Radiology will be the most likely specialty to get affected by AI per ChatGPT

[Stone Cold]

Below is the response from ChatGPT

Question: Hypothetically, which medical specialty is most likely to be replaced by AI given the recent advancements?



ChatGPT Response:

It's important to note that no medical specialty can be fully replaced by AI. However, certain aspects of medical specialties can be automated using AI, allowing healthcare professionals to focus on more complex tasks and improve patient care.

That being said, some medical specialties may see more AI integration than others. Radiology is one specialty that is already heavily utilizing AI. AI algorithms can analyze medical images and detect abnormalities, allowing radiologists to make more accurate diagnoses.

Pathology is another specialty that is starting to see more AI integration. AI can analyze tissue samples and help pathologists make more accurate diagnoses.
Other specialties that could potentially benefit from AI integration include cardiology (for analysis of heart rhythms and ECGs), neurology (for analysis of brain scans), and ophthalmology (for analysis of retinal scans).

However, it's important to note that AI is not a replacement for healthcare professionals. Rather, it's a tool that can help them provide better care. Ultimately, healthcare is a human-driven industry, and the knowledge, skills, and empathy of healthcare professionals are essential for providing quality patient care.

---

Members don't see this ad.

 

[Mo991]

"Neurology for analysis of brain scans"
Tells you a lot about chatgpt no?
You should try chatgpt in a language other than english. You will see how it fails big time on simple questions.
meaning: it is a language model, nothing more. No "understanding" is going in the background.

 

[MetroPath]

Below is the response from ChatGPT

Question: Hypothetically, which medical specialty is most likely to be replaced by AI given the recent advancements?



ChatGPT Response:

It's important to note that no medical specialty can be fully replaced by AI. However, certain aspects of medical specialties can be automated using AI, allowing healthcare professionals to focus on more complex tasks and improve patient care.

That being said, some medical specialties may see more AI integration than others. Radiology is one specialty that is already heavily utilizing AI. AI algorithms can analyze medical images and detect abnormalities, allowing radiologists to make more accurate diagnoses.

Pathology is another specialty that is starting to see more AI integration. AI can analyze tissue samples and help pathologists make more accurate diagnoses.
Other specialties that could potentially benefit from AI integration include cardiology (for analysis of heart rhythms and ECGs), neurology (for analysis of brain scans), and ophthalmology (for analysis of retinal scans).

However, it's important to note that AI is not a replacement for healthcare professionals. Rather, it's a tool that can help them provide better care. Ultimately, healthcare is a human-driven industry, and the knowledge, skills, and empathy of healthcare professionals are essential for providing quality patient care.

View attachment 368321

Check this company out:

Medical Imaging Cloud AI for Radiology | Arterys

And this:

Cardiac Care Pathway Management | Mpirik

Learn more about how Cardiac Intelligence® software utilizes artificial intelligence to identify patients at risk with cardiac disease.

www.mpirik.com

 

[vm26]

Below is the response from ChatGPT

Question: Hypothetically, which medical specialty is most likely to be replaced by AI given the recent advancements?



ChatGPT Response:

It's important to note that no medical specialty can be fully replaced by AI. However, certain aspects of medical specialties can be automated using AI, allowing healthcare professionals to focus on more complex tasks and improve patient care.

That being said, some medical specialties may see more AI integration than others. Radiology is one specialty that is already heavily utilizing AI. AI algorithms can analyze medical images and detect abnormalities, allowing radiologists to make more accurate diagnoses.

Pathology is another specialty that is starting to see more AI integration. AI can analyze tissue samples and help pathologists make more accurate diagnoses.
Other specialties that could potentially benefit from AI integration include cardiology (for analysis of heart rhythms and ECGs), neurology (for analysis of brain scans), and ophthalmology (for analysis of retinal scans).

However, it's important to note that AI is not a replacement for healthcare professionals. Rather, it's a tool that can help them provide better care. Ultimately, healthcare is a human-driven industry, and the knowledge, skills, and empathy of healthcare professionals are essential for providing quality patient care.

