We need a thread on why AI WILL NOT replace radiologists in the next several decades. I’ll start.

Started by SeisK
This forum made possible through the generous support of SDN members, donors, and sponsors. Thank you.
Get help with your application

Use all the free resources available to you from SDN: articles, guides, expert advising, forums discussions, and school research.

Browse articles Compare schools My application plan Read Interview Feedback
Practice for interviews Ask an Advisor
Advertisement - Members don't see this ad
Dude I've been seeing that same video in different variations for two decades+ now. Talk about beating a dead horse. When can I actually buy a goddamn self driving car?
 
As someone who just matched DR this year and will be starting intern year soon, I'm hopeful that the tech bros are going to overpromise and underdeliver like they usually seem to do. Radiology seems easy to replace on paper, but I'm hopeful that it's not. Doing six years of training and then having to do a second residency would not be ideal.
 
Maybedoc1 said:
As someone who just matched DR this year and will be starting intern year soon, I'm hopeful that the tech bros are going to overpromise and underdeliver like they usually seem to do. Radiology seems easy to replace on paper, but I'm hopeful that it's not. Doing six years of training and then having to do a second residency would not be ideal.
Click to expand...
Right there with you, no clue what I would retrain in
 
Maybedoc1 said:
As someone who just matched DR this year and will be starting intern year soon, I'm hopeful that the tech bros are going to overpromise and underdeliver like they usually seem to do. Radiology seems easy to replace on paper, but I'm hopeful that it's not. Doing six years of training and then having to do a second residency would not be ideal.
Click to expand...
They are. Radiology is difficult because it requires integration of multiple knowledge domains as well as the visual aspect of things. Just wait till you see your first few complex cases and you’ll have an understanding of just how opaque image interpretation can be.
 
It's funny because so far AI seems far more a threat to the artistic or creative fields right now, where hallucinations and making **** up are A-OK.
 
LegendaryBibimbap said:
It's funny because so far AI seems far more a threat to the artistic or creative fields right now, where hallucinations and making **** up are A-OK.
Click to expand...
I think artists, at least at the higher end in the true creative fields, will never really have anything to worry about since there's always going to be a hunger for art produced by other humans and the appreciation of art is entirely subjective anyway. With a field like rads, the inverse is true--AI is going to be useless if it's not correct, but once it's shown that AI is superior to humans and even human+AI on an objective scale there's no reason to keep humans around and you now have a moral obligation to get humans out of the way completely. Not saying this will happen anytime soon but it's an interesting thought expeirment.
 
Last edited:
Maybedoc1 said:
As someone who just matched DR this year and will be starting intern year soon, I'm hopeful that the tech bros are going to overpromise and underdeliver like they usually seem to do. Radiology seems easy to replace on paper, but I'm hopeful that it's not. Doing six years of training and then having to do a second residency would not be ideal.
Click to expand...
It's not gonna be like you're done with residency and suddenly there's no jobs. The field will slowly change as it always has over decades. The market may not be as hot as it is now but you'll have a job til you die, it just might not look the same as it did when you first started. Besides, nothing really is truly insulated from being changed by AI and if you want to do manual labor like surgery go for it if you're that scared. I don't want that kind of lifestyle personally.
 
SeisK said:
This post is made by a recent study here that delineates 1/6 medical students interested in radiology decide not to after learning about AI minimally from other attendings (almost certainly in specialties that do not generate radiology reports), and it’s something that keeps coming up, at first amusingly, but now it’s slowly become annoying.

Radiology is the best specialty. We deal with essentially no crap that other specialties have to on a day to day basis, we’re extraordinarily efficient, we deal with ALL the type of things you learn about in med school (even those pesky lysosomal storage diseases you were told never mattered), you are directly exposed to the applications of the coolest modern physical and technological sciences, and you’re paid appropriately for it unlike a large swath of the rest of medicine.

My motivation in this is, well, I’m a jealous guy. I want all the smart, driven, charismatic people to come to my specialty and in their (necessarily) naive state as young influenceable medical students I think a bunch of smooth-brained window-lickers (with the utmost respect) are dissuading them from this thing. So I want to start a thread on why this is so horribly mistaken.

