- Joined
- Oct 6, 2018
- Messages
- 326
- Reaction score
- 442
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. “
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. “
Last edited:
