I just want to point out that the term "A.I." gets used pretty loosely in these articles, as if A.I. is a monolithic commodity that you plugin to your software to make it do chatGPT.
The example in the article is an in house developed "A.I." to help radiologists assess images. Digging a bit deeper it seems they are using mostly old CNN type architectures with a few million parameters.[1]
I think it still remains to be seen what a 1T+ parameter transformer trained specifically for radiology will do. I think anyone would be confident that a locally run CNN will not hold a candle to it.
There are many, many papers and projects out there about tuning foundation models on various types of medical imaging data, and many organized efforts to produce large medical imaging datasets to feed that training. This stuff is well-known in the trenches and can improve on the older, smaller CNNs in some ways, but not in a way that's produced any step change in automated capabilities. People are certainly working on it!
> I think it still remains to be seen what a 1T+ parameter transformer trained specifically for radiology will do
Does image processing of something like this scale with parameters?
It makes sense that language continues to scale as the vector space is huge. Even models that generate images scale as the problem space is so large.
But here there are only so many pixels in the image and they are a lot more uniform. You likely can't have 1T images to train on, so wouldn't the model just overfit and basically memorize every image it has ever seen?
Unfortunately, VLMs don’t seem to work. Even the best ones (eg, Gemini 2.5 Pro) hallucinate to a laughable degree. I’ve been seeing them make up large-scale features in images that absolutely no human would ever mistake.
What does work is narrow AI trained to do very specific tasks (such as medical image segmentation). But I don’t think that type of AI will benefit much from massive scaling.
Medical records remain a walled garden mostly because of HIPAA. Unlike all the AI development that has managed to skirt the copyright law to train large models, training a 1T+ parameter transformer on a large enough dataset will need a lot of consumers to give up their medical records.
I've noticed a funny trend where people will bring up an ethical concern and some dweeb will immediately roll out the Chinese strawman like it justifies their incredibly dubious position.
"We should violate medical privacy laws on an enormous scale because China" Okay buddy.
I wasn't suggesting that we should, but in my lifetime we went from hospital admissions and discharges being published in the local paper, to that information being some of the most closely guarded with (theoretically) severe penalties for disclosure.
I think it was AIDS and people not wanting to be outed that changed things.
But whatever the reason, we now have HIPAA and it adds a huge amount of complexity to the management of medical records, and a huge amount of complexity to any research that needs access to a broad spectrum of medical case histories (such as training AI). Other countries don't have these concerns, and will outcompete the US on developing these capabilities.
Concerns about HIPAA as an obstacle to research are largely a red herring. Almost no clinically useful research can be accomplished by just processing large volumes of historical patient charts. In practice the researchers will have to gather additional data from study subjects so they're going to need to obtain consent anyway.
People keep claiming that other countries will out compete us in medical research and it keeps not happening.
yes but you can train relatively dumb AI on high volumes of historical data and presenting factors. This is typically just a 'diagnostic assist', but often significantly better than a human
I did work on a diagnostic assist tool developed by a large pharma which diagnosed between Asthma and COPD. GPs get this right 52% time, specialists just over 60 and AI came in over 80%...using 12 relatively mundane input variables. I believe there are a lot of these situations, but not clear pathway to FDA approval
If we become worse off than China, it's because we squandered a massive decades long lead and actively moved critical industrial processes over there, not because we didn't let AI companies scrape HIPAA data.
The Red Scare never ended for a whole lot of Americans. And my foot is more communist than modern China is. But you know, being afraid of Communism is just as, if not in fact easier, when you have no clue what Communism actually is.
If copyright stayed at 7+7, I would agree that AI companies were "skirting" copyright law, but as I see it life + 70 violates the "limited" terminology in the Constitution by denying people the right to create derivative works.
IP holders did this to themselves by extending the term beyond any reasonable interpretation of "limited".
Hell, Jack Valenti wanted to extend it to "forever minus a day"
IMO, copyright law in its current form is unconstitutional and unnatural.
Epic probably has enough breadth on their own to do it, although they don’t actually have an image display system that I know of, so perhaps they don’t have access to the raw images - only the written reports.
From a summary on HHS.gov it says "De-Identified Health Information. There are no restrictions on the use or disclosure of de-identified health information." Maybe someone with more knowledge could expand on the limitations of what counts as "De-Identified" but I think that might work. I followed the reference and nothing in the CFR jumps out at me, but I'm not a lawyer so who knows.
"The radiologist will be with you soon" also gets used pretty loosely in this article, since a radiologist is not someone you are likely to speak with. They sit in their home office in Montana clicking through images all day, and send thru their diagnosis to a different practitioner at the office you are visiting.
A radiologist practicing in the USA can live anywhere in the world, but they must be trained in the USA, licensed in the state(s) where they read from, credentialed at the hospitals or facilities they read from, and carry medical malpractice insurance.
The point being, hospitals can't throw cheap Indian or Filipino radiologists at the worklist.
Not sure about other countries, but RAs do outsource work to RRAs(1) and then sign the final product. This is relatively new as Medicare/Medicaid didn't recognize RRAs as mid-level practitioners until 2019. I suspect the answer is yes for ARRT-certified RRAs, but it will be state-specific (similar to the answer above about RAs).
For a while it seemed that "AI" topics were treated as niche and news articles used relatively specific terms of art like deep learning or machine learning. Somehow ChatGPT changed that and now everything is just AI.
ChatGPT changed that because it, along with other current advances, brought AI out of its niche and into widespread mainstream usage. The laboratory the source article covers (https://www.mayo.edu/research/labs/radiology-informatics/ove...) publishes some of the technical detail you're interested in, but in 2025 it's also interesting to people less familiar with theoretical computer science. The doctors quoted in the source article, using programs that can calculate kidney volume or detect blood clots, don't necessarily want or need to know the difference between "artificial intelligence" and "deep learning".
I do my best to filter ai topics here using ublock, and it's blocked about a third of the posts, and yet every day there's a new variation, whether a new technique/technology, a new product (open source or commercial), etc... that I have to add to the filter. It's really offputting. Yet apparently this is the only site besides lobste.rs that has programming related content that is of any depth that I know of.
I really wish hacker news had the ability to tag posts and let me filter (and tag them generally not have to get specific on each new buzzword or product).
I'll speak for myself, for me it's illustrated by the spam sketch from Monty Python. It crowds out everything, including my mental space. The hype and disconnect from reality make it intriguing to my brain so I am very likely to end up reading the comments about it, but it's usually a waste of my time. Software does what it does and my opinions (or anyone else's) just don't matter. I should focus my time on learning and building instead of trying to understand hype.
I would say generally speaking that people who assume AI will replace somebody else's job believe that these jobs are merely mechanical and there is no high-level reasoning involved that would basically require AGI (when that comes about nobody is safe). So the model of the AI radiologist assumes the only job of a radiologist is to classify images, which is pretty vulnerable to near-future disruption.
I imagine, given the training involved, the job involves more than just looking at pictures? This is what I would like to see explained.
The analogy would be the "95% of code is written by AI" stat that gets trotted out, replacing code with image evaluation. Yes AI will write the code but someone has to tell the AI what to write which is the tricky part.
>> jobs are merely mechanical and there is no high-level reasoning involved
This is a very binary way of thinking about it. More usual is that components of many professions are mechanical and can be automated, while other components are not mechanical and thus harder to automate. Regardless, if some % of the mechanical work goes away, it is unlikely that human workers just work less. Instead, they will work just as much and the overall demand for workers is reduced by %
Between 1985 and 2025 we went from programming in 8086 assembly language to high level languages like python, typescript and go. These automate a lot of the drudgery of programming in asm, so why has the overall demand for programmers not diminish (in fact it increased massively)?
Can you describe a driving scenario where the correct action couldn't be determined "mechanically"? Are you thinking of something like the trolley problem?
That is such a contrived phenomenon, it has taken decades of lobbying and destruction of political accountability to create the conditions where a person considered sane would touch that idea instead of immediately skipping over to driverless trains.
Incredibly wasteful gimmick, I don't get why the usians are still struggling away at it now that the chinese seem to have already done it.
I don't know where you've gotten the idea that cars are something you can "skip over" on the way to trains. Public transit is great, and the US should do it better, but it doesn't obsolete every other form of transit. The vast majority of people in every country who can afford access to cars use them regularly.
If (as acknowledged in the article) AI automates at least part of the work of radiologists (e.g. tool that "saves her 15 to 30 minutes each time she examines a kidney image"), don't you fear that the demand of radiologists will decline? Even if some are still needed, surely if a hospital needs X reports per day and now Y radiologists are sufficient to provide them rather than the current Z (Y<Z), that should be something for people considering your career to take into account?
On the other hand, how much of your confidence in not being replaced stems from AI not being able to do the work, and how much from legal/societal issues (a human needing to be legally responsible for the diagnoses)? Honestly the description in the article of what a radiologist does "Radiologists do far more than study images. They advise other doctors and surgeons, talk to patients, write reports and analyze medical records. After identifying a suspect cluster of tissue in an organ, they interpret what it might mean for an individual patient with a particular medical history, tapping years of experience" doesn't strike me as anything impossible for AI within a few years, now that models are multimodal and they can work with increasing amounts of text (e.g. medical histories).
No. There is no area of medicine where a boost in productivity will cause doctors to have idle time. The wait times may decrease, throughput may increase, diagnosis accuracy may improve, even costs may decrease (press x to doubt) but no, there will never be a case where we will need less radiologists.
Wonderful insight that I'd never considered. Talk to almost anyone in America and they'll tell you about a health issue that they or their family are deferring due to lack of access. Waiting months or years to simply talk to a specialist, let alone find one that can help, is sadly the norm. Patients rightly feel it's a waste of their time so won't even seek treatment.
Remove that barrier to access and we won't see a shiny new streamlined medical system but rather a flood of new patients requiring even more bureaucracy to manage.
100% why I bothered to stick w/ OneMedical after the Amazon acquisition. I tried to get a "New Dr/New Patient" appointment with any local doctor I could find that took my insurance, and the closest I could get was 3 weeks away.
I'm not worried about medical folks having "time to clean" any time soon.
To be clear, we also need more medical professionals in general -- they're not keeping up with the population and it's making us all less healthy. Three to six months, or more, in the SF bay area for some critical appointments is really unacceptable, but there's not really an option given supply and demand.
I'm sure this will all get better with captain brainworms at the helm.
>> No. There is no area of medicine where a boost in productivity will cause doctors to have idle time. The wait times may decrease, throughput may increase, diagnosis accuracy may improve, even costs may decrease (press x to doubt) but no, there will never be a case where we will need less radiologists.
I dont think this is how market participants may think about it. If costs decrease, some group of radiologists will drop out of the market. We may not "need" less radiologists, but we're signaling we need less of them by not paying them as much as before.
Much like I still "need" a photographer, but short of weddings, I'm not willing to pay as much as before. I may well acquire a photogrpher for a portrait, but it would have to be priced competitively to a selfie.
You're making the common mistake of thinking that healthcare is some sort of normal market subject to simple supply-and-demand economics rules. The reality that supply is heavily constrained by CMS funding for residency slots, prices are (mostly) fixed by a few large payers, and patient demand is effectively infinite. There is a serious shortage of radiologists already and it's getting worse as the population grows older and sicker. If AI tools make radiologists more productive then more imaging studies will be ordered.
>> supply is heavily constrained by CMS funding for residency slots
I keep hearing this argument. Then I look at an Insurance Explanation of Benefits statement. A Radiologist might make $1-2k/day in billings. If you are in a balance-billing state, whatever insurance doesnt pay gets forward billed to the patient. On a standard 252 day workyear, that is $250-500k/yr in billings. The average resident salary is 70k, lets assume 100k with benefits.
Of course there is plenty of overhead, but from the math i'm seeing, the average Radiology Resident is a 150-400k net revenue center. Is the overhead really greater than 150-400k/yr/resident?
What am I missing, why would a profit center need "CMS funding"? From what my doctor friends tell me, the real bottleneck is the unwillingness of AMA and ABR to open up more radiology residency spots (artificial supply constraint) with "CMS funding" a boogeyman and red herring.
Contrary to what your doctor friends might have told you, the AMA has no power to restrict the number of residency slots. They are a private membership organization with no regulatory or accreditation authority. At one time they did lobby Congress to restrict graduate medical education funding but have since reversed that position.
A minority of funding does come from sources other than CMS. Teaching hospitals are largely free to add more residency slots if they want to. The fact that most hospitals don't do this indicates that GME programs are largely unprofitable.
>> You cannot bill for a resident's work, without an attending signing off on it.
Isnt that the case for almost every industry? I was a management consultant and our Partner signed off on the work before billing clients. Architects bill clients, but only when the lead architect signs off on the plans. Same for civil engineers. Same for magazines where the editor signs off on most major pieces.
Faster CPUs, better screens, helpful IDEs, heck even Gen-AI itself did not reduce the need for software engineers let alone decrease the costs. As mentioned in another comment, The Jevons Paradox implies that in certain industries, increased productivity may actually lead to more consumption (therefore propping up the demand / cost) despite not being intuitive so.
The only industries that has observed the opposite effect I can think of so far are translators and stock photographers. Maybe also proof readers - but is that gen Ai or did spellcheckers already kill that branch?
> Unless I am mistaken, the work of radiologists is more defined.
When I had cancer I had to go to an Interventional Radiologist. Never heard of it before. But they use X-Ray in real-time (flouroscopy?) to guide surgery. Pretty neat.
I've authored and contributed to several open source projects over the years, and I'm currently doing a deep dive in CAD/CAM after buying a CNC machine.
I help my practice where I can, and have written a little utility to make generating reports easier, but I would have to quit my job if I was trying to take on the absolutely enormous task of radiology computer aides diagnosis. And I need my job!
> I would have to quit my job if I was trying to take on the absolutely enormous task of radiology computer aides diagnosis
I think about this sort of thing every so often. The difference one person can make by creating a decent piece of software might be one of the best ways of getting a return on investment (in human hours, if not in money).
>> The only industries that has observed the opposite effect I can think of so far are translators and stock photographers. Maybe also proof readers - but is that gen Ai or did spellcheckers already kill that branch?
if the cost for preventative scans goes down, demand will rise. medical demand is incredibly constrained by price. people skip all kinds of tests they need because they can't afford it. the radiologists will have more work to do, not less.
There is a national shortage of radiologists in the US, with many hospital systems facing a backlog of unread cases measuring in the thousands. And, the baby boomers are starting to retire, it's only going to get worse. We aren't training enough new radiologists, which is a different discussion.
Askl to your question on where my confidence stems from, there are both legal reasons and 'not being able to do the work' reasons.
Legal is easy, the most powerful lobby in most states are trial attorneys. They simply won't allow a situation where liability cannot be attached to medical practice. Somebody is getting sued.
As to what I do day to day, I don't think I'm just matching patterns. I believe what I do takes general intelligence. Therefore, when AI can do my job, it can do everyone else's job as well.
> We aren't training enough new radiologists, which is a different discussion.
About that, I think the AMA is ultimately going to be a victim of its own success. It achieved its goal of creating a shortage of medical professionals and enriching the existing ones. I don't think any of their careers are in danger.
However, long term, I think magic (in the form of sufficiently advanced technology) is going to become cost effective at the prices that the health care market is operating at. First the medical professionals will become wholly dependent on it, then everyone will ask why we need to pay these people eye-watering sums of money to ask the computers questions when we can do that ourselves, for free.
The trial lawyer angle doesn't seem accurate. Did trial lawyers prevent pregnancy tests from rolling out? COVID tests? Or any other automatic diagnostic, as long as it was reasonably accurate?
Not as far as I know. Once an automated diagnostic is reasonably accurate, it replaces humans doing the work manually. The same would be true of anything else that can be automatically detected.
No comment on whether radiology is close to that yet, although I don't think a few-million-param neural network would tell us much one way or another.
Are you aware of any states in the US that have made it harder to sue doctors for malpractice?
My point, which I made poorly, is this: There's a reason doctors that went to medical school in India and trained as Radiologists in India can't read US cases remotely for a fraction of the cost of US trained and licensed radiologists.
It's not because the systems to read remotely don't exist.
It's not because they're poorly trained or bad doctors.
No, it's because the AMA lobbies to protect American doctors' jobs, and refuses to license them to practice in the US. Of course you can still sue for medical malpractice regardless of citizenship. Trial lawyers have nothing to do with it, American doctors who don't want competition are entirely to blame.
A big wrinkle in AI evangelism is that proponents don’t understand the concept of human judgment as a “learned” skill - it takes practice and AI models / systems do not suffer consequences the way humans do. They have no emotions and therefore can not “understand” the implications of their decisions.
For context, generative AI music is basically unlistenable. I’ve yet to come across a convincing song, let alone 30 seconds worth of viable material. The AI tools can help musicians in their workflow, but they have no concept of human emotion or expression and it shows. Interpreting a radiology problem is more like an art form than a jigsaw puzzle, otherwise it would’ve been automated long ago (like a simple blood test). Like you note, the legal system in the US prides itself on “accountability” (said tongue in cheek) and AI suffers no consequences.
Just look how well AI worked in the United Healthcare deployment involving medical care and money. Hint: stock is still falling.
It's not really my genre, so my judgment is perhaps clouded. Also, I find the dumb lyrics entertaining and they were probably written by a human (though obviously an AI could be prompted to do just as well). I am a fan of unique character in vocals and I love that it pronounces "A-R-A" as "ah-ahr-ah", but the little bridge at 1:40 does nothing for me.
> A big wrinkle in AI evangelism is that proponents don’t understand the concept of human judgment as a “learned” skill
Which is ironic given how much variation in output quality there is based on the judgement of the person using the LLM (work scope, domain, prompt quality, etc.)
You say "Legal is easy, the most powerful lobby in most states are trial attorneys."
The most powerful lobby in this case is the ABR which carefully constricts coveted residency spots in Radiology to create an artificial scarcity and keep up incomes. It is the opposite of, say, technology, where we have no gatekeeper and supply increases.
The ABR will say that Medicare doesnt fund enough residency spots, but all you need to do is look at an EoB and see that a week of residency billings covers the entire cost of the resident.
If a teaching hospital with an existing radiology residency program wanted to add one more spot, does the ABR have any power to stop them? If Medicare offered more funding to a teaching hospital to add more spots would the hospital turn it down?
I worked on an autocontouring model but we could not get very high accuracy for it to be adopted commercially. The algorithm would work for some organs but would totally freak-out on the others. And if the anatomy was out of norm then it would not work at all. This was 5 years ago, I see Siemens [0] has a similar tool. I remember shadowing a dosimetrist contouring all the Organs-At-Risk (OAR) and it took about 3-4 hours to contour one CT image of thoracic region. Do you know how much better the autocontouring tools have become?
May be in the West. However more practical countries like China with a huge population and clear incentive to provide healthcare to a large population at reduced cost will have incentives to balance accuracy and usefulness in a better way.
My personal opinion is that a lot of Medical professionals are simply gatekeeping at this point of time and using legal definitions to keep changing goalposts.
However this is a theme that will keep on repeating in all domains and I do feel that gradual change is better than sudden, disruptive change.
AI in healthcare is going to add so many layers of indirection for malpractice lawsuits. You'll spend years and lots of $$$ trying to figure out who the defendant would ultimately be, only for it to end up being a LLC that unfortunately just filed for bankruptcy.
The worry isn't that you'll find an AI sitting on the chair that a radiologist used to sit. It's that the entire field of radiology gets reduced down to a button click on a software.
The other doctors will still be there for you to sue.
So the question is, “what if people bought an x-ray machine (affordably available on Amazon)and started using it without training on radiological safety”?
Here in New Zealand you get a licence after purchasing the equipment, and require the machine spec, date of manufacture and serial number to get the licence.
It does require a radiologist name on the paperwork as they are the one with the radiation licence. However it is possible to get one if not a radiologist (dentists do, and radiographers have).
Being licensed to use the equipment is the hard bit, as insurance companies require accreditation which is hard to get.
Isn't that 90% of going to get scan is right now? You'll still need the "shop" to provide the equipment and the tech with the training to know what/where to scan, but you might get the results a bit faster? Are the radiologists the chokepoint now, or is it the techs?
That's the way it already works in many cases, just like with outpatient surgery clinics and other outpatient specialist practices. There is a critical difference, though, because radiology also has sub-specialities and someone focused on orthopedics probably isn't the one you'd want reading your cardiology images, nor would you want your ophthalmologic radiologist trying to diagnose a brain CT.
I initially read this as ‘Medical Imaging tech who is an entrepreneur.’ I now think you’re a radiologist?
Any particular interests?
Fixing the shiite RIS/PACS world and the hell of hl7 would make me happy. I’m an MR tech, and just finished trying to make a scan description ‘MRI Cervical Spine + Right Brachial Plexis’
Are AI models able to detect abnormalities that even an experienced radiologist can't see? i.e. something that would look normal to a human eye but AI correctly flags it for investigation? Or are all AI detections 'obvious' to human eyes and simply a confirmation? I suspect the latter since it was human annotated images the model was trained on.
For example, let's say I'm looking at a chest x-ray. There is a pneumonia at the left lung base and I am clever enough to notice it. 'Aha', I think, congratulating myself at making the diagnosis and figuring out why the patient is short of breath.
But, in this example, I stop looking closely at the X-ray after noticing the pneumonia, so I miss a pneumothorax at the right lung apex.
I have made a mistake radiologists call 'satisfaction of search'.
My 'search' for the patient's problem was 'satisfied' by finding the pneumonia, and because I am human and therefore fundamentally flawed, I stopped looking for a second clinically relevant diagnosis.
An AI module that detects a pneumothorax is not prone to this type of error. So it sees something I did not. But it doesn't see something that I can't see. I just didn't look.
> I have made a mistake radiologists call 'satisfaction of search'.
Ah, now I have a name for it.
When I've chased a bug and fixed a problem I found that would cause the observed problem behavior, but haven't yet proven the behavior is corrected, I'm always careful to specify that "I fixed a problem, but I don't know if I fixed the problem". Seems similar: found and fixed a bug that could explain the issue, but that doesn't mean there's not another one that, independently, would also cause the same observed problem.
I've been going to RSNA for over 25 years, in all that time, the best I've seen from any model presented to me was the smack the radiologist on the head and say, "you dummy, you should have seen that!" model.
That is, the models spot pathologies that 99.9999% of rads would spot anyway if not overworked, tired, or in a hurry. But, addressing the implication of your question, the value is actually in spotting a pathology that 99.9999% of rads would never spot. In all my years developing medical imaging startups and software, I've never seen it happen.
I'm sure it's a matter of training data, but I don't know if it's a surmountable problem. How do you get enough training data for the machine to learn and reliably catch those exceptions?
One Harvard study trained an AI that could reliably determine someone's race from a chest X-ray. AIs can be trained to see things we can't.
The difficulty is likely in making a good training dataset of labeled images with pathologies radiologists couldn't see. I imagine in some cases (like cancer), you may happen to have an earlier CT scan or X-ray from the patient where the pathology is not quite yet detectable to the human eye.
I suspect that radiologists could identify race from a plain chest x-ray if they were given the patient’s race and asked to start noticing the difference. They just aren’t doing it because, if it’s important, you can just look at the patient.