—-

This is a post I made on auntminnie on a related thread, which I think really drives the point home. I’d love to hear other’s thoughts (doesn’t matter how thought out or not). This comes from a background in not a small amount of literature review, and clinical trial research.

“
We don’t really have successful models that can predict the future well in economic terms, and when that happens emotions run rife and dominate the conversation.

You have computer scientists and software developers that immediately show their futorology bias by repeatedly spouting “radiologists will be obsolete” even today, despite what little AI has been implemented probably isn’t saving anyone any time, and the only RoI comes in the form of higher quality reads.

You have radiologists on the other hand, who possibly in an ego-defense kind of way, state “AI will only assist us not replace us” when assisting you is tantamount to replacing you. If I need five radiologists instead of ten to get through a list in a day, I’ve replaced five with the implementation.

But the fact of the matter is, actual clinical implementation of algorithms and reproducibility studies have not matched trial studies in accuracy, and will continue to not do so for the next several decades at least, for many reasons:

I’ll start with the obvious: no radiologist is being replaced until radiologist+AI is better than radiologist in a large scale, heterogenous population. I’ll go more into that below. Starting with that:

1. Edge cases are not a negligible proportion of our studies. Even if they were, there are no studies or software present that assess the accuracy of an AI in determining edge-scenarios, so how am I going to know “you don’t need to look at this study” even if AI surpasses my ability? This is why AI that is a “normal identifier” is far away. Far away. FAR. AWAY.
2. Training datasets are not generalizable because of subtle differences in the scanners underlying the data acquisition, and heterogenous datasets are proprietary making it extremely difficult sometimes to acquire larger datasets to train your algorithms. There are some efforts to overcome this, but five large homogenous datasets do not a heterogenous sample make.
3. The Black Box problem. This is tied to problem 2. There’s often something else consistently on the image that may demonstrate why something is going to happen that’s coincidentally tied to the pathology, that we can’t identify. “Who cares if the diagnoses are accurate?” I do MFer, because if in a multivariate analysis we account for this hidden “black box variable” and find the machine is now worse than humans, I’m not going to use the thing. I have no idea if there are black box variables in your algorithm to even begin knowing how to set up a multivariate analysis in its elimination. This right here is almost certainly why clinical implementation of extremely promising algorithms have been milquetoast. Frankly, there’s s*** I can’t see that the thing is using to cheat. When you employ the algorithm in another population that doesn’t have that hidden variable, it fails. Two ways of getting around this are localizers to help the radiologist figure out what the AI is seeing, and testing the algorithm on an extremely heterogenous population (lots of different types of patients, lots of different types of scanners, lots of different types of clinical settings in acquisitions).
4. AI is exceptionally vulnerable to artifacts that are trivial to us.
5. AI does not reproduce human-level sensitivity or specificity on cross-sectional imaging, which is likely our most important work as it’s here we often truly make diagnoses, whereas in planar imaging we only provide descriptions that lean in favor of diagnoses.

Additionally, here are the bigger deals:
6. Greater accuracy doesn’t save anyone any time. Or at least it morally shouldn’t. AI+Radiologist surpassing radiologist performance assumes the radiologist hasn’t changed their behavior in the presence of AI, unless the software has accounted for that behavior in its pre-release trial. A radiologist going through studies quicker because they have AI on board isn’t reproducing the study conditions, so its conclusions can’t be guaranteed to extrapolate, and the person suffering that decision is the patient. Because of this, AI doesn’t actually yield a RoI for the radiology practice when used. Then again, there are a lot of dubious radiologist practices out there, and they’re becoming dubiouser with private equity expansion.

Finally:
7. No prospective trials. This is a big deal, probably the biggest. Nothing, I mean nothing in any field of medicine becomes or supplants the standard of care until you have a large, national-scale, large AND SUFFICIENTLY HETEROGENOUS sample population randomized clinical trial demonstrating the new method surpasses the old in terms of morbidity and mortality years down the line—NOT FOR MODALITIES AS A WHOLE, but for the thousands of specific pathologies picked up on that modality. There is a lot of groundwork to be done before you’ll let the experimental arm be put at risk of the study going wrong. You do this by performing quite exhaustive retrospective studies analyzing variables important to the outcome, and for AI that’s a lot of variables. Additionally and most importantly, this is also overcome by making the experimental population arm be “existing standard + new intervention,” which I’ll again remind you doesn’t replace a single radiologist. After this case is met can you maybe attempt to use the “new intervention” alone without the existing standard. Even a single such Phase 3 trial takes YEARS, and a simple search of clinicaltrials.gov will show that there is not even a phase 1 trial of ANY imaging modality AI versus radiologist. The FDA will NEVER clear these devices as standard of care until a Phase 3 looks gorgeous and published on the front page of NEJM, and right now we don’t even know yet how to set up an appropriately sampled population for such a phase 3 as, again, generalizability is an enormous issue (you’d have to sure any new variant of image acquisition is covered). Keep in mind though that while this is the biggest deal, it is the BIGGEST deal. Once an AI has overcome this hurdle for a specific pathology, the radiologist has lost. If AI says “acute interstitial edematous pancreatitis” and AI > AI + Radiologist for this pathology, that’s what goes in the report even if you don’t see it.

And again, I’ll remind you. You set up clinical trails NOT FOR MODALITIES AS A WHOLE. But for specific pathologies. You need a phase 1 for acute interstitial edematous pancreatitis, acute necrotizing pancreatitis, chronic pancreatitis, pancreatic adenocarcinoma… and so on. For the thousands of such diagnoses a radiologist is required to identify and describe. That’s a lot of work for a small group of software devs who don’t know what pancreatitis is.

Given the above, and probably because private equity would prefer modest short term return than huge long term return, the AI software we do see is relatively small, sold to radiologists rather than providers directly, and is always advertised as an adjunct to the standard of care rather than any kind of replacement for it lest they suffer the FDA and litigation’s wrath.

And I’ll remind everyone finally that all of this will reduce the need for radiologists, but still will not replace them. I see the future of radiology one that is much more data / mathematics / physical science driven as the number and complexity of imaging modalities grows and as the importance of AI grows. We have to become experts on it. We have to become as familiar with the language of AI implementation into healthcare as the oncologist is with their various chemotherapies, and the subtleties of using them depending on the context of what cancer. We really should be the experts and keepers of this, and become as familiar with it as the computer scientists themselves. For the benefit of our patients. Learn it, not because you fear it (if you’re new you don’t have much to fear) but because you want to employ it to save your patient’s lives. “
Click to expand...
I'm interested in your thoughts on a couple updates on the field

Here was a prospective trial demonstrating the efficacy of a generalist model AI for neuro MRI. It wasn't comparing its accuracy to radiologists but showed high accuracy among 52 diagnoses in a real clinical environment: Learning neuroimaging models from health system-scale data - PMC

Here's a proposed pathway for approval of generalist AI models that circumvents the “one prospective trial for every pathology" model mentioned in your original comment and essentially treats AI as a physician with clinical benchmarks and training before approval. It was referenced in the first paper:
 
Last edited:
Advertisement - Members don't see this ad
the AI takeover of radiology is inevitable and we will start to see it sooner than later, likely significant changes in the next 5 years. the AI naysayers can continue to bury their heads in the sands. if you are an attending, save money. if you a resident, think of a plan B. if you are a med student, avoid radiology. its simple.
 
radrad5 said:
the AI takeover of radiology is inevitable and we will start to see it sooner than later, likely significant changes in the next 5 years. the AI naysayers can continue to bury their heads in the sands. if you are an attending, save money. if you a resident, think of a plan B. if you are a med student, avoid radiology. its simple.
Click to expand...
Radstradamus coming in hot. How did the 5 year prediction turn out for Hinton?
 
odyssey2 said:
I'm interested in your thoughts on a couple updates on the field

Here was a prospective trial demonstrating the efficacy of a generalist model AI for neuro MRI. It wasn't comparing its accuracy to radiologists but showed high accuracy among 52 diagnoses in a real clinical environment: Learning neuroimaging models from health system-scale data - PMC

Here's a proposed pathway for approval of generalist AI models that circumvents the “one prospective trial for every pathology" model mentioned in your original comment and essentially treats AI as a physician with clinical benchmarks and training before approval. It was referenced in the first paper:
Click to expand...
My opinion has not changed in the subject!
 
radrad5 said:
the AI takeover of radiology is inevitable and we will start to see it sooner than later, likely significant changes in the next 5 years. the AI naysayers can continue to bury their heads in the sands. if you are an attending, save money. if you a resident, think of a plan B. if you are a med student, avoid radiology. its simple.
Click to expand...
This is just cope from someone who probably didn't go into rads because they were so scared of AI and now they hate whatever specialty they ended up in. AI hasn't done s**** in practice and probably won't do s**** for a VERY long time. No actual radiologist takes it seriously and guess who controls whether or not any of the AI tools actually get implemented? Radiologists.
 
hopefulraddaddypls said:
Would it be wise to look to switch out of rads? -Current R1
Click to expand...
Frankly, if youre asking this question as a second semester R1 and you still haven't realized how BS the AI threat is you are 1. Just looking for an excuse to leave rads cuz you hate it or 2. Just that dumb. In either case, yes you should leave. You're gonna have a tough career ahead of you
 
Radguy12 said:
Frankly, if youre asking this question as a second semester R1 and you still haven't realized how BS the AI threat is you are 1. Just looking for an excuse to leave rads cuz you hate it or 2. Just that dumb. In either case, yes you should leave. You're gonna have a tough career ahead of you
Click to expand...
Is it really BS? Money keeps pouring into it
 
hopefulraddaddypls said:
Is it really BS? Money keeps pouring into it
Click to expand...
The US economy is literally based on pouring money into BS. Do you remember the 2008 recession? Have you heard of the dot com bubble? The great depression? Enron? NFTs? Crypto? This is just the next iteration. You are a second semester R1 so I assume an adult in their mid to late 20s. If you have not realized this and are being swayed by the Internet and losers who are not even radiologists, I really do think you should consider switching specialties. Go break your back in the OR or talk to whiny patients all day in clinic. Everyone will happily take you because you'll confirm their biases and justify their decision to not do rads. You are too dumb to be a radiologist and I mean that in the nicest way possible.
 
Last edited:
Also, count my lung nodules. Where is my lung nodule counter? I have been promised a lung nodule counter over and over again but I still don't have one. I need it to compare to chest CTs from up to 5 years back automatically and place a circle around each nodule so i can clearly see then. It should be accurate enough to where I don't have to go back and confirm anything and I need to not be legally liable if I miss a nodule when using it. It is the simplest task in radiology. Do this and I will quit my job and go work at McDonalds
 
I think you guys are taking this the wrong way, all you need is for AI to reliably ID 1 thing to tank the field: Normal scans.

That's it. The field needs to evolve in some way to adapt to that eventuality
 
Last edited:
Doc mu said:
I think all you guys are taking this the wrong way, all you need is for AI to reliably ID 1 thing to tank the field: Normal scans.

That's it. The field needs to evolve in som way to whether that eventuality
Click to expand...
lol that 1 thing is actually many things. A normal chest x ray is different than a normal abd x ray then multiply that out by all the many different types of studies including different protocols and make sure it doesn’t miss things that look slightly different when they’re in a suboptimal contrast timing phase or have artifact.

Are you a radiologist?
 
Doc mu said:
I think all you guys are taking this the wrong way, all you need is for AI to reliably ID 1 thing to tank the field: Normal scans.

That's it. The field needs to evolve in som way to whether that eventuality
Click to expand...

A high sensitivity for normal scans is mathematically identical / the same as a high specificity for every pathology. This is not a trivial task.
 
Advertisement - Members don't see this ad
Steve_Zissou said:
lol that 1 thing is actually many things. A normal chest x ray is different than a normal abd x ray then multiply that out by all the many different types of studies including different protocols and make sure it doesn’t miss things that look slightly different when they’re in a suboptimal contrast timing phase or have artifact.

Are you a radiologist?
Click to expand...
Yeah and you think this is exceptionally difficult because...?

"doesn’t miss things that look slightly different when they’re in a suboptimal contrast timing phase or have artifact" - easy, if in doubt, flag it for the radiologist to see, majority will still be filtered out.
 
SeisK said:
A high sensitivity for normal scans is mathematically identical / the same as a high specificity for every pathology. This is not a trivial task.
Click to expand...
Those aren’t identical tasks. ‘Normal vs not-normal’ is a binary triage problem; pathology-specific specificity is a multi-label diagnostic problem. An AI can be useful by confidently clearing only obviously normal scans and routing everything equivocal or abnormal to a radiologist. It doesn’t need to distinguish every pathology correctly to take a large chunk of normal-volume work off the table
 
Doc mu said:
Yeah and you think this is exceptionally difficult because...?

"doesn’t miss things that look slightly different when they’re in a suboptimal contrast timing phase or have artifact" - easy, if in doubt, flag it for the radiologist to see, majority will still be filtered out.
Click to expand...
What makes you think it’s not exceptionally difficult? Has it been solved in any real capacity yet?
 
Doc mu said:
Those aren’t identical tasks. ‘Normal vs not-normal’ is a binary triage problem; pathology-specific specificity is a multi-label diagnostic problem. An AI can be useful by confidently clearing only obviously normal scans and routing everything equivocal or abnormal to a radiologist. It doesn’t need to distinguish every pathology correctly to take a large chunk of normal-volume work off the table
Click to expand...
They are mathematically identical, yes. High sensitivity to normal is isomorphic to high specificity for all abnormal things simultaneously, and vice versa. Review your contingency tables.
 
Last edited:
Doc mu said:
It is already being used in europe for chest Xrays
Click to expand...
That’s where you have to be careful. Do you know where and for what it’s being used? Is it being used in hospital populations or in pre-screening like they sometimes do for jobs where the pretest probability is near zero for any clinically significant findings? What’s the sensitivity and specificity for it in its use case?

I’m not arguing that it won’t improve and that at some point it won’t be able to read a normal CXR as normal. I’m arguing that being able to do that for a variety of modalities in a large population is a supremely difficult series of very different tasks that will take many years to solve if it’s actually solvable. The claims are always exorbitant and the harbingers are always assured, but so far they seem to consistently be wrong when it comes to radiology, including people who are far more expert at understanding AI and ML than you, unless you happen to be Geoffrey Hinton.
 
SeisK said:
They are mathematically identical, yes. High sensitivity to normal is isomorphic to high specificity for all abnormal things simultaneously, and vice versa. Review your contingency tables.
Click to expand...
So you are implying that acheiving both in practice with AI is equally difficult?
 
Steve_Zissou said:
That’s where you have to be careful. Do you know where and for what it’s being used? Is it being used in hospital populations or in pre-screening like they sometimes do for jobs where the pretest probability is near zero for any clinically significant findings? What’s the sensitivity and specificity for it in its use case?

I’m not arguing that it won’t improve and that at some point it won’t be able to read a normal CXR as normal. I’m arguing that being able to do that for a variety of modalities in a large population is a supremely difficult series of very different tasks that will take many years to solve if it’s actually solvable. The claims are always exorbitant and the harbingers are always assured, but so far they seem to consistently be wrong when it comes to radiology, including people who are far more expert at understanding AI and ML than you, unless you happen to be Geoffrey Hinton.
Click to expand...
What makes it supremely difficult, AI is already being used for narrow diagnostic problems currently with decent success/numbers why would triaging scans by filtering out normals be much different?
 
Doc mu said:
So you are implying that acheiving both in practice with AI is equally difficult?
Click to expand...
Yes. An algorithm that to high specificity excludes any and all pathologies of clinical relevance is extraordinarily difficult.
 
SeisK said:
Yes. An algorithm that to high specificity excludes any and all pathologies of clinical relevance is extraordinarily difficult.
Click to expand...
The level of excellence you are implying is not necessary, when uncertain send it to the radiologist.

This would still cull the majority of the volume in most community centers
 
Doc mu said:
What makes it supremely difficult, AI is already being used for narrow diagnostic problems currently with decent success/numbers why would triaging scans by filtering out normals be much different?
Click to expand...
My dude, you’re asking me why it’s difficult, I’m telling you that no one has even remotely figured it out so clearly it’s not the low hanging fruit you think it is. Once again, are you a radiologist? Do you even understand the basic complexities of what makes radiology hard, or hell even the basics of radiology? Per your posting history it sounds like you’re maybe a first year resident in an unrelated field, if so then it makes sense why you wouldn’t understand how difficult it can be to suss out normal in many cases.