View attachment 368321

Eventually AI will replace us but by that time, there will be no pilots, bus & truck drivers, cashiers, bank tellers, realtors, pharmacists etc...also I see the mid-level + AI combo replacing physicians in the clinical setting before it replaces radiologists.

 

[confused_doc]

I definitely see AI and midlevels replacing most physicians with a head physician within each specialty taking most of the liability. But you're deluded if you don't realise radiology, then pathology will be the first specialties to be majorly automated by AI. Hiwever, they'll always be a few radiologist to take on the liability and to interpret the complex cases that AI can't interpret.

 

[SeisK]

AI just found a new publicity gimmick in the form of NLP. The hurdles AI has to overcome to replace a radiologist haven’t changed, and ChatGPT being a thing didn’t make existing AI algorithms any better at their shortcomings. It’s just garnered more public attention, so people are talking about it more, but even ChatGPT doesn’t stand up to scrutiny - it comes to false conclusions, generalizes, and when called out on its mistakes, lies and fabricates sources. The analogy is an image-generating algorithm coming up with a beautiful appearing renaissance painting, but the continuity of body surfaces at overlying interfaces don’t line up, body proportions are slightly off, and they have six free-floating fingers. It would be like asking “why does this thing have six fingers” and the app telling you “humans have six fingers, here’s my source: ___ “.

More handwaving bloviating and screeching on all sides, and yet most importantly STILL no clinical trial results. It’s great at approximating subconscious processes, but not a prefrontal cortex. And medicine is like 80% PFC.

 

[vm26]

I definitely see AI and midlevels replacing most physicians with a head physician within each specialty taking most of the liability. But you're deluded if you don't realise radiology, then pathology will be the first specialties to be majorly automated by AI. Hiwever, they'll always be a few radiologist to take on the liability and to interpret the complex cases that AI can't interpret.

Guess we will see how/when the dominos fall...

 

[qxrt]

I definitely see AI and midlevels replacing most physicians with a head physician within each specialty taking most of the liability. But you're deluded if you don't realise radiology, then pathology will be the first specialties to be majorly automated by AI. Hiwever, they'll always be a few radiologist to take on the liability and to interpret the complex cases that AI can't interpret.

I don't get it. Here we are, sitting around while you guys claim for years and years that AI will take over radiology. And yet, here we are, nowhere near that, while midlevels are taking over huge swaths of emergency medicine/primary care, non-surgical specialties are impressed by ChatGPT giving solid recommendations for how to manage, say, GI bleeds, etc. Show me the money or GTFO. Stop making promises about things that apparently are perpetually in a state of "happening within the next few years." It literally costs us nothing to sit around waiting for you to show us a working example instead of blowing hot air year after year.

Remember when Geoffrey Hinton, apparently one of the so-called fathers of AI, said in 2016 that "“We should stop training radiologists now. It's just completely obvious that within five years, deep learning is going to do better than radiologists.” Yeah, turns out he may be knowledgable in AI, but he knows very little about the practice of radiology.

Is talk and promises all AI can deliver in radiology?

 

[SeisK]

I don't get it. Here we are, sitting around while you guys claim for years and years that AI will take over radiology. And yet, here we are, nowhere near that, while midlevels are taking over huge swaths of emergency medicine/primary care, non-surgical specialties are impressed by ChatGPT giving solid recommendations for how to manage, say, GI bleeds, etc. Show me the money or GTFO. Stop making promises about things that apparently are perpetually in a state of "happening within the next few years." It literally costs us nothing to sit around waiting for you to show us a working example instead of blowing hot air year after year.

Remember when Geoffrey Hinton, apparently one of the so-called fathers of AI, said in 2016 that "“We should stop training radiologists now. It's just completely obvious that within five years, deep learning is going to do better than radiologists.” Yeah, turns out he may be knowledgable in AI, but he knows very little about the practice of radiology.

Is talk and promises all AI can deliver in radiology?

My impression is that a whole bunch of people realized the mistake they made when they decided NOT to go into it, even when it wasn’t hard to get into, and their method of overcoming their cognitive dissonance is by assuring themselves and everyone else that not going into it was the right decision because doom is right around the corner.

Has been for the past decade. Will always be. It’s like the Flying Spaghetti Monster, I can’t disprove AI won’t undergo some miraculous change in the next five years that suddenly makes it better than everyone at everything.

 

[point542]

I don't get it. Here we are, sitting around while you guys claim for years and years that AI will take over radiology. And yet, here we are, nowhere near that, while midlevels are taking over huge swaths of emergency medicine/primary care, non-surgical specialties are impressed by ChatGPT giving solid recommendations for how to manage, say, GI bleeds, etc. Show me the money or GTFO. Stop making promises about things that apparently are perpetually in a state of "happening within the next few years." It literally costs us nothing to sit around waiting for you to show us a working example instead of blowing hot air year after year.

Remember when Geoffrey Hinton, apparently one of the so-called fathers of AI, said in 2016 that "“We should stop training radiologists now. It's just completely obvious that within five years, deep learning is going to do better than radiologists.” Yeah, turns out he may be knowledgable in AI, but he knows very little about the practice of radiology.

Is talk and promises all AI can deliver in radiology?

Totally agreed that midlevels + AI would encroach significantly on clinical specialties before it does on DR. Many fields already have a growing midlevel problem. Once they are aided with AI, it will significantly worsen encroachment.

 

[radrad5]

AI just found a new publicity gimmick in the form of NLP. The hurdles AI has to overcome to replace a radiologist haven’t changed, and ChatGPT being a thing didn’t make existing AI algorithms any better at their shortcomings. It’s just garnered more public attention, so people are talking about it more, but even ChatGPT doesn’t stand up to scrutiny - it comes to false conclusions, generalizes, and when called out on its mistakes, lies and fabricates sources. The analogy is an image-generating algorithm coming up with a beautiful appearing renaissance painting, but the continuity of body surfaces at overlying interfaces don’t line up, body proportions are slightly off, and they have six free-floating fingers. It would be like asking “why does this thing have six fingers” and the app telling you “humans have six fingers, here’s my source: ___ “.

More handwaving bloviating and screeching on all sides, and yet most importantly STILL no clinical trial results. It’s great at approximating subconscious processes, but not a prefrontal cortex. And medicine is like 80% PFC.

dude, that miraculous change just happened. chat gpt. you are totally underplaying this. computers are good at pattern recognition. chat gpt can describe images. what human rads do is use pattern recognition to describe images.

radiology will be dead in 10-15 years. so will a lot of other white collar jobs. we will benefit from regulatory hurdles etc but c'mon man it seems obvious.

 

[SeisK]

dude, that miraculous change just happened. chat gpt. you are totally underplaying this. computers are good at pattern recognition. chat gpt can describe images. what human rads do is use pattern recognition to describe images.

radiology will be dead in 10-15 years. so will a lot of other white collar jobs. we will benefit from regulatory hurdles etc but c'mon man it seems obvious.

no, it really didn’t. it’s an NLP gimmick. AI‘s inability to recognize a C-spine fracture or appendicitis with high sensitivity or specificity remains unchanged by GLP being trained to have what seems like a conversation with you.

Literally everything you said after the first “chat gpt” is what people were saying 10 years ago.

I’m frustrated by my inability to teach chatgpt that ”hyperintense on DWI” does not mean “increased diffusion.”

 

[CallMeBigJeff]

I definitely see AI and midlevels replacing most physicians with a head physician within each specialty taking most of the liability. But you're deluded if you don't realise radiology, then pathology will be the first specialties to be majorly automated by AI. Hiwever, they'll always be a few radiologist to take on the liability and to interpret the complex cases that AI can't interpret.

You're the guy hyping up AI in that other thread. Are you a radiologist bro? Because based on your replies you don't offer specific insight as to why one would "be deluded to not realized radiology will be one of the first specialties automated by AI". Blanket statements like "interpret the complex cases" is hilarious because almost every other study has some **** limitation (poor positioning, motion, artifact, prior with no available report or one of the aforementioned limitations) that you need a radiologist to use their brain for.

 

[confused_doc]

You're the guy hyping up AI in that other thread. Are you a radiologist bro? Because based on your replies you don't offer specific insight as to why one would "be deluded to not realized radiology will be one of the first specialties automated by AI". Blanket statements like "interpret the complex cases" is hilarious because almost every other study has some **** limitation (poor positioning, motion, artifact, prior with no available report or one of the aforementioned limitations) that you need a radiologist to use their brain for.

The reason why radiology is most at risk is simply because the images are already digitized so out of all the specialties in medicine it has the largest volume of structured data to train AI models on.

And absolutely every study has a limitation, I build models for image segmentation of medical images, and we regularly encounter differences in model performance when something as trivial as different scanners are used for the exact same dataset. But as mentioned before my work is to build algorithms that force the model to understand the deep features that make up images regardless of any rotations, differences in color hue, cropping of images etc. It's a field of active research in ML and every day there seems to be another breakthrough increasing the generalisability of these models.

 

[Steve_Zissou]

The reason why radiology is most at risk is simply because the images are already digitized so out of all the specialties in medicine it has the largest volume of structured data to train AI models on.

And absolutely every study has a limitation, I build models for image segmentation of medical images, and we regularly encounter differences in model performance when something as trivial as different scanners are used for the exact same dataset. But as mentioned before my work is to build algorithms that force the model to understand the deep features that make up images regardless of any rotations, differences in color hue, cropping of images etc. It's a field of active research in ML and every day there seems to be another breakthrough increasing the generalisability of these models.

I don’t think the first question was answered. Are you a radiologist?

 

[Neopolymath]

I don’t think the first question was answered. Are you a radiologist?

Post history says former UK trainee in another field.

 

[confused_doc]

I don’t think the first question was answered. Are you a radiologist?

Nope not a radiologist. I'm a medical doctor and AI researcher who builds computer vision models for a living.

 

[confused_doc]

Post history says former UK trainee in another field.

I'm actually US based now.

 

[bronx43]

Nope not a radiologist. I'm a medical doctor and AI researcher who builds computer vision models for a living.

What kind of timeline are you predicting for AI takeover of radiology and clinical medicine?

 

[Steve_Zissou]

The reason why radiology is most at risk is simply because the images are already digitized so out of all the specialties in medicine it has the largest volume of structured data to train AI models on.

And absolutely every study has a limitation, I build models for image segmentation of medical images, and we regularly encounter differences in model performance when something as trivial as different scanners are used for the exact same dataset. But as mentioned before my work is to build algorithms that force the model to understand the deep features that make up images regardless of any rotations, differences in color hue, cropping of images etc. It's a field of active research in ML and every day there seems to be another breakthrough increasing the generalisability of these models.

It seems to me that pathology or dermatology would go first, given that they mainly work with 2d models which would likely be much easier to interpret than the 3D ones that rads works with.

Also, I’ll take any non-radiologist take with a grain of salt since it’s really difficult to understand the complexities of the field (any medical field) unless you’re immersed in it for quite a while.

I’m just an intern though, so it’s all guesses for me

 

[bronx43]

It seems to me that pathology or dermatology would go first, given that they mainly work with 2d models which would likely be much easier to interpret than the 3D ones that rads works with.

Also, I’ll take any non-radiologist take with a grain of salt since it’s really difficult to understand the complexities of the field (any medical field) unless you’re immersed in it for quite a while.

I’m just an intern though, so it’s all guesses for me

To be fair, every field has immense complexity if you get down into the very fine granular scale of diagnosis and treatment. But, the way the system is going, it doesn’t look like such perfection matters too much. A midlevel isn’t as good for anesthesia or clinical medicine but the system puts up with their shortcomings. Everything has devolved into “good enough.”

Theoretically if a ML algorithm can be 95% as good as a human for imaging then I would bet money the corporatists would run with it and gladly accept whatever medicolegal repercussions arise from the 5% imperfection.

 

[Steve_Zissou]

To be fair, every field has immense complexity if you get down into the very fine granular scale of diagnosis and treatment. But, the way the system is going, it doesn’t look like such perfection matters too much. A midlevel isn’t as good for anesthesia or clinical medicine but the system puts up with their shortcomings. Everything has devolved into “good enough.”

Theoretically if a ML algorithm can be 95% as good as a human for imaging then I would bet money the corporatists would run with it and gladly accept whatever medicolegal repercussions arise from the 5% imperfection.

The issue is not having a good grasp of what the field really entails to produce good results. I sure as **** didn’t have a good idea of what medicine really entailed to do it semi-competently until this year when I’ve spent 6 months on wards, 6 days a week, for 12 hours each day. I doubt someone without formal radiology training has a real understanding of what is required for a good read and the amount of variables that are accounted for in each read, but I could be wrong.

Probably, but I remain of the opinion that there’s no need to worry until we start seeing it take pathologist and dermatologist jobs.

 

[confused_doc]

What kind of timeline are you predicting for AI takeover of radiology and clinical medicine?

I'll make a detailed comment once I'm done with work. Also just to be clear I haven't actually claimed that AI is going to takeover anything, in the other thread I said if it was to takeover, all specialties would be replaced, but radiology would be one of the specialties to be greatly automated first and that's because of the huge head start in investment in computer vision research applied to radiology as well as the sheer volume of digitized data in radiology with which to train the models.

Also, I’ll take any non-radiologist take with a grain of salt since it’s really difficult to understand the complexities of the field (any medical field) unless you’re immersed in it for quite a while.

Sure, but by your same reasoning you have to take a radiologist's opinion with a grain of salt as most of them don't work in AI or have a deep understanding of the algorithms.

 

[confused_doc]

The issue is not having a good grasp of what the field really entails to produce good results. I sure as **** didn’t have a good idea of what medicine really entailed to do it semi-competently until this year when I’ve spent 6 months on wards, 6 days a week, for 12 hours each day. I doubt someone without formal radiology training has a real understanding of what is required for a good read and the amount of variables that are accounted for in each read, but I could be wrong.

Probably, but I remain of the opinion that there’s no need to worry until we start seeing it take pathologist and dermatologist jobs.

Also just a quick aside it's way more difficult for computers to understand color images than it is for them to understand black and white images.

 

[Steve_Zissou]

I'll make a detailed comment once I'm done with work. Also just to be clear I haven't actually claimed that AI is going to takeover anything, in the other thread I said if it was to takeover, all specialties would be replaced, but radiology would be one of the specialties to be greatly automated first and that's because of the huge head start in investment in computer vision research applied to radiology as well as the sheer volume of digitized data in radiology with which to train the models.


Sure, but by your same reasoning you have to take a radiologist's opinion with a grain of salt as most of them don't work in AI or have a deep understanding of the algorithms.

I agree, I have very minuscule knowledge on AI and algorithms (Dennetts Consciousness explained and the UW online AI in radiology coursework that I did on my own).

A radiologist, who has no experience with AI, likely won’t have the best take but likely a better take than the AI person with little to no radiology experience. I say that because this is AI trying to solve a radiology problem, not radiology trying to solve an AI problem.

I mean, remember 2016 when Hinton said AI would replace radiologists within 5 years and that they should stop training radiologists? There’s a good example of an AI expert with little radiology knowledge being completely wrong. So they were either wrong because they knew too little about AI, or wrong because they know very little about radiology and the complexity of it.



I do think it’ll change practice significantly within the next decade, but the chicken littles that pop up here and on Reddit get a little tiresome.