There are a lot of things in medicine that aren’t in literature, but are well-known among certain practitioners. I’m an anesthesiologist and practice in an area with a large African-American population. About 10-15% (rough guess) of people of West African descent will have a ridiculously strong salivary response to certain drugs (muscarinic agonists). As in, after one dose their mouths will be full of saliva in seconds. We don’t have East Africans for comparison, so I can’t say it’s a pan-Bantu thing, but I have seen it in a Nigerian who lived here. Not in the literature, but we all know it. I had a (EDIT: non-anesthesia) colleague ask me about a hypersecretory response from such a drug. I said, oh, was he black? Yes, how did I know? Because we give those drugs all the time and have eyes. It’s very rare to see in European-descended populations.
It's possible humans could learn to do this, but I'm skeptical they could do it this well. According to the article, humans experts couldn't tell race from the chest X-rays and the researchers couldn't figure out how the AI was detecting race. They fed it corrupted images to figure out what information it was relying on. It was robust against both low pass and high pass filters.
That's the reason rads never train to determine race from a chest x-ray.
BTW, models don't need to train that either. Because if it's important, it's recorded, along with a picture, in the guy's medical record.
I'd just like to gently suggest that determining someone's race from an X-Ray instead of, say, their photograph, is maybe not how we should be burning training cycles if we want to push medical imaging forward. Human radiologists had that figured out ages ago.
You're being snide about the Harvard/MIT researchers being idiots doing useless research because they don't realize radiologists can just look at the patients face, but that's obviously not what happened. They were trying to see if AI could introduce racial bias. They're not releasing an AI to tell radiologists the race of their patients.
According to the article, human experts could not tell race from chest X-rays, while the AI could to do reliably. Further, it could still determine race when given an X-ray image passed through a low pass filter or high pass filter, showing that it's not relying solely on fine detail or large features to do so.
2. because it isnt a competitive market (basically, the American Board of Radiologists controls standards of practice and can slow down technologies seen as competitive to human doctors)?
3. or perhaps 1 doesnt happen because outsiders know the market is guarded by the market itsself?
What's accurate in this article? It's very vague, it can be tldred into "we won't go anywhere, although AI does more and more of our work"
> Radiologists do far more than study images. They advise other doctors and surgeons, talk to patients, write reports and analyze medical records. After identifying a suspect cluster of tissue in an organ, they interpret what it might mean for an individual patient with a particular medical history, tapping years of experience.
AI will do that more efficiently, and probably already does. "tapping years of experience" is just data in training set.
> A.I. can also automatically identify images showing the highest probability of an abnormal growth, essentially telling the radiologist, “Look here first.” Another program scans images for blood clots in the heart or lungs, even when the medical focus may be elsewhere.
> “A.I. is everywhere in our workflow now,” Dr. Baffour said.
> “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.”
Maybe you'll be able to happily retire because inertia, but overall it looks like elevator operator job.
>> Therefore, when I lose my job to AI, so does everyone else.
Not quite right? Some fields are licensed, regulated, and have appointments -- and others are not. AI is most keenly focused on fields w/o licensure barriers
On the one hand, you’re totally right. The job takes general intelligence.
On the other hand, a lot of jobs take general intelligence. You’re right about that too.
It’s difficult to guess the specifics of your life, but: maybe you’ve engaged a real estate agent. Some people use no real estate agent. Some have a robo agent. No AI involved. Maybe you have written a will. Some people go online and spend $500 on templates from Trust & Will, others spend $3,000 on a lawyer to fill in the templates for them, some don’t do any of that at all. Even in medicine, you know, a pharma rep has to go and convince someone to add their thing to the guidelines, and you can look back at the time between the study and adoption as, well people were intelligent and there was demand, but doctors were not doing so and so thing due to lack of essentially sales. I mean you don’t have to be generally intelligent to know that flossing is good for you, and yet: so few people floss! That would maybe not put tons of dentists out of business. But people are routinely doing (or not doing) professional services stuff not for any good (or bad) reason at all.
Clearly the thing going on in the professional services economy isn’t about general intelligence - there’s already lots of stuff that is NOT happening long before AI changes the game. It’s all cultural.
If you’ve gotten this far without knowing what I am talking about… listen, who knows what’s going to happen? Clearly a lot of behavior is not for any good reason.
How do you know where the ball is going to go for culture? Personally I think it’s a kind of arrogant position: “I’m a member of the guild, and from my POV, if my profession is replaced, so is everyone else’s.” Arrogance is not an attractive culture, it’s an adversarial one! And you could say inertia, and yet: look who’s running the HHS! There are kids right now, that I know in my real life, who look like you or me, who went to fancy Ivy League school, and they are vaccine skeptical. What about inertia and general intelligence then? So I’ll just say, you know, putting yourself out here on this forum, being all like, “I will AMA, I am the voice,” and then to be so arrogant: you are your own evidence for why maybe it won’t last 10 years.
I jumped into this thread to share my thoughts, and my thoughts alone, because I'm not sure there are a lot of radiologists on HN. I certainly don't speak for all radiologists.
But, I would submit to you, that rapid, radical changes to the practice of medicine are rare, if not impossible.
I don't think it will go away as long as we have third party paying for the costs and AMA controlling competition.
If I had to pay $500 or whatever to get a scan, and instead I could get my data, send it to a model and only follow up if it came back bad, I would. But now someone else pays and there are laws and regulations that prevent people from controlling their data, or at least make it difficult. Kind of weird I have a file on me that I have never seen.
Nope. You have an absolute legal right to obtain copies of your medical images and other data. Who paid for it is irrelevant. The provider can charge you a nominal fee for copying the files but they can't keep it from you.
Respectfully, it doesn't matter what you expect or think. What matters is this:
- If the law allows AI to replace you.
- If the hospital/company thinks [AI cost + AI-caused law suits] will be less expensive that [your salary + law suites caused by you].
I'm almost in the same situation as you are. I have 22 years left until retirement and I'm thinking I should change my career before I'm too old to do it.
The original author of the paper about the technological singularity [1] defines it as simply the point where predictions break down.
If AI gets to the point where it is truly replacing radiologists and programmers wholesale, it is difficult to tell anyone what to do about it today, because that's essentially on the other side of the singularity from here. Who knows what the answer will be?
(Ironically, the author of that paper, being also a science fiction author, is also responsible for morphing the singularity into "the rapture for nerds" in his own sci-fi writing. But I find the original paper's definition to have more utility in the current world.)
I don't think robotics is progressing at nearly the same pace as AI so for a while there will still be a bunch of manual labor for us to fight over. :-)
I really think we were doing things the “right way” before I left: providing AI analyses from various vendors as overlays for slides, and being able to pre-flag slides with obvious cases of cancer or other infection. (These analyses typically being custom algorithms provided by vendors through APIs we collaborate on, not LLM output.)
With things like this simply built into an existing, established pathology platform, I’m pretty confident a team of 4-5 pathologists could do the work of 6-8, with better quality output, similar to how Copilot and other tooling speeds us up as developers. At $300-400k/yr/doctor, that’s considerable savings (and the opportunity to allocate more doctors in specialties that aren’t easily automated).
However, for lots of reasons, it seems the market doesn’t necessarily agree with me on the value potential of this approach in this field (which can certainly be a self-fulfilling prophecy).
>the value potential of this approach in this field
As a medical school dropout, I think the largest "negative" in value assessment would be the inability to shift malpractice liability over to a human practitioner (i.e. can you insure a company's faulty AI/LLM?) — yet today I just read a HN article on insurers offering a new product: insuring chat/LLM fuck-ups.
As exciting as this all is, I'm still so glad I didn't become a radiologist!
I think I could have been happy with the physician lifestyle, but I knew I wouldn't make good life decisions surviving/getting through residency. I am glad to have survived (e.g. career change).
I've spent two decades as an electrician, and as the story goes: it is what it is... but also what you make of it. An incredible Good Fortune keeps me from having to work full-time, but as I've entered my forties I know I want to get in an air conditioned office (before too long).
Perhaps I'll first move to a country where healthcare isn't tied to full-time employment. Best of luck to you in your Interpretations.
> “People should stop training radiologists now,” Geoffrey Hinton said, adding that it was “just completely obvious” that within five years A.I. would outperform humans in that field.
I'm starting to notice a strong trend where non-domain experts will confidently assert that X field will be replaced by AI.
Hinton is a domain expert in AI, case in point Scott Aaronson cites AI revolutionizing protein-folding research, putting many protein-folding academics out of jobs.
I'd argue that AI experts and CS theorists are in a better position to speak about technological change than the domain experts who are trained/selected mainly to practice the status quo.
> putting many protein-folding academics out of jobs.
Has this actually happened? I'd wager it's more like in OP's article, AI as a helper that speeds up research, enabling bigger or different research questions. Reliable prediction not of protein shape, but of protein-protein or protein-mRNA interactions, for example.
Isn't this a common trend with CS? It seems like there's just constant over hype that drives booms and crashes, even to such a degree that if something was there it can't even be built. We have billionaires and hundred millionaires that made their money on VR, blockchain, and Web3. The wealth comes even when no product does and that seems like a big issue that's being exploited
> One pathology AI founder told me that it wasn’t hospitals or diagnostic labs that showed the most promise. It was R&D groups within Big Pharma. Those scientists and executives wanted new tooling. They were often sitting on massive internal datasets, had real budget allocated to experimental tech, and — critically — had a clear ROI if your model helped shave months off a study or more precisely target the right patient cohort. Most importantly, pharma didn’t care as much about the regulatory headaches, as they weren’t using your model to diagnose patients
Really nice article, thanks for sharing. With digital pathology ai startups going bust left and right it's quite an accurate analysis. One thing it's missing in the "product" path is preanalytical differences - same tissue processed in different labs can produce wildly different pixels.
The key to the power of the LLM is that the training process can learn effectively from vast corpora of unlabelled text. Unfortunately, there is no comparably vast database of medical images.
In order to "crack" radiology, the AI companies would need to launch an enormous data collection program involving thousands of hospitals across the world. Every time you got an MRI or X-Ray, you would sign some disclosure form that allowed your images to be anonymously submitted to the central data repository. This kind of project is very easy to describe, but very difficult to execute.
I agree with you, but here is where things get tricky:
Everyday I see something on a scan yhat I've never seen before. And, possibly, no one has ever seen before. There is tremendous variation in human anatomy and pathology.
So what do I do? I use general intelligence. I talk to the patient. I talk to the referring doctor. I compare with other studies, across modalities and time.
Another way of saying this that might not trigger your alarm bells is that even a perfect image analyzer is not enough to replace a radiologist. The job consists of much more than just analyzing images.
>Unfortunately, there is no comparably vast database of medical images.
Even a tiny hospital with radiology services will produce many thousands of images with accompanying descriptions every year. And you are allowed to anonymize and do research on these things in many places as neither image nor accompanying description is a personal identifier.
So this is yet another Hinton-ish prediction, any time soon radiologist are going dodo. This time LLMs will crack the nut that image recognition have failed at for 20 years.
Where LLMs have succeeded is in doing hot takes that miss the mark, they should be really good at cornering the "prematurely predicting demise of radiologist"-market
It's organised into regional NHS primary care and acute trusts, which usually share their records these days. It's still substantially more integrated than in many other countries.
AI is not easily going to "steal" any properly gatekept jobs that require a licensed practitioner.
Machine learning technology can be used to search for artifacts in medical diagnostic images that humans might miss. The flagged situations still have to be interpreted by a radiologist.
The thing I find most interesting about ML in radiology is that a computer can observe the entire dynamic range of the sensor at once. A human will only look at a window or a compressed window.
So are datasets/currently available data the limitation here?
Let's say a major healthcare leak occurred, involving millions of images and associated doctor notes, diagnostics, etc... would this help advance the field or is it some algorithmic issue?
People get tripped up by thinking "there is a subset of what I do that only humans can do and so that means AI will not eliminate my profession entirely and my job is safe".
Let's assume for now that it's true that AI can't do a certain subset of your work. Your profession won't be eliminated from the earth, that's true. But if 80% of your work can be done by AI, 80% of your work will be done by AI. There will still be humans kept around for that remaining 20%, but fewer of them will be needed.
The demand for radiologists is effectively infinite. Right now the healthcare system is supply constrained. If AI tools reduce the work of reading an image then more imaging studies will be ordered.
Also, many radiologists do interventional procedures directly with patients. We're a long, long way from being able to significantly automate that work.
So I recently needed an MRI to evaluate a rotator cuff injury. Because MRIs are expensive, insurnace required that I do 6 weeks of physical therapy first to see if that would resolve it.
Is this because the imaging process is so expensive, or because the radiologist who reads the image is so expensive?
If demand for an expensive imaging process increases, it will get even more expensive.
MRI machines and radiologist reads are both expensive. These are among the largest areas of medical spending for most health plans (and those costs are largely passed on to subscribers and employers).
Many rotator cuff injuries do improve to some extent with targeted physical therapy. You can accomplish a lot by strengthening the stabilizer muscles, and by learning how to set your shoulder joint in the correct position before taking a load. So the insurer's position is reasonable for most cases. If the injury is more severe then your doctor can submit a request to skip the normal step therapy process.
If there is one thing that NN can do amazingly well it is pattern recognition. The biggest blockers to adoption are the EHR providers, regulations and just inertia. It will happen, the only question is when and to a small degree, how.
EHR vendors aren't the major blocker here. Most of them have pretty good APIs now (at least for reading USCDI data from patient charts). But the data quality as entered by users is often crap: when you're training it's going to be garbage in / garbage out unless you put enormous efforts into cleansing. And the raw images are often not in the EHR but rather in a separate PACS.
From what I have seen trying to build out new features into this area, EHR vendors often act like monopolies blocking a lot of competition by controlling any data exchange. The vendor lock-in is strong and very harmful to improving process and, ultimately, harms patient care.
What specifically are they blocking? The most popular EHRs from vendors like Epic and Oracle have extensive interoperability APIs for multiple open industry standards. In general you can read all USCDI data pretty easily. Writing data is more limited, for understandable reasons. And if you want to distribute a SMART on FHIR app then you'll have to go through some sort of approval process.
The strong vendor lock-in is unavoidable given how deeply an EHR gets embedded into any provider organization. No matter what regulators and standards bodies do, it will never be easy to switch.
For small models, yes. But for the kind of massive multimodal models getting trained these days, the concept of "a pattern on an object" will exist. It doesn't need to have seen a tiger sofa before. But tiger + sofa + patterns on an object is enough to not run from the tiger print.
Reminder that according to https://openai.com/index/healthbench/, not only do current AI models provide far better recommendations than doctors, but even if doctors know the AI recommendations as a starting point, human revision does not bring any improvement.
Honestly this doesn't surprise me, considering the quality of the average doctor.
“Radiologists do far more than study images. They advise other doctors and surgeons, talk to patients, write reports and analyze medical records. After identifying a suspect cluster of tissue in an organ, they interpret what it might mean for an individual patient with a particular medical history, tapping years of experience.”
Now think about how much of software development is typing out the code vs talking to people, getting a clear definition of the problem, debugging, etc. (I would love an LLM that could debug problems in production — but all they can do is tell me stuff I already know). Then layer on that there are far more ideas for what should be built than you have time to actually build in every organization I’ve ever worked in.
I’m not worried about my job. I’m more worried my coworkers won’t realize what a great tool this is and my company will be left in the dust.
Fwiw I think this is not true right now. It might be true in the future, I guess. Today you have to learn some pretty specific techniques and they don’t come naturally to most engineers I think. They are accustomed to one-shot Google searches, not prompt engineering. Yes, you can learn how to prompt engineer, but it takes trial and error. A lot of my co-workers today still believe the LLM is useless because it hallucinates API definitions. They very much are not used to tools that require the level of hand-holding you need today.
Either teaching them is going to get easier, which is somewhat plausible actually — or you think LLMs are going to get dramatically easier to prompt with more reliable output. I actually think that’s somewhat implausible. Log(compute) is a pretty hard number to move past a certain point and it seems we’ve hit that point already. RLHF is great, but it doesn’t scale well and the result is that some tools are hard to build. Giving me millions of context tokens is neat, but if it’s wildly expensive and confuses the model on top of that, it’s not actually useful.
I do think there is room to improve and of course there are new techniques to be discovered. The models may even help discover some of them, although I am skeptical that they will find Everest by hill climbing when they’re currently on Pike’s Peak. Or maybe we really need to find Olympus Mons.
This reminds me of the idea that Human-Chess partnerships would be the ultimate manifestation of Chess genius. I'm not sure whether the idea is still holding on but engines are so far ahead of human play that I doubt a human in the loop can add anything these days given how devastatingly far ahead the engines are and the advent of machine learning techniques.
Chess reminds me more of programming given the set of defined rules in each. However, I'm biased as I work in radiology and program more as a hobby. So far I've seen way more tools to help me code than to accurately detect radiologic findings.
You can compare best image synthesis and image understanding from two decades ago (SIFT / HOG), from a decade ago (CNN, SdA) and now (Transformer). Very rapid progress that went from being able to unreliability recognize a face to getting to outperforming human professionals (see MMMU) is quite remarkable.
AIUI, and I may be wrong, but each of the mentioned technologies was a "breakthrough" technology - not iterative improvement. Along this vein, I was wondering if there was some promising, novel research OP was aware of for image understanding.
The most recent of which you mentioned, Transformers, is used by both LLMs and image synthesis/understanding. The parent posits that while computer vision lags behind NLP, this may not continue. While your comment points out that image synthesis and understanding has improved over time, I'm not sure I follow the argument that it may soon leapfrog or even catch up with LLMs (i.e. text understanding and synthesis.)
My understanding is that the way that chess.com and other online services detect cheating is by comparing the human-made moves to a "perfect" version of what the chess engine would play.
Which gives credence to your theory that people aren't bringing much to the table.
Glad to hear that this is one area that AI is conclusively useful.
Still not clear that the already superhuman capabilities of AI won't still fully supplement radiologist interpretive skills with every additional bit of training data that comes in.
I just want to point out that the term "A.I." gets used pretty loosely in these articles, as if A.I. is a monolithic commodity that you plugin to your software to make it do chatGPT.
The example in the article is an in house developed "A.I." to help radiologists assess images. Digging a bit deeper it seems they are using mostly old CNN type architectures with a few million parameters.[1]
I think it still remains to be seen what a 1T+ parameter transformer trained specifically for radiology will do. I think anyone would be confident that a locally run CNN will not hold a candle to it.
[1]https://mayo-radiology-informatics-lab.github.io/MIDeL/index...
There are many, many papers and projects out there about tuning foundation models on various types of medical imaging data, and many organized efforts to produce large medical imaging datasets to feed that training. This stuff is well-known in the trenches and can improve on the older, smaller CNNs in some ways, but not in a way that's produced any step change in automated capabilities. People are certainly working on it!
> I think it still remains to be seen what a 1T+ parameter transformer trained specifically for radiology will do
Does image processing of something like this scale with parameters?
It makes sense that language continues to scale as the vector space is huge. Even models that generate images scale as the problem space is so large.
But here there are only so many pixels in the image and they are a lot more uniform. You likely can't have 1T images to train on, so wouldn't the model just overfit and basically memorize every image it has ever seen?
You are right, but I meant it more as a total radiologist, scans the images and can put it into medical context.
Unfortunately, VLMs don’t seem to work. Even the best ones (eg, Gemini 2.5 Pro) hallucinate to a laughable degree. I’ve been seeing them make up large-scale features in images that absolutely no human would ever mistake.
What does work is narrow AI trained to do very specific tasks (such as medical image segmentation). But I don’t think that type of AI will benefit much from massive scaling.
If the meat machine can learn it, then the silicon machine can learn it.
Medical records remain a walled garden mostly because of HIPAA. Unlike all the AI development that has managed to skirt the copyright law to train large models, training a 1T+ parameter transformer on a large enough dataset will need a lot of consumers to give up their medical records.
UK's doing it:
>Concerns raised over AI trained on 57 million NHS medical records
https://www.newscientist.com/article/2479302-concerns-raised...
Underreported competitive advantage of NHS is their unified and comprehensive data
China won't have an issue with it.
The US won't have an issue with it. You haven't realized what the US has become.
A place where it's still possible to innovate, despite the best efforts of the copyright lobby?
I've noticed a funny trend where people will bring up an ethical concern and some dweeb will immediately roll out the Chinese strawman like it justifies their incredibly dubious position.
"We should violate medical privacy laws on an enormous scale because China" Okay buddy.
I wasn't suggesting that we should, but in my lifetime we went from hospital admissions and discharges being published in the local paper, to that information being some of the most closely guarded with (theoretically) severe penalties for disclosure.
I think it was AIDS and people not wanting to be outed that changed things.
But whatever the reason, we now have HIPAA and it adds a huge amount of complexity to the management of medical records, and a huge amount of complexity to any research that needs access to a broad spectrum of medical case histories (such as training AI). Other countries don't have these concerns, and will outcompete the US on developing these capabilities.
Concerns about HIPAA as an obstacle to research are largely a red herring. Almost no clinically useful research can be accomplished by just processing large volumes of historical patient charts. In practice the researchers will have to gather additional data from study subjects so they're going to need to obtain consent anyway.
People keep claiming that other countries will out compete us in medical research and it keeps not happening.
yes but you can train relatively dumb AI on high volumes of historical data and presenting factors. This is typically just a 'diagnostic assist', but often significantly better than a human
Better how? Do you have a citation for that?
I did work on a diagnostic assist tool developed by a large pharma which diagnosed between Asthma and COPD. GPs get this right 52% time, specialists just over 60 and AI came in over 80%...using 12 relatively mundane input variables. I believe there are a lot of these situations, but not clear pathway to FDA approval
Is that worse than being worse off technologically than China and the consequences that bears
Do you really feel life has been horrible in all these countries which have been worse off technologically than the US for the past decades?
Being first is not the same as being happy.
If we become worse off than China, it's because we squandered a massive decades long lead and actively moved critical industrial processes over there, not because we didn't let AI companies scrape HIPAA data.
Because the end result is either paying China to eval scans, buy models from China that do it, or settle for inferior scan reading by humans.
The Red Scare never ended for a whole lot of Americans. And my foot is more communist than modern China is. But you know, being afraid of Communism is just as, if not in fact easier, when you have no clue what Communism actually is.
If copyright stayed at 7+7, I would agree that AI companies were "skirting" copyright law, but as I see it life + 70 violates the "limited" terminology in the Constitution by denying people the right to create derivative works.
IP holders did this to themselves by extending the term beyond any reasonable interpretation of "limited".
Hell, Jack Valenti wanted to extend it to "forever minus a day"
IMO, copyright law in its current form is unconstitutional and unnatural.
Epic probably has enough breadth on their own to do it, although they don’t actually have an image display system that I know of, so perhaps they don’t have access to the raw images - only the written reports.
Right, some of that data is in Epic Cosmos.
https://cosmos.epic.com/
Epic does not have the ability to use the data as they see fit, even if they have the data available. That is why I brought up HIPAA.
When it comes to test results, does HIPAA prevent anonymization, such that all would be provided is test input & eventual result?
From a summary on HHS.gov it says "De-Identified Health Information. There are no restrictions on the use or disclosure of de-identified health information." Maybe someone with more knowledge could expand on the limitations of what counts as "De-Identified" but I think that might work. I followed the reference and nothing in the CFR jumps out at me, but I'm not a lawyer so who knows.
HHS has published clear guidelines on de-identification.
https://www.hhs.gov/hipaa/for-professionals/special-topics/d...
"The radiologist will be with you soon" also gets used pretty loosely in this article, since a radiologist is not someone you are likely to speak with. They sit in their home office in Montana clicking through images all day, and send thru their diagnosis to a different practitioner at the office you are visiting.
Which means the field has a larger issue since they can just as easily sit in the Philippines or Brazil or India.
There's a reason they can't.
Why not? Where I work scans are reported overseas. The radiologists are qualified in the country they are reporting for however.
Ok, I should have been more clear.
A radiologist practicing in the USA can live anywhere in the world, but they must be trained in the USA, licensed in the state(s) where they read from, credentialed at the hospitals or facilities they read from, and carry medical malpractice insurance.
The point being, hospitals can't throw cheap Indian or Filipino radiologists at the worklist.
Can a certified radiologist farm out work to cheap assistants in other countries, and then put their stamp on the result?
Not sure about other countries, but RAs do outsource work to RRAs(1) and then sign the final product. This is relatively new as Medicare/Medicaid didn't recognize RRAs as mid-level practitioners until 2019. I suspect the answer is yes for ARRT-certified RRAs, but it will be state-specific (similar to the answer above about RAs).
(1)https://www.arrt.org/pages/earn-arrt-credentials/credential-...
This is illegal, and you will go to jail if you do this with a Medicare or Medicaid patient and it is discovered.
Why? If they agree with it, is it that different to a software generate report that they sign?
So what you're saying is they can, they just don't at the moment for temporary, localized legal reasons.
For a while it seemed that "AI" topics were treated as niche and news articles used relatively specific terms of art like deep learning or machine learning. Somehow ChatGPT changed that and now everything is just AI.
ChatGPT changed that because it, along with other current advances, brought AI out of its niche and into widespread mainstream usage. The laboratory the source article covers (https://www.mayo.edu/research/labs/radiology-informatics/ove...) publishes some of the technical detail you're interested in, but in 2025 it's also interesting to people less familiar with theoretical computer science. The doctors quoted in the source article, using programs that can calculate kidney volume or detect blood clots, don't necessarily want or need to know the difference between "artificial intelligence" and "deep learning".
I do my best to filter ai topics here using ublock, and it's blocked about a third of the posts, and yet every day there's a new variation, whether a new technique/technology, a new product (open source or commercial), etc... that I have to add to the filter. It's really offputting. Yet apparently this is the only site besides lobste.rs that has programming related content that is of any depth that I know of. I really wish hacker news had the ability to tag posts and let me filter (and tag them generally not have to get specific on each new buzzword or product).
Why do you want to read about tech, but specially non-AI tech?
I'll speak for myself, for me it's illustrated by the spam sketch from Monty Python. It crowds out everything, including my mental space. The hype and disconnect from reality make it intriguing to my brain so I am very likely to end up reading the comments about it, but it's usually a waste of my time. Software does what it does and my opinions (or anyone else's) just don't matter. I should focus my time on learning and building instead of trying to understand hype.
Agreed
Working radiologist here, 20 years experience.
This article is surprisingly accurate. I fully expect to finish my career without being 'replaced' by AI.
Happy to debate/answer questions :-)
I would say generally speaking that people who assume AI will replace somebody else's job believe that these jobs are merely mechanical and there is no high-level reasoning involved that would basically require AGI (when that comes about nobody is safe). So the model of the AI radiologist assumes the only job of a radiologist is to classify images, which is pretty vulnerable to near-future disruption.
I imagine, given the training involved, the job involves more than just looking at pictures? This is what I would like to see explained.
The analogy would be the "95% of code is written by AI" stat that gets trotted out, replacing code with image evaluation. Yes AI will write the code but someone has to tell the AI what to write which is the tricky part.
>> jobs are merely mechanical and there is no high-level reasoning involved
This is a very binary way of thinking about it. More usual is that components of many professions are mechanical and can be automated, while other components are not mechanical and thus harder to automate. Regardless, if some % of the mechanical work goes away, it is unlikely that human workers just work less. Instead, they will work just as much and the overall demand for workers is reduced by %
Between 1985 and 2025 we went from programming in 8086 assembly language to high level languages like python, typescript and go. These automate a lot of the drudgery of programming in asm, so why has the overall demand for programmers not diminish (in fact it increased massively)?
We already have AI taxis (in specific limited areas, but still). Driving isn't something I'd usually call "merely mechanical".
Can you describe a driving scenario where the correct action couldn't be determined "mechanically"? Are you thinking of something like the trolley problem?
Driving (in US) is considered unskilled labor.
That is such a contrived phenomenon, it has taken decades of lobbying and destruction of political accountability to create the conditions where a person considered sane would touch that idea instead of immediately skipping over to driverless trains.
Incredibly wasteful gimmick, I don't get why the usians are still struggling away at it now that the chinese seem to have already done it.
I don't know where you've gotten the idea that cars are something you can "skip over" on the way to trains. Public transit is great, and the US should do it better, but it doesn't obsolete every other form of transit. The vast majority of people in every country who can afford access to cars use them regularly.
> The vast majority of people in every country who can afford access to cars use them regularly.
Well... that's just not true.
Why are you so sure about that?
100%
If (as acknowledged in the article) AI automates at least part of the work of radiologists (e.g. tool that "saves her 15 to 30 minutes each time she examines a kidney image"), don't you fear that the demand of radiologists will decline? Even if some are still needed, surely if a hospital needs X reports per day and now Y radiologists are sufficient to provide them rather than the current Z (Y<Z), that should be something for people considering your career to take into account?
On the other hand, how much of your confidence in not being replaced stems from AI not being able to do the work, and how much from legal/societal issues (a human needing to be legally responsible for the diagnoses)? Honestly the description in the article of what a radiologist does "Radiologists do far more than study images. They advise other doctors and surgeons, talk to patients, write reports and analyze medical records. After identifying a suspect cluster of tissue in an organ, they interpret what it might mean for an individual patient with a particular medical history, tapping years of experience" doesn't strike me as anything impossible for AI within a few years, now that models are multimodal and they can work with increasing amounts of text (e.g. medical histories).
No. There is no area of medicine where a boost in productivity will cause doctors to have idle time. The wait times may decrease, throughput may increase, diagnosis accuracy may improve, even costs may decrease (press x to doubt) but no, there will never be a case where we will need less radiologists.
Which may take us to a sort of “Jevens Paradox” kind of place except for medical care.
Like there are times already where I’ve put off or not sought medical care because of the hassle involved.
If I could just waltz into the office and get an appointment and have an issue seen to same day I would probably do it more often.
Wonderful insight that I'd never considered. Talk to almost anyone in America and they'll tell you about a health issue that they or their family are deferring due to lack of access. Waiting months or years to simply talk to a specialist, let alone find one that can help, is sadly the norm. Patients rightly feel it's a waste of their time so won't even seek treatment.
Remove that barrier to access and we won't see a shiny new streamlined medical system but rather a flood of new patients requiring even more bureaucracy to manage.
100% why I bothered to stick w/ OneMedical after the Amazon acquisition. I tried to get a "New Dr/New Patient" appointment with any local doctor I could find that took my insurance, and the closest I could get was 3 weeks away.
I'm not worried about medical folks having "time to clean" any time soon.
To be clear, we also need more medical professionals in general -- they're not keeping up with the population and it's making us all less healthy. Three to six months, or more, in the SF bay area for some critical appointments is really unacceptable, but there's not really an option given supply and demand.
I'm sure this will all get better with captain brainworms at the helm.
>> No. There is no area of medicine where a boost in productivity will cause doctors to have idle time. The wait times may decrease, throughput may increase, diagnosis accuracy may improve, even costs may decrease (press x to doubt) but no, there will never be a case where we will need less radiologists.
I dont think this is how market participants may think about it. If costs decrease, some group of radiologists will drop out of the market. We may not "need" less radiologists, but we're signaling we need less of them by not paying them as much as before.
Much like I still "need" a photographer, but short of weddings, I'm not willing to pay as much as before. I may well acquire a photogrpher for a portrait, but it would have to be priced competitively to a selfie.
You're making the common mistake of thinking that healthcare is some sort of normal market subject to simple supply-and-demand economics rules. The reality that supply is heavily constrained by CMS funding for residency slots, prices are (mostly) fixed by a few large payers, and patient demand is effectively infinite. There is a serious shortage of radiologists already and it's getting worse as the population grows older and sicker. If AI tools make radiologists more productive then more imaging studies will be ordered.
>> supply is heavily constrained by CMS funding for residency slots
I keep hearing this argument. Then I look at an Insurance Explanation of Benefits statement. A Radiologist might make $1-2k/day in billings. If you are in a balance-billing state, whatever insurance doesnt pay gets forward billed to the patient. On a standard 252 day workyear, that is $250-500k/yr in billings. The average resident salary is 70k, lets assume 100k with benefits.
Of course there is plenty of overhead, but from the math i'm seeing, the average Radiology Resident is a 150-400k net revenue center. Is the overhead really greater than 150-400k/yr/resident?
What am I missing, why would a profit center need "CMS funding"? From what my doctor friends tell me, the real bottleneck is the unwillingness of AMA and ABR to open up more radiology residency spots (artificial supply constraint) with "CMS funding" a boogeyman and red herring.
Contrary to what your doctor friends might have told you, the AMA has no power to restrict the number of residency slots. They are a private membership organization with no regulatory or accreditation authority. At one time they did lobby Congress to restrict graduate medical education funding but have since reversed that position.
https://savegme.org/
A minority of funding does come from sources other than CMS. Teaching hospitals are largely free to add more residency slots if they want to. The fact that most hospitals don't do this indicates that GME programs are largely unprofitable.
You cannot bill for a resident's work, without an attending signing off on it.
I like residents, I actually started a new radiology residency program.
They are a net drain on productivity in radiology, due to the teaching obligations
>> You cannot bill for a resident's work, without an attending signing off on it.
Isnt that the case for almost every industry? I was a management consultant and our Partner signed off on the work before billing clients. Architects bill clients, but only when the lead architect signs off on the plans. Same for civil engineers. Same for magazines where the editor signs off on most major pieces.
Faster CPUs, better screens, helpful IDEs, heck even Gen-AI itself did not reduce the need for software engineers let alone decrease the costs. As mentioned in another comment, The Jevons Paradox implies that in certain industries, increased productivity may actually lead to more consumption (therefore propping up the demand / cost) despite not being intuitive so.
The only industries that has observed the opposite effect I can think of so far are translators and stock photographers. Maybe also proof readers - but is that gen Ai or did spellcheckers already kill that branch?
Software engineers could spread out into new areas and find new things to do. Unless I am mistaken, the work of radiologists is more defined.
> Unless I am mistaken, the work of radiologists is more defined.
When I had cancer I had to go to an Interventional Radiologist. Never heard of it before. But they use X-Ray in real-time (flouroscopy?) to guide surgery. Pretty neat.
> Software engineers could spread
This is key. Radiologists that can program (or software engineers that do medicine) make a massive difference.
When someone explains what they want and it’s built for them, you get a very different product to that built by the user.
I want more radiologists that can write software please.
I'm a radiologist that can program!
I've authored and contributed to several open source projects over the years, and I'm currently doing a deep dive in CAD/CAM after buying a CNC machine.
I help my practice where I can, and have written a little utility to make generating reports easier, but I would have to quit my job if I was trying to take on the absolutely enormous task of radiology computer aides diagnosis. And I need my job!
> I would have to quit my job if I was trying to take on the absolutely enormous task of radiology computer aides diagnosis
I think about this sort of thing every so often. The difference one person can make by creating a decent piece of software might be one of the best ways of getting a return on investment (in human hours, if not in money).
>> The only industries that has observed the opposite effect I can think of so far are translators and stock photographers. Maybe also proof readers - but is that gen Ai or did spellcheckers already kill that branch?
Voice-over performers. Sketch artists. Audiobook actors. Tutors (see: https://www.google.com/finance/quote/CHGG:NYSE?window=5Y)
if the cost for preventative scans goes down, demand will rise. medical demand is incredibly constrained by price. people skip all kinds of tests they need because they can't afford it. the radiologists will have more work to do, not less.
There is a national shortage of radiologists in the US, with many hospital systems facing a backlog of unread cases measuring in the thousands. And, the baby boomers are starting to retire, it's only going to get worse. We aren't training enough new radiologists, which is a different discussion.
Askl to your question on where my confidence stems from, there are both legal reasons and 'not being able to do the work' reasons.
Legal is easy, the most powerful lobby in most states are trial attorneys. They simply won't allow a situation where liability cannot be attached to medical practice. Somebody is getting sued.
As to what I do day to day, I don't think I'm just matching patterns. I believe what I do takes general intelligence. Therefore, when AI can do my job, it can do everyone else's job as well.
> We aren't training enough new radiologists, which is a different discussion.
About that, I think the AMA is ultimately going to be a victim of its own success. It achieved its goal of creating a shortage of medical professionals and enriching the existing ones. I don't think any of their careers are in danger.
However, long term, I think magic (in the form of sufficiently advanced technology) is going to become cost effective at the prices that the health care market is operating at. First the medical professionals will become wholly dependent on it, then everyone will ask why we need to pay these people eye-watering sums of money to ask the computers questions when we can do that ourselves, for free.
I agree with you on all points. The only question is how long will it take?
The trial lawyer angle doesn't seem accurate. Did trial lawyers prevent pregnancy tests from rolling out? COVID tests? Or any other automatic diagnostic, as long as it was reasonably accurate?
Not as far as I know. Once an automated diagnostic is reasonably accurate, it replaces humans doing the work manually. The same would be true of anything else that can be automatically detected.
No comment on whether radiology is close to that yet, although I don't think a few-million-param neural network would tell us much one way or another.
Are you aware of any states in the US that have made it harder to sue doctors for malpractice?
My point, which I made poorly, is this: There's a reason doctors that went to medical school in India and trained as Radiologists in India can't read US cases remotely for a fraction of the cost of US trained and licensed radiologists.
It's not because the systems to read remotely don't exist.
It's not because they're poorly trained or bad doctors.
Itsty because they can't be sued.
No, it's because the AMA lobbies to protect American doctors' jobs, and refuses to license them to practice in the US. Of course you can still sue for medical malpractice regardless of citizenship. Trial lawyers have nothing to do with it, American doctors who don't want competition are entirely to blame.
A big wrinkle in AI evangelism is that proponents don’t understand the concept of human judgment as a “learned” skill - it takes practice and AI models / systems do not suffer consequences the way humans do. They have no emotions and therefore can not “understand” the implications of their decisions.
For context, generative AI music is basically unlistenable. I’ve yet to come across a convincing song, let alone 30 seconds worth of viable material. The AI tools can help musicians in their workflow, but they have no concept of human emotion or expression and it shows. Interpreting a radiology problem is more like an art form than a jigsaw puzzle, otherwise it would’ve been automated long ago (like a simple blood test). Like you note, the legal system in the US prides itself on “accountability” (said tongue in cheek) and AI suffers no consequences.
Just look how well AI worked in the United Healthcare deployment involving medical care and money. Hint: stock is still falling.
>For context, generative AI music is basically unlistenable. I’ve yet to come across a convincing song, let alone 30 seconds worth of viable material.
This one pops into my head every couple months:
https://youtube.com/watch?v=4gYStWmO1jQ
It's not really my genre, so my judgment is perhaps clouded. Also, I find the dumb lyrics entertaining and they were probably written by a human (though obviously an AI could be prompted to do just as well). I am a fan of unique character in vocals and I love that it pronounces "A-R-A" as "ah-ahr-ah", but the little bridge at 1:40 does nothing for me.
You may have missed the month or so where this[1] AI-generated track (remixed by a person, but nonetheless) dominated pop culture.
[1] https://www.youtube.com/watch?v=1uW_AUwEv-0
> A big wrinkle in AI evangelism is that proponents don’t understand the concept of human judgment as a “learned” skill
Which is ironic given how much variation in output quality there is based on the judgement of the person using the LLM (work scope, domain, prompt quality, etc.)
You say "Legal is easy, the most powerful lobby in most states are trial attorneys."
The most powerful lobby in this case is the ABR which carefully constricts coveted residency spots in Radiology to create an artificial scarcity and keep up incomes. It is the opposite of, say, technology, where we have no gatekeeper and supply increases.
The ABR will say that Medicare doesnt fund enough residency spots, but all you need to do is look at an EoB and see that a week of residency billings covers the entire cost of the resident.
IMHO the ABR isn't quite as powerful as you're indicating, but in general I agree.
For what it's worth, I started a new residency program to train more radiologist, so I do have some skin in the game.
If a teaching hospital with an existing radiology residency program wanted to add one more spot, does the ABR have any power to stop them? If Medicare offered more funding to a teaching hospital to add more spots would the hospital turn it down?
Basically, the ABR needs to approve a residency program for the residents that graduate from the program to be eligible to sit for the boards.
I don't think they would ever do it, but technically the ABR could stop them.
There is a perpetual shortage of these types of technicians, so it is unlikely that demand for those jobs will drop.
I worked on an autocontouring model but we could not get very high accuracy for it to be adopted commercially. The algorithm would work for some organs but would totally freak-out on the others. And if the anatomy was out of norm then it would not work at all. This was 5 years ago, I see Siemens [0] has a similar tool. I remember shadowing a dosimetrist contouring all the Organs-At-Risk (OAR) and it took about 3-4 hours to contour one CT image of thoracic region. Do you know how much better the autocontouring tools have become?
[0] https://www.siemens-healthineers.com/en-us/radiotherapy/soft...
Segmenting algorithms is massively improved from 5 years ago. How much the actual SOTA is used on the work floor though is always a hit and miss.
Agreed, and from what I understand radiation oncology treatment planning has benefited from this.
As my wife says: "Until it's as easy to sue AI as it is doctors, we probably won't see AI replacing doctors."
May be in the West. However more practical countries like China with a huge population and clear incentive to provide healthcare to a large population at reduced cost will have incentives to balance accuracy and usefulness in a better way.
My personal opinion is that a lot of Medical professionals are simply gatekeeping at this point of time and using legal definitions to keep changing goalposts.
However this is a theme that will keep on repeating in all domains and I do feel that gradual change is better than sudden, disruptive change.
AI in healthcare is going to add so many layers of indirection for malpractice lawsuits. You'll spend years and lots of $$$ trying to figure out who the defendant would ultimately be, only for it to end up being a LLC that unfortunately just filed for bankruptcy.
The worry isn't that you'll find an AI sitting on the chair that a radiologist used to sit. It's that the entire field of radiology gets reduced down to a button click on a software.
The other doctors will still be there for you to sue.
What if ppl just bought the equipment and did the scans at home?
So the question is, “what if people bought an x-ray machine (affordably available on Amazon)and started using it without training on radiological safety”?
I assume followed shortly by "what is this weird red splotch on my skin?"
Have you priced out a CT scanner and MRI?
Will you be able to source a radioactive source for your x-rays?
The X-Ray source is the X-ray machine. You may be referring to nuclear medicine which injects radioactive stuff, or radiation therapy.
DIY radiation therapy would be a whole new level.
> The X-Ray source is the X-ray machine.
Healthcare-grade x-ray tubes to put in your (expensive) x-ray machine are not something you easily can obtain without a license.
Here in New Zealand you get a licence after purchasing the equipment, and require the machine spec, date of manufacture and serial number to get the licence.
It does require a radiologist name on the paperwork as they are the one with the radiation licence. However it is possible to get one if not a radiologist (dentists do, and radiographers have).
Being licensed to use the equipment is the hard bit, as insurance companies require accreditation which is hard to get.
Fair. But what if ppl instead got scans in "radiology shops" without waiting for a specialist? Specialists are expensive.
Isn't that 90% of going to get scan is right now? You'll still need the "shop" to provide the equipment and the tech with the training to know what/where to scan, but you might get the results a bit faster? Are the radiologists the chokepoint now, or is it the techs?
That's the way it already works in many cases, just like with outpatient surgery clinics and other outpatient specialist practices. There is a critical difference, though, because radiology also has sub-specialities and someone focused on orthopedics probably isn't the one you'd want reading your cardiology images, nor would you want your ophthalmologic radiologist trying to diagnose a brain CT.
Medical Imaging tech entrepreneur here.
Been going to RSNA for longer than you've been a radiologist. In all that time, I've never come across an AI that I felt was fit for purpose.
I wholeheartedly agree with you.
Many many reasons for this, and I'm happy to chime in from the tech side of things and fill in any blanks outside your knowledge domain.
> Medical Imaging tech entrepreneur here
I initially read this as ‘Medical Imaging tech who is an entrepreneur.’ I now think you’re a radiologist?
Any particular interests?
Fixing the shiite RIS/PACS world and the hell of hl7 would make me happy. I’m an MR tech, and just finished trying to make a scan description ‘MRI Cervical Spine + Right Brachial Plexis’
You can’t. Description too long.
Basically, they are all currently hot garbage.
In my practice we have:
1. Mammography CAD
2. Pulmonary nodule CAD
3. Intracranial hemorrhage CAD
4. Intracranial vessel CAD
They all stink! I would love it if they were useful, radiology is insanely difficult and I could use the help!
(great username: a radiologist with "seesthruya")
Are AI models able to detect abnormalities that even an experienced radiologist can't see? i.e. something that would look normal to a human eye but AI correctly flags it for investigation? Or are all AI detections 'obvious' to human eyes and simply a confirmation? I suspect the latter since it was human annotated images the model was trained on.
Depends on what you mean by 'see'.
For example, let's say I'm looking at a chest x-ray. There is a pneumonia at the left lung base and I am clever enough to notice it. 'Aha', I think, congratulating myself at making the diagnosis and figuring out why the patient is short of breath.
But, in this example, I stop looking closely at the X-ray after noticing the pneumonia, so I miss a pneumothorax at the right lung apex.
I have made a mistake radiologists call 'satisfaction of search'.
My 'search' for the patient's problem was 'satisfied' by finding the pneumonia, and because I am human and therefore fundamentally flawed, I stopped looking for a second clinically relevant diagnosis.
An AI module that detects a pneumothorax is not prone to this type of error. So it sees something I did not. But it doesn't see something that I can't see. I just didn't look.
This is definitely a thing.
https://www.npr.org/sections/health-shots/2013/02/11/1714096...
I'm skeptical to the claim that AI isn't prone to this sort of error, though. AI loves the easy answer.
AI is overloaded. An LLM loves the easy answer, but that's not what is underlying an image classification model.
> I have made a mistake radiologists call 'satisfaction of search'.
Ah, now I have a name for it.
When I've chased a bug and fixed a problem I found that would cause the observed problem behavior, but haven't yet proven the behavior is corrected, I'm always careful to specify that "I fixed a problem, but I don't know if I fixed the problem". Seems similar: found and fixed a bug that could explain the issue, but that doesn't mean there's not another one that, independently, would also cause the same observed problem.
It's also called inattentional blindness.
https://en.wikipedia.org/wiki/Inattentional_blindness
I've been going to RSNA for over 25 years, in all that time, the best I've seen from any model presented to me was the smack the radiologist on the head and say, "you dummy, you should have seen that!" model.
That is, the models spot pathologies that 99.9999% of rads would spot anyway if not overworked, tired, or in a hurry. But, addressing the implication of your question, the value is actually in spotting a pathology that 99.9999% of rads would never spot. In all my years developing medical imaging startups and software, I've never seen it happen.
I don't expect to see it in my lifetime.
I'm sure it's a matter of training data, but I don't know if it's a surmountable problem. How do you get enough training data for the machine to learn and reliably catch those exceptions?
I have a fairly strong background in tech, and I've been programming computers since 1979 when my dad bought me a TRS-80. Tape drives FTW!
I agree with almost everything you've said here.
Except 'not in my lifetime', because I plan on living for a very long time, and who knows what those computer nerds will come up with eventually ;-)
One Harvard study trained an AI that could reliably determine someone's race from a chest X-ray. AIs can be trained to see things we can't.
The difficulty is likely in making a good training dataset of labeled images with pathologies radiologists couldn't see. I imagine in some cases (like cancer), you may happen to have an earlier CT scan or X-ray from the patient where the pathology is not quite yet detectable to the human eye.
I suspect that radiologists could identify race from a plain chest x-ray if they were given the patient’s race and asked to start noticing the difference. They just aren’t doing it because, if it’s important, you can just look at the patient.
There are a lot of things in medicine that aren’t in literature, but are well-known among certain practitioners. I’m an anesthesiologist and practice in an area with a large African-American population. About 10-15% (rough guess) of people of West African descent will have a ridiculously strong salivary response to certain drugs (muscarinic agonists). As in, after one dose their mouths will be full of saliva in seconds. We don’t have East Africans for comparison, so I can’t say it’s a pan-Bantu thing, but I have seen it in a Nigerian who lived here. Not in the literature, but we all know it. I had a (EDIT: non-anesthesia) colleague ask me about a hypersecretory response from such a drug. I said, oh, was he black? Yes, how did I know? Because we give those drugs all the time and have eyes. It’s very rare to see in European-descended populations.
It's possible humans could learn to do this, but I'm skeptical they could do it this well. According to the article, humans experts couldn't tell race from the chest X-rays and the researchers couldn't figure out how the AI was detecting race. They fed it corrupted images to figure out what information it was relying on. It was robust against both low pass and high pass filters.
Sigh.
Rads can see a person's race. We look at them.
That's the reason rads never train to determine race from a chest x-ray.
BTW, models don't need to train that either. Because if it's important, it's recorded, along with a picture, in the guy's medical record.
I'd just like to gently suggest that determining someone's race from an X-Ray instead of, say, their photograph, is maybe not how we should be burning training cycles if we want to push medical imaging forward. Human radiologists had that figured out ages ago.
You're being snide about the Harvard/MIT researchers being idiots doing useless research because they don't realize radiologists can just look at the patients face, but that's obviously not what happened. They were trying to see if AI could introduce racial bias. They're not releasing an AI to tell radiologists the race of their patients.
According to the article, human experts could not tell race from chest X-rays, while the AI could to do reliably. Further, it could still determine race when given an X-ray image passed through a low pass filter or high pass filter, showing that it's not relying solely on fine detail or large features to do so.
>> Happy to debate/answer questions :-)
Curious -- do you think that is because
1. the technology isnt there, or
2. because it isnt a competitive market (basically, the American Board of Radiologists controls standards of practice and can slow down technologies seen as competitive to human doctors)?
3. or perhaps 1 doesnt happen because outsiders know the market is guarded by the market itsself?
What's accurate in this article? It's very vague, it can be tldred into "we won't go anywhere, although AI does more and more of our work"
> Radiologists do far more than study images. They advise other doctors and surgeons, talk to patients, write reports and analyze medical records. After identifying a suspect cluster of tissue in an organ, they interpret what it might mean for an individual patient with a particular medical history, tapping years of experience.
AI will do that more efficiently, and probably already does. "tapping years of experience" is just data in training set.
> A.I. can also automatically identify images showing the highest probability of an abnormal growth, essentially telling the radiologist, “Look here first.” Another program scans images for blood clots in the heart or lungs, even when the medical focus may be elsewhere. > “A.I. is everywhere in our workflow now,” Dr. Baffour said. > “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.”
Maybe you'll be able to happily retire because inertia, but overall it looks like elevator operator job.
What's so special about radiology?
There's nothing special about radiology. And I do believe inertia will carry me through the end of my career, which has approximately 10 years left.
However, it's my opinion that my job takes general intelligence, not just pattern matching.
Therefore, when I lose my job to AI, so does everyone else.
>> Therefore, when I lose my job to AI, so does everyone else.
Not quite right? Some fields are licensed, regulated, and have appointments -- and others are not. AI is most keenly focused on fields w/o licensure barriers
You are correct!
I should have said when an AI can do my job it can do anyone's job.
On the one hand, you’re totally right. The job takes general intelligence.
On the other hand, a lot of jobs take general intelligence. You’re right about that too.
It’s difficult to guess the specifics of your life, but: maybe you’ve engaged a real estate agent. Some people use no real estate agent. Some have a robo agent. No AI involved. Maybe you have written a will. Some people go online and spend $500 on templates from Trust & Will, others spend $3,000 on a lawyer to fill in the templates for them, some don’t do any of that at all. Even in medicine, you know, a pharma rep has to go and convince someone to add their thing to the guidelines, and you can look back at the time between the study and adoption as, well people were intelligent and there was demand, but doctors were not doing so and so thing due to lack of essentially sales. I mean you don’t have to be generally intelligent to know that flossing is good for you, and yet: so few people floss! That would maybe not put tons of dentists out of business. But people are routinely doing (or not doing) professional services stuff not for any good (or bad) reason at all.
Clearly the thing going on in the professional services economy isn’t about general intelligence - there’s already lots of stuff that is NOT happening long before AI changes the game. It’s all cultural.
If you’ve gotten this far without knowing what I am talking about… listen, who knows what’s going to happen? Clearly a lot of behavior is not for any good reason.
How do you know where the ball is going to go for culture? Personally I think it’s a kind of arrogant position: “I’m a member of the guild, and from my POV, if my profession is replaced, so is everyone else’s.” Arrogance is not an attractive culture, it’s an adversarial one! And you could say inertia, and yet: look who’s running the HHS! There are kids right now, that I know in my real life, who look like you or me, who went to fancy Ivy League school, and they are vaccine skeptical. What about inertia and general intelligence then? So I’ll just say, you know, putting yourself out here on this forum, being all like, “I will AMA, I am the voice,” and then to be so arrogant: you are your own evidence for why maybe it won’t last 10 years.
All good points! Nobody knows the future!
I jumped into this thread to share my thoughts, and my thoughts alone, because I'm not sure there are a lot of radiologists on HN. I certainly don't speak for all radiologists.
But, I would submit to you, that rapid, radical changes to the practice of medicine are rare, if not impossible.
I don't think it will go away as long as we have third party paying for the costs and AMA controlling competition.
If I had to pay $500 or whatever to get a scan, and instead I could get my data, send it to a model and only follow up if it came back bad, I would. But now someone else pays and there are laws and regulations that prevent people from controlling their data, or at least make it difficult. Kind of weird I have a file on me that I have never seen.
> Kind of weird I have a file on me that I have never seen.
Can’t you ask for it?
Nope. You have an absolute legal right to obtain copies of your medical images and other data. Who paid for it is irrelevant. The provider can charge you a nominal fee for copying the files but they can't keep it from you.
https://www.hhs.gov/hipaa/for-individuals/guidance-materials...
At the same time it can be a handy tool to be a first cut at triage.
It's really not a matter of "full replacement or bust".
This is true!
Respectfully, it doesn't matter what you expect or think. What matters is this:
I'm almost in the same situation as you are. I have 22 years left until retirement and I'm thinking I should change my career before I'm too old to do it.> it doesn't matter what you expect or think
Can you please edit out swipes like that from your HN posts? (Prepending "respectfully" doesn't help much.) This is in the site guidelines: https://news.ycombinator.com/newsguidelines.html.
The rest of your comment is just fine of course.
I agree with you fully.
And, I didn't say I would never be replaced. I said I would finish my career, which is approximately 10 more years at this point.
What career would you change to that would be safe, given the conditions you provided and your time horizon?
The original author of the paper about the technological singularity [1] defines it as simply the point where predictions break down.
If AI gets to the point where it is truly replacing radiologists and programmers wholesale, it is difficult to tell anyone what to do about it today, because that's essentially on the other side of the singularity from here. Who knows what the answer will be?
(Ironically, the author of that paper, being also a science fiction author, is also responsible for morphing the singularity into "the rapture for nerds" in his own sci-fi writing. But I find the original paper's definition to have more utility in the current world.)
[1]: https://accelerating.org/articles/comingtechsingularity
I think that if AI can replace software engineers then AI can replace any job because the domain of software engineering is pretty much everything.
I don't think robotics is progressing at nearly the same pace as AI so for a while there will still be a bunch of manual labor for us to fight over. :-)
Not really. AI advances are boosting progress in robotics. Many US and China robot companies are developing general androids.
Plumber, car mechanic, something that human hands can do but robots can't.
Crime?
I formerly worked at https://corista.com, dealing specifically with integrations with AI pathology algorithm vendors like https://www.qritive.com and https://visiopharm.com. (Note: we dealt specifically with microscopy rather than radiology.)
I really think we were doing things the “right way” before I left: providing AI analyses from various vendors as overlays for slides, and being able to pre-flag slides with obvious cases of cancer or other infection. (These analyses typically being custom algorithms provided by vendors through APIs we collaborate on, not LLM output.)
With things like this simply built into an existing, established pathology platform, I’m pretty confident a team of 4-5 pathologists could do the work of 6-8, with better quality output, similar to how Copilot and other tooling speeds us up as developers. At $300-400k/yr/doctor, that’s considerable savings (and the opportunity to allocate more doctors in specialties that aren’t easily automated).
However, for lots of reasons, it seems the market doesn’t necessarily agree with me on the value potential of this approach in this field (which can certainly be a self-fulfilling prophecy).
>the value potential of this approach in this field
As a medical school dropout, I think the largest "negative" in value assessment would be the inability to shift malpractice liability over to a human practitioner (i.e. can you insure a company's faulty AI/LLM?) — yet today I just read a HN article on insurers offering a new product: insuring chat/LLM fuck-ups.
As exciting as this all is, I'm still so glad I didn't become a radiologist!
For what it's worth, I almost dropped out of medical school in my second year to go back to a career in tech. It was a close thing.
Still, I have enjoyed my time in radiology. It is insanely challenging to do well, and has been very rewarding.
I think I could have been happy with the physician lifestyle, but I knew I wouldn't make good life decisions surviving/getting through residency. I am glad to have survived (e.g. career change).
I've spent two decades as an electrician, and as the story goes: it is what it is... but also what you make of it. An incredible Good Fortune keeps me from having to work full-time, but as I've entered my forties I know I want to get in an air conditioned office (before too long).
Perhaps I'll first move to a country where healthcare isn't tied to full-time employment. Best of luck to you in your Interpretations.
> “People should stop training radiologists now,” Geoffrey Hinton said, adding that it was “just completely obvious” that within five years A.I. would outperform humans in that field.
I'm starting to notice a strong trend where non-domain experts will confidently assert that X field will be replaced by AI.
Hinton is a domain expert in AI, case in point Scott Aaronson cites AI revolutionizing protein-folding research, putting many protein-folding academics out of jobs.
I'd argue that AI experts and CS theorists are in a better position to speak about technological change than the domain experts who are trained/selected mainly to practice the status quo.
> putting many protein-folding academics out of jobs.
Has this actually happened? I'd wager it's more like in OP's article, AI as a helper that speeds up research, enabling bigger or different research questions. Reliable prediction not of protein shape, but of protein-protein or protein-mRNA interactions, for example.
Isn't this a common trend with CS? It seems like there's just constant over hype that drives booms and crashes, even to such a degree that if something was there it can't even be built. We have billionaires and hundred millionaires that made their money on VR, blockchain, and Web3. The wealth comes even when no product does and that seems like a big issue that's being exploited
https://archive.is/e2phZ
For a more technical dive into startups that have been chugging along with AI in pathology for ~decade, check out this:
https://www.owlposting.com/p/what-happened-to-pathology-ai-c...
> One pathology AI founder told me that it wasn’t hospitals or diagnostic labs that showed the most promise. It was R&D groups within Big Pharma. Those scientists and executives wanted new tooling. They were often sitting on massive internal datasets, had real budget allocated to experimental tech, and — critically — had a clear ROI if your model helped shave months off a study or more precisely target the right patient cohort. Most importantly, pharma didn’t care as much about the regulatory headaches, as they weren’t using your model to diagnose patients
Really nice article, thanks for sharing. With digital pathology ai startups going bust left and right it's quite an accurate analysis. One thing it's missing in the "product" path is preanalytical differences - same tissue processed in different labs can produce wildly different pixels.
<3
The key to the power of the LLM is that the training process can learn effectively from vast corpora of unlabelled text. Unfortunately, there is no comparably vast database of medical images.
In order to "crack" radiology, the AI companies would need to launch an enormous data collection program involving thousands of hospitals across the world. Every time you got an MRI or X-Ray, you would sign some disclosure form that allowed your images to be anonymously submitted to the central data repository. This kind of project is very easy to describe, but very difficult to execute.
I agree with you, but here is where things get tricky:
Everyday I see something on a scan yhat I've never seen before. And, possibly, no one has ever seen before. There is tremendous variation in human anatomy and pathology.
So what do I do? I use general intelligence. I talk to the patient. I talk to the referring doctor. I compare with other studies, across modalities and time.
I reason. I synthesize. I think.
So my point is, basically, radiology takes AGI.
"Takes AGI?" It's the same argument people used to make about NLP. By now, it's proven wrong.
Another way of saying this that might not trigger your alarm bells is that even a perfect image analyzer is not enough to replace a radiologist. The job consists of much more than just analyzing images.
>Unfortunately, there is no comparably vast database of medical images.
Even a tiny hospital with radiology services will produce many thousands of images with accompanying descriptions every year. And you are allowed to anonymize and do research on these things in many places as neither image nor accompanying description is a personal identifier.
So this is yet another Hinton-ish prediction, any time soon radiologist are going dodo. This time LLMs will crack the nut that image recognition have failed at for 20 years.
Where LLMs have succeeded is in doing hot takes that miss the mark, they should be really good at cornering the "prematurely predicting demise of radiologist"-market
They’ll have better luck in countries like the U.K. where medical data is at least somewhat more organised by virtue of being under the NHS umbrella
I don't think the UK NHS has a single centralized PACS or EHR.
It's organised into regional NHS primary care and acute trusts, which usually share their records these days. It's still substantially more integrated than in many other countries.
"The staff has grown 55 percent since Dr. Hinton’s forecast of doom, to more than 400 radiologists."
Wonder what other forecasts of doom he is wrong about :|.
They are also all driving themselves to the hospital instead of using self-driving cars. Different forecaster though
[dead]
AI is not easily going to "steal" any properly gatekept jobs that require a licensed practitioner.
Machine learning technology can be used to search for artifacts in medical diagnostic images that humans might miss. The flagged situations still have to be interpreted by a radiologist.
So will it cost me less than $1.5k the next time I need an x-ray?
of course not, that would be silly. it will cost $1.5k + the AI analysis fee
The thing I find most interesting about ML in radiology is that a computer can observe the entire dynamic range of the sensor at once. A human will only look at a window or a compressed window.
This is a very key point. Perhaps that discrepancy can be leveraged in image generation to save time.
It already is, which is why rads input window/level settings.
So are datasets/currently available data the limitation here?
Let's say a major healthcare leak occurred, involving millions of images and associated doctor notes, diagnostics, etc... would this help advance the field or is it some algorithmic issue?
People get tripped up by thinking "there is a subset of what I do that only humans can do and so that means AI will not eliminate my profession entirely and my job is safe".
Let's assume for now that it's true that AI can't do a certain subset of your work. Your profession won't be eliminated from the earth, that's true. But if 80% of your work can be done by AI, 80% of your work will be done by AI. There will still be humans kept around for that remaining 20%, but fewer of them will be needed.
The demand for radiologists is effectively infinite. Right now the healthcare system is supply constrained. If AI tools reduce the work of reading an image then more imaging studies will be ordered.
Also, many radiologists do interventional procedures directly with patients. We're a long, long way from being able to significantly automate that work.
So I recently needed an MRI to evaluate a rotator cuff injury. Because MRIs are expensive, insurnace required that I do 6 weeks of physical therapy first to see if that would resolve it.
Is this because the imaging process is so expensive, or because the radiologist who reads the image is so expensive?
If demand for an expensive imaging process increases, it will get even more expensive.
When you get a medical test like an MRI, there is both a technical fee and a professional fee that is paid.
For a shoulder MRI, the technical fee, which goes to the facility that performs the scan, maybe $1,500.
The professional fee, that goes to the radiologist that interprets the scan, maybe $80.
Patients often only see a global fee.
MRI machines and radiologist reads are both expensive. These are among the largest areas of medical spending for most health plans (and those costs are largely passed on to subscribers and employers).
Many rotator cuff injuries do improve to some extent with targeted physical therapy. You can accomplish a lot by strengthening the stabilizer muscles, and by learning how to set your shoulder joint in the correct position before taking a load. So the insurer's position is reasonable for most cases. If the injury is more severe then your doctor can submit a request to skip the normal step therapy process.
This is 100% true and very insightful, but not obvious to people who don't work in healthcare.
If there is one thing that NN can do amazingly well it is pattern recognition. The biggest blockers to adoption are the EHR providers, regulations and just inertia. It will happen, the only question is when and to a small degree, how.
EHR vendors aren't the major blocker here. Most of them have pretty good APIs now (at least for reading USCDI data from patient charts). But the data quality as entered by users is often crap: when you're training it's going to be garbage in / garbage out unless you put enormous efforts into cleansing. And the raw images are often not in the EHR but rather in a separate PACS.
From what I have seen trying to build out new features into this area, EHR vendors often act like monopolies blocking a lot of competition by controlling any data exchange. The vendor lock-in is strong and very harmful to improving process and, ultimately, harms patient care.
What specifically are they blocking? The most popular EHRs from vendors like Epic and Oracle have extensive interoperability APIs for multiple open industry standards. In general you can read all USCDI data pretty easily. Writing data is more limited, for understandable reasons. And if you want to distribute a SMART on FHIR app then you'll have to go through some sort of approval process.
The strong vendor lock-in is unavoidable given how deeply an EHR gets embedded into any provider organization. No matter what regulators and standards bodies do, it will never be easy to switch.
For an ML model, a sofa with a tiger pattern might be a tiger, if in its training dataset tiger stripes always means tiger.
It does not have common sense.
For small models, yes. But for the kind of massive multimodal models getting trained these days, the concept of "a pattern on an object" will exist. It doesn't need to have seen a tiger sofa before. But tiger + sofa + patterns on an object is enough to not run from the tiger print.
Reminder that according to https://openai.com/index/healthbench/, not only do current AI models provide far better recommendations than doctors, but even if doctors know the AI recommendations as a starting point, human revision does not bring any improvement.
Honestly this doesn't surprise me, considering the quality of the average doctor.
“Radiologists do far more than study images. They advise other doctors and surgeons, talk to patients, write reports and analyze medical records. After identifying a suspect cluster of tissue in an organ, they interpret what it might mean for an individual patient with a particular medical history, tapping years of experience.”
Now think about how much of software development is typing out the code vs talking to people, getting a clear definition of the problem, debugging, etc. (I would love an LLM that could debug problems in production — but all they can do is tell me stuff I already know). Then layer on that there are far more ideas for what should be built than you have time to actually build in every organization I’ve ever worked in.
I’m not worried about my job. I’m more worried my coworkers won’t realize what a great tool this is and my company will be left in the dust.
that great tool does not need ten years of experience to use. your coworkers will be able to catch up quite easily
Fwiw I think this is not true right now. It might be true in the future, I guess. Today you have to learn some pretty specific techniques and they don’t come naturally to most engineers I think. They are accustomed to one-shot Google searches, not prompt engineering. Yes, you can learn how to prompt engineer, but it takes trial and error. A lot of my co-workers today still believe the LLM is useless because it hallucinates API definitions. They very much are not used to tools that require the level of hand-holding you need today.
Either teaching them is going to get easier, which is somewhat plausible actually — or you think LLMs are going to get dramatically easier to prompt with more reliable output. I actually think that’s somewhat implausible. Log(compute) is a pretty hard number to move past a certain point and it seems we’ve hit that point already. RLHF is great, but it doesn’t scale well and the result is that some tools are hard to build. Giving me millions of context tokens is neat, but if it’s wildly expensive and confuses the model on top of that, it’s not actually useful.
I do think there is room to improve and of course there are new techniques to be discovered. The models may even help discover some of them, although I am skeptical that they will find Everest by hill climbing when they’re currently on Pike’s Peak. Or maybe we really need to find Olympus Mons.
Tl;dr — maybe!
This reminds me of the idea that Human-Chess partnerships would be the ultimate manifestation of Chess genius. I'm not sure whether the idea is still holding on but engines are so far ahead of human play that I doubt a human in the loop can add anything these days given how devastatingly far ahead the engines are and the advent of machine learning techniques.
Chess reminds me more of programming given the set of defined rules in each. However, I'm biased as I work in radiology and program more as a hobby. So far I've seen way more tools to help me code than to accurately detect radiologic findings.
So far. Computer vision is currently lagging NLP, but I wouldn't expect that to last.
> but I wouldn't expect that to last
Do you have any links to research or work being done on computer vision that leads you to this conclusion? Would love to check it out!
You can compare best image synthesis and image understanding from two decades ago (SIFT / HOG), from a decade ago (CNN, SdA) and now (Transformer). Very rapid progress that went from being able to unreliability recognize a face to getting to outperforming human professionals (see MMMU) is quite remarkable.
AIUI, and I may be wrong, but each of the mentioned technologies was a "breakthrough" technology - not iterative improvement. Along this vein, I was wondering if there was some promising, novel research OP was aware of for image understanding.
The most recent of which you mentioned, Transformers, is used by both LLMs and image synthesis/understanding. The parent posits that while computer vision lags behind NLP, this may not continue. While your comment points out that image synthesis and understanding has improved over time, I'm not sure I follow the argument that it may soon leapfrog or even catch up with LLMs (i.e. text understanding and synthesis.)
My understanding is that the way that chess.com and other online services detect cheating is by comparing the human-made moves to a "perfect" version of what the chess engine would play.
Which gives credence to your theory that people aren't bringing much to the table.
TL;DR: “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.”
Glad to hear that this is one area that AI is conclusively useful.
Still not clear that the already superhuman capabilities of AI won't still fully supplement radiologist interpretive skills with every additional bit of training data that comes in.