludes any AI assisted diagnosis.
Malpractice insurance tends to exclude the diabetic retinopathy one too.. the vendor has to provide insurance.
But the equipment is operated by a person, and the diagnostic report has to be signed off by a person, who has a malpractice insurance policy for personal injury attorneys to go after.
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.
For real though how close are we to a product that takes an order for an ED or inpatient CT A/P, protocols it then reads the images and can read the chart and spits out a dictated report without any human intervention that ends up usable as is even 90% of the time.
Right, the last 10% will be expensive or you accept a potential 10% failure rate.
Maybe I should have said 5%. 90% was a made up threshold. How close are we to even a basic “level 5”, ED doc puts in order for indication: “concern for sepsis, lol”, rad tech does their thing and a finished read appears, with no additional human involved except for maybe a review but not even 50% of the time is any addendum needed.
We are not close at all.
So you're telling me the reason an extremely expensive yet totally redundant cost in the healthcare infrastructure will remain in place is because of regulatory capture?
You're probably right.
However, it is also because, in matters of life or death, as a diagnosis from a radiologist can be, we often seek a second opinion, perhaps even a third.
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.
Provided enough political will (and you know that this can be correlated to many factors, like lobbying), laws can be changed.
Is there a risk that radiologists miss stuff because they get a pre-written report by AI that pushes them in a certain direction?
Maybe, but there is already that risk of some influence from other doctors, patients, nurses and general circumstances.
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.
I am actively researching this friction and others like it. I would love it if you happened to have recommendations for literature that 3rd parties can use to corroborate your experience (I’ve found some, but this is harder to uncover than I expected as I’m not in the field)
So you studied like 8 years of med school [1] and 2 years of CS? Damn! That’s a lot.
[1] I don’t know the US system so it’s just a guess
4 years undergrad (CS, math, bio)
4 years med school
2 years computer science
6 years of residency (intern year, 4 years of DR, 1 year of IR)
16 years...
Why do you need the first 4 years undergrad - in places like the UK you can go straight to study medicine from secondary school at age ~18?
The belief is -- and it is one that I share -- that this makes for more well rounded, human physicians.
Additionally, a greater depth of thinking leads to better diagnosticians, and physician-scientists as well (IMO).
Now, all of this is predicated on the traditional model of the University education, not the glorified jobs training program that it has slowly become.
Cynically, it's also a way for the US system to gatekeep "poor" people from entering professions like medicine and law because of the extra tuition fees (and opportunity time-cost) needed to complete their studies.
I am a natural skeptic, but in this case I think it is just an accident of history how different systems developed.
FWIW, although this is not well known, many medical schools offer combined BA/MD degrees, ranging from 4-8 years:
https://students-residents.aamc.org/medical-school-admission...
When I went 20 years ago, my school did not require a bachelor's degree and would admit exceptional students after 2 years of undergraduate coursework. However I think this has now gone away everywhere due to AAMC criteria
In Australia, Medicine was/is typically an undergrad degree.
In the mid-90s my school started offering a Bachelor of Biomedical Science which was targeted at two audiences - people who wanted to go into Medicine from a research, not clinical perspective, and people who wanted to practice medicine in the US (specifically because it was so laborious for people to get credentialed in the US with a foreign medical degree, that people were starting to say "I will do my pre-med in Australia, and then just go to a US medical school").
When I was in Australia and applying to study medicine (late 90s):
Course acceptance is initially driven by academic performance, and ranked scoring.
To get into Medicine at Monash and Melbourne Universities, you'd need a TER (Tertiary Entrance Ranking) of 99.8 (i.e. top 0.2% of students). This number was derived by course demand and capacity.
But, during my time, Monash was known for having a supplementary interview process with panel and individual interviews - the interview group was composed of faculty, practicing physicians not affiliated with the university, psychologists, and lay community members - specifically with the goal of looking for those well-rounded individuals.
It should also be noted that though "undergrad", there's little difference in the roadmap. Indeed when I was applying, the MBBS degree (Bachelor of Medicine and Surgery) was a six-year undergrad (soon revised to five), with similar post grad residency and other requirements for licensure and unrestricted practice.
Me:
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
Add to that that the demand for imaging is not fixed. Even if somehow imaging became a lot cheaper to do with AI, then likely we would just get more imaging done instead of having fewer radiologists.
“Unless the law changes”
Famous last words.