 

[Steve_Zissou]

The real question is, when will AI replace programmers?

ChatGPT Will Replace Programmers Within 10 Years

Predicting The End of Manmade Software

levelup.gitconnected.com

 

[radrad5]

I'll make a detailed comment once I'm done with work. Also just to be clear I haven't actually claimed that AI is going to takeover anything, in the other thread I said if it was to takeover, all specialties would be replaced, but radiology would be one of the specialties to be greatly automated first and that's because of the huge head start in investment in computer vision research applied to radiology as well as the sheer volume of digitized data in radiology with which to train the models.


Sure, but by your same reasoning you have to take a radiologist's opinion with a grain of salt as most of them don't work in AI or have a deep understanding of the algorithms.

curious about your timeline prediction

- nervous trainee with large burden of student loans...

 

[vm26]

curious about your timeline prediction

- nervous trainee with large burden of student loans...

This is all tricky to predict. I understand the rationale behind the belief that AI will replace radiologists 1st but one can speculate that a field like pathology is more susceptible as there is essentially no interaction with clinicians. AI + mid-levels in the ED/primary care setting could also go 1st.

My group has been using AIDoc which is decent but fairly limited in how it can be used. The AI I would love to have now is one the auto-populates my report with appropriate f/u guidelines for any incidentaloma I find (after cross-checking with pts demographics/risk factors from the EMR and checking with ACR guidelines etc), or one that automatically communicates critical findings and documents this in my report (and is accepted from a medico-legal standpoint) before I am done dictating. That would be awesome and time-saving

 

[radrad5]

i think even with an (AI) (very) optimistic timeline its still probably 10 years out

- 2-5 years to actually develop the technology capable of interpreting imaging across all modalities
- 2-5 years of implementation, prospective trials
- 2-5 year phase of AI+human

hard to imagine all of that happening in <10 years, but i am far from an expert

 

[vm26]

i think even with an (AI) (very) optimistic timeline its still probably 10 years out

- 2-5 years to actually develop the technology capable of interpreting imaging across all modalities
- 2-5 years of implementation, prospective trials
- 2-5 year phase of AI+human

hard to imagine all of that happening in <10 years, but i am far from an expert

A major hurdle or point of no return will be when AI is held to a legal standard equal to or greater than that of a human radiologist.

 

[qxrt]

i think even with an (AI) (very) optimistic timeline its still probably 10 years out

- 2-5 years to actually develop the technology capable of interpreting imaging across all modalities
- 2-5 years of implementation, prospective trials
- 2-5 year phase of AI+human

hard to imagine all of that happening in <10 years, but i am far from an expert

The bolded alone is very unlikely to happen within the next 10 years. The ability for AI to "interpret" CT/MR/US or even "just" radiographs is vastly overestimated. I say this as someone working in a group actively incorporating AI technology like Aidoc into our workflow.

 

[SeisK]

The bolded alone is very unlikely to happen within the next 10 years. The ability for AI to "interpret" CT/MR/US or even "just" radiographs is vastly overestimated. I say this as someone working in a group actively incorporating AI technology like Aidoc into our workflow.

Clinical trials are also pathology specific. AI is good for some things, it is hilariously bad for others. Therefore the 2-5 years of tech development, then 2-5 years for trial implementation, while that timeline is presumably true, it is probably only true for specific pathologies, and needs to be tested on specific pathologies only.

Of which there are thousands to tens of thousands, some of which AI will be easy to implement, and others very much not so.

This will not be a fast process, even if implementation is straightforward, which it won’t be.

 

[Clean7795]

I would argue that other specialties with significant independent/semi-independent midlevel encroachment are most at risk. They have demonstrated that physician-level expertise is only needed in specific cases, and as such the AI only needs to be better than a NP.

In radiology, most diseases require either general intelligence or the ability to compare to prior exams/other modalities. Technological advancement past these barriers is not an inevitability. Self-driving cars cannot advance past the heterogeneity in real-life driving environments.