The A.I. Radiologist Will Not Be with You Soon
nytimes.com
nytimes.com
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...
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?
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.
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.
(1)https://www.arrt.org/pages/earn-arrt-credentials/credential-...
https://www.hhs.gov/hipaa/for-professionals/special-topics/d...
"We should violate medical privacy laws on an enormous scale because China" Okay buddy.
Being first is not the same as being happy.
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.
People keep claiming that other countries will out compete us in medical research and it keeps not happening.
>Concerns raised over AI trained on 57 million NHS medical records
https://www.newscientist.com/article/2479302-concerns-raised...
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.
This article is surprisingly accurate. I fully expect to finish my career without being 'replaced' by AI.
Happy to debate/answer questions :-)
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.
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.
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.
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 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 ;-)
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.
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.
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.
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.
Firstly, that doesn't tell us whether there is bias in data. That tells us whether or not there is bias in their data.
Secondly, it tells us it can train to spot things that human rads do not train to spot. It tells us nothing at all about whether or not an AI can train to spot things a human rad also trains to spot, but can't.
Human rads don't train to spot race. Why? Because they don't need to do so. Human rads do train to spot pathologies in as early a stage as possible. I've never seen an AI spot one at an earlier stage than the best human rads can. But I have seen several AIs fail to spot pathologies that even human rads at the median could spot.
That's the state of play today. And it's likely to remain that way for a long, long time. Human rads will be needed to review this work not because they are human, but right now, it's because human rads are just better. At the top end, human rads are not only better, but are manifestly superior.
They aren't studying bias in data. They were studying bias in AI. The data used was 850,000 chest x-rays from 6 publically available datasets. They aren't studying whether this dataset differs from the general public or has some kind of racial bias; that's irrelevant to the study.
> it tells us it can train to spot things that human rads do not train to spot
You're kidding yourself if you think you could determine someone's race with 97%+ accuracy from a chest x-ray if only you trained at it. The study authors (who are themselves a mix of radiologists and computer scientists) claim that radiologists widely believe it to be nearly impossible to determine race from a chest x-ray. No one is ever going to try to train radiologists to distinguish race from chest x-rays, so you'll always be able to hold out hope that maybe humans could do it with enough training. But your hope is based on nothing; you don't have a shred of evidence that radiologists could ever do this.
> I've never seen an AI spot one at an earlier stage than the best human rads can.
According to the article, AIs aren't trained to do this, because we don't have datasets to train this. You need a dataset where the disease is correctly labeled despite the best radiologists not being able to see it in the x-ray. Trained with a good enough dataset, they'd be able to see things we miss.
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.
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!
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.
Will you be able to source a radioactive source for your x-rays?
DIY radiation therapy would be a whole new level.
Healthcare-grade x-ray tubes to put in your (expensive) x-ray machine are not something you easily can obtain without a license.
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.
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.
The other doctors will still be there for you to sue.
> 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?
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.
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.
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
I should have said when an AI can do my job it can do anyone's job.
- 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].
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
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.
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.
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).
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.
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.
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.
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.)
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.
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.
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.
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.
For what it's worth, I started a new residency program to train more radiologist, so I do have some skin in the game.
I don't think they would ever do it, but technically the ABR could stop them.
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.
I'm sure this will all get better with captain brainworms at the helm.
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.
I'm not worried about medical folks having "time to clean" any time soon.
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.
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?
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'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 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).
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.
Voice-over performers. Sketch artists. Audiobook actors. Tutors (see: https://www.google.com/finance/quote/CHGG:NYSE?window=5Y)
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.
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
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.
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.
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.
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.
Well... that's just not true.
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 %
[0] https://www.siemens-healthineers.com/en-us/radiotherapy/soft...
It's really not a matter of "full replacement or bust".
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?
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.
Can’t you ask for it?
https://www.hhs.gov/hipaa/for-individuals/guidance-materials...
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).
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!
Still, I have enjoyed my time in radiology. It is insanely challenging to do well, and has been very rewarding.
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.
I'm starting to notice a strong trend where non-domain experts will confidently assert that X field will be replaced by AI.
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.
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.
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
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.
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.
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
Wonder what other forecasts of doom he is wrong about :|.
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.
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?
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.
Also, many radiologists do interventional procedures directly with patients. We're a long, long way from being able to significantly automate that work.
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.
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.
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.
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.
It does not have common sense.
Honestly this doesn't surprise me, considering the quality of the average doctor.
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.
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!
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!
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.)
Which gives credence to your theory that people aren't bringing much to the table.
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.