Explain to me why it hasn’t clearly been solved if it’s so easy. Give me more than a headline level read of why you think so. Your prior comment “Europe is using it” was basically that without any actually diving into what they’re using, who they’re using it on, and how useful it actually is, so do the work here if you’re going to make a claim.
 
Doc mu said:
The level of excellence you are implying is not necessary, when uncertain send it to the radiologist.

This would still cull the majority of the volume in most community centers
Click to expand...

It won’t happen without an upheaval in how the FDA clears regulatory devices. What you’re talking about is still a Class 3 device, which still requires a prospective clinical trial. And for assessing normals that requires an abundance of pathologies assessed for, with outcomes determined over decades.

I absolutely guarantee you that level of excellence IS necessary. It is a precedent that already exists in practice by legislative implementation.
 
Doc mu said:
What makes it supremely difficult, AI is already being used for narrow diagnostic problems currently with decent success/numbers why would triaging scans by filtering out normals be much different?
Click to expand...
Not it's not. It's obvious you're not a radiologist lmao. Ask anyone in this thread if they are actually using AI on a daily basis. The answer is resoundingly no. Tell me youve never dictated a scan without actually telling me.

Your gonna have to find a different way to process your regret for not going into radiology. I recommend radical acceptance
 
Radguy12 said:
Not it's not. It's obvious you're not a radiologist lmao. Ask anyone in this thread if they are actually using AI on a daily basis. The answer is resoundingly no. Tell me youve never dictated a scan without actually telling me.

Your gonna have to find a different way to process your regret for not going into radiology. I recommend radical acceptance
Click to expand...
For some reason there seems to be this trend of people who know next to nothing about radiology (and maybe think they do because they’ve looked at a few radiographs and CTs) and also know nothing about AI that make it a point to loudly and confidently declare that AI will take over portions or all of radiology and will do so easily. This is a recurring theme both here and on Reddit, and really I think it might be the same thing you see on Facebook with old people who very confidently proclaim very incorrect things based on fundamentally incorrect heuristics or headlines/memes they’ve seen.
 
Steve_Zissou said:
For some reason there seems to be this trend of people who know next to nothing about radiology (and maybe think they do because they’ve looked at a few radiographs and CTs) and also know nothing about AI that make it a point to loudly and confidently declare that AI will take over portions or all of radiology and will do so easily. This is a recurring theme both here and on Reddit, and really I think it might be the same thing you see on Facebook with old people who very confidently proclaim very incorrect things based on fundamentally incorrect heuristics or headlines/memes they’ve seen.
Click to expand...
People love headlines, but never actually read the articles. And even if they do, they fail to critically think about them. No one in rads is using AI except for a handful of people at ivory tower institutions who are themselves either consultants for or have a stake in one of these BS radiology AI companies. And by using I mean literally telling people they're using it without it be useful in any capacity. Its a non-sequitur but for some reason everyone outside of radiology seems convinced radiology is dead. More screeners for me I guess 🤷
 
Advertisement - Members don't see this ad
odyssey2 said:
he’s mentioning something that rules out negative screeners reliably. Probably doesn’t exist but these admins *think* they do
Click to expand...
Yeah that often seems to be the case, even with admin here. They told us our fracture detection system would have 99% or some absurd number for sensitivity and specificity, and it most definitely does not have that and often misses obvious fractures or mislabels random things (skin folds, lines overlying patient, etc) as fractures. They were convinced that it would save us a ton of time, and possibly be a screener to triage fracture plain films pre-radiologist but it hasn’t worked out that way.
 
You must log in or register to reply here.

Similar threads

A
AI assisted mammogram readings
Replies
8
Views
2K
Melkor
M
M
Radiologists, at what point in training did you know you made the right (or even wrong) specialty choice?
Replies
18
Views
7K
hopefulraddaddypls
H
A
Question for radiologists
Replies
5
Views
3K
maxxor
M
GDX
Job opportunity for Radiologists (remote work)
Replies
7
Views
5K
Cognovi
Cognovi
P
Med-Gemini beats Radiologists
Replies
4
Views
4K
Radology
R