I find this fundamentally pessimistic. I would wage that within 20 years AI will solve 80%+ of these issues and will be the dominant form of reading imagery.
That's a long timeline for not much progress! If AI has only solved 80% of these problems in 20 years, we're in a much slower world than probably most people reading this blog are anticipating.
I think though that everyone else is more interested in the question of, "Will AI reduce the demand for radiologists on a five year timeline, not a 20 year one," though.
>Human radiologists spend a minority of their time on diagnostics and the majority on other activities, like talking to patients and fellow clinicians.
This really depends on the radiologist. The 2012 study mentioned was only looking at staff radiologists at hospitals. However, I'm not sure this is representative of the majority of radiologists; many nowadays work fully remotely, reviewing images with essentially zero patient interaction and limited interaction with other clinicians.
I do strongly believe that Jevons paradox should inform our expectations when it comes to the impact of AI on the labor market, but I'm not convinced it's useful for explaining this specific situation. The regulatory and reimbursement hurdles discussed seem much more influential, and we should expect professionals lacking these moats (such as software developers) to be impacted differently.
I expect a couple of words need to be added here: "But demand for human labor is higher than ever [so far]." And here will be the reason: "average income . . . over 48 percent higher than the average salary in 2015."
This was a clear and balanced article but misses one fundamental truth about AI in radiology that is not reflected in the AI literature - in the real world, very few of these clinical applications work well at all. The most successful AI software in radiology is either used to improve image quality (AI-based image reconstruction for CT and MR) or triage (look for asymmetry on a brain CT, for example, and push anything with asymmetry to the top the list - i.e. very primitive algorithms but ones that are reproducible). There is not a single application for diagnostic purposes that is worth spending money on.
I am speaking as an AI-enthusiast who has extensively tested many of these applications. They do not work. One might suspect that in 2025 we would at least have an AI-based CAD to detect pulmonary nodules on a lung cancer screening CT that is effective. It is a matter of finding a white dot on a black background, made a bit more challenging by vessels or bronchi. Companies have been working on this since 2000 but the lung CAD applications today are barely more advanced than what I beta tested in 2002. Certainly there is no product on the market that I would advocate spending any money for.
I have to laugh when I read about how a certain model is extremely accurate in diagnosing a disease, because the widespread practice of testing these in controlled environments with curated data does not translate to the real world. Truly it is a case of the literature (and I believe one can reasonably say that the AI literature is relatively shoddy and not as reliable as other disciplines) not matching real-world outcomes at all. One gets the sense that researchers are living in a bubble.
That is why demand for radiologists has not dropped.
As someone else mentioned, the widest adopted most successful "AI" tools in radiology are neural net-based image reconstructions, particularly for MRI, but other modalities (CT, PET) are adopting them too. These tools allow you to roughly cut imaging time in half with similar quality. Since the scans are shorter, you can do more of them, this has understandable helped increase imaging volumes and made the radiologists shortage worse. These tools have had the fastest time from research papers, to commercial product, to widespread deployment that I have ever seen.
This is just the latest development that has helped MRIs (and other modalities) to get cheaper, dramatically better, and dramatically faster. All of which has increased utilization. Radiology is already in the middle of Jevon's Paradox, and every indication is that efficiency gains from AI will only accelerate this.
The diffusion of these "narrow AI" algorithms is not a good analogy for the diffusion of general AI. The latter is designed to be a plug-in replacement for a human worker. Almost by definition, it will (if achieved) be able to diffuse much more quickly.
This analysis is fundamentally wrong. The writers expectations, like most people, are for linear AI development. It's our mature to assume this. Technology is exponential in development and we only know this in hindsight when the evidence cannot be denied.
A Radiology career is 20 to 30 years. AI capability doubles every 6 months to 1 year. As AI gets more and more intelligence (models get better) AND as AI's get 'plugged in' to the huge volume of every day scans and interpretation of these scans by human radiologists, the AI gets smarter and more independent. It's not if the AI will take over as an independent entity in Xray and MRI (and similar technologies), it a matter of when. And when is sometime in the next 5 to 10 years.
The cycle is predictable. First as an aid to the radiologist, then lead with radiologist review. The beginning of the end of this process is AI all the way for 80% of 'routine' work with 20% sent to 'elite' radiologists for review and the end is 99.6% of the work as AI without review.
And insurance companies will support this effort enthusiastically. 90% lower cost, higher equipment utilization, standardized results and deskilling of necessary human labor. Like current medicine there will be mistakes and occasional tragedies but these are just a cost of business and with AI this cost is less.
=====
Those of us who have been around awhile have seen this happen in many areas. In the 1960s and 1970s acres of clerks in insurance offices and similar paper processing businesses (banks, wall street, and various back office operations) were reduced in stages over time where we now have an app and less than 1% of prior staffing. Same with machines replacing the factory workers. The UAW in auto manufacturing work has had a greater than 80% reduction in membership from the 70s.
And AI is accelerating the rate of change. The best news is that dramatically low cost medical testing can vastly expand the world wide market and the residual jobs for radiologists might be remote workers in America. That will help a bit.
Timothy B. Lee: somehow my reply ended up as a regular entry and not as a reply. Please see my comment that begins "... If the most extreme predictions..." as my reply to your comment
"... If the most extreme predictions about the effect of AI on employment and wages were true, then radiology should be the canary in the coal mine.
But demand for human labor is higher than ever. In 2025, American diagnostic radiology residency programs offered a record 1,208 positions across all radiology specialties, a four percent increase from 2024, and the field’s vacancy rates are at all-time highs...."
This quote from the article and the general tone of the article seemed to me to be a very clear statement that Radiology was a safe career. Most of the article was positive on the tone for radilogy services and the inability for AI to fully grow into a replacement stack for human radiologists.
Radiology is essentially pattern recognition of computer friendly inputs with a high amount of statistical analysis (in the analysis sample and in the training material). Very much an early target for AI because there are no high hurdle interfaces to create or mechanical systems requiring displacement as part of the elimination process. Radiology is an early and relatively easy target that will have strong institutional support for quick adoption (relative) from medical corporations and insurance companies. The need for both insurance companies and medical corporations to lower costs to lead or follow competitors is a virtual guarantee of a quick execution cycle.
This article is very well researched and interesting to read. Thank you.
It follows from the “AI is Normal Technology” crowd. The AI itself is very powerful, but we’re just not seeing the real world impacts of it yet because the technology needs to flow through all of our social systems before it can truly be impactful.
I still believe that AI will take on a bigger role in diagnoses, but it will take much longer to get there.
"Out of the models in 2024 that reported the number of sites where they were tested, 38 percent were tested on data from a single hospital."
DNN need vast training data sets. This seems then to be one of the main problems.
The way healthcare is set up prevents us from having large, anonimized, data sets.
As in most European countries scans are mostly paid with public money, data sets (scan + doctor's comments) across the full population should be made available to companies in those countries.
I find this fundamentally pessimistic. I would wage that within 20 years AI will solve 80%+ of these issues and will be the dominant form of reading imagery.
That's a long timeline for not much progress! If AI has only solved 80% of these problems in 20 years, we're in a much slower world than probably most people reading this blog are anticipating.
Sure but it also means a massive reduction in the number of radiologists. Because 20% of the remaining problems are like 5% of the cases.
I think though that everyone else is more interested in the question of, "Will AI reduce the demand for radiologists on a five year timeline, not a 20 year one," though.
>Human radiologists spend a minority of their time on diagnostics and the majority on other activities, like talking to patients and fellow clinicians.
This really depends on the radiologist. The 2012 study mentioned was only looking at staff radiologists at hospitals. However, I'm not sure this is representative of the majority of radiologists; many nowadays work fully remotely, reviewing images with essentially zero patient interaction and limited interaction with other clinicians.
I do strongly believe that Jevons paradox should inform our expectations when it comes to the impact of AI on the labor market, but I'm not convinced it's useful for explaining this specific situation. The regulatory and reimbursement hurdles discussed seem much more influential, and we should expect professionals lacking these moats (such as software developers) to be impacted differently.
Agreed. The I know a couple of radiologists and they are 100% WFH.
I know a bunch of other doctors and they say their radiology consults are basically all remote.
I expect a couple of words need to be added here: "But demand for human labor is higher than ever [so far]." And here will be the reason: "average income . . . over 48 percent higher than the average salary in 2015."
This was a clear and balanced article but misses one fundamental truth about AI in radiology that is not reflected in the AI literature - in the real world, very few of these clinical applications work well at all. The most successful AI software in radiology is either used to improve image quality (AI-based image reconstruction for CT and MR) or triage (look for asymmetry on a brain CT, for example, and push anything with asymmetry to the top the list - i.e. very primitive algorithms but ones that are reproducible). There is not a single application for diagnostic purposes that is worth spending money on.
I am speaking as an AI-enthusiast who has extensively tested many of these applications. They do not work. One might suspect that in 2025 we would at least have an AI-based CAD to detect pulmonary nodules on a lung cancer screening CT that is effective. It is a matter of finding a white dot on a black background, made a bit more challenging by vessels or bronchi. Companies have been working on this since 2000 but the lung CAD applications today are barely more advanced than what I beta tested in 2002. Certainly there is no product on the market that I would advocate spending any money for.
I have to laugh when I read about how a certain model is extremely accurate in diagnosing a disease, because the widespread practice of testing these in controlled environments with curated data does not translate to the real world. Truly it is a case of the literature (and I believe one can reasonably say that the AI literature is relatively shoddy and not as reliable as other disciplines) not matching real-world outcomes at all. One gets the sense that researchers are living in a bubble.
That is why demand for radiologists has not dropped.
As someone else mentioned, the widest adopted most successful "AI" tools in radiology are neural net-based image reconstructions, particularly for MRI, but other modalities (CT, PET) are adopting them too. These tools allow you to roughly cut imaging time in half with similar quality. Since the scans are shorter, you can do more of them, this has understandable helped increase imaging volumes and made the radiologists shortage worse. These tools have had the fastest time from research papers, to commercial product, to widespread deployment that I have ever seen.
This is just the latest development that has helped MRIs (and other modalities) to get cheaper, dramatically better, and dramatically faster. All of which has increased utilization. Radiology is already in the middle of Jevon's Paradox, and every indication is that efficiency gains from AI will only accelerate this.
The diffusion of these "narrow AI" algorithms is not a good analogy for the diffusion of general AI. The latter is designed to be a plug-in replacement for a human worker. Almost by definition, it will (if achieved) be able to diffuse much more quickly.
This analysis is fundamentally wrong. The writers expectations, like most people, are for linear AI development. It's our mature to assume this. Technology is exponential in development and we only know this in hindsight when the evidence cannot be denied.
A Radiology career is 20 to 30 years. AI capability doubles every 6 months to 1 year. As AI gets more and more intelligence (models get better) AND as AI's get 'plugged in' to the huge volume of every day scans and interpretation of these scans by human radiologists, the AI gets smarter and more independent. It's not if the AI will take over as an independent entity in Xray and MRI (and similar technologies), it a matter of when. And when is sometime in the next 5 to 10 years.
The cycle is predictable. First as an aid to the radiologist, then lead with radiologist review. The beginning of the end of this process is AI all the way for 80% of 'routine' work with 20% sent to 'elite' radiologists for review and the end is 99.6% of the work as AI without review.
And insurance companies will support this effort enthusiastically. 90% lower cost, higher equipment utilization, standardized results and deskilling of necessary human labor. Like current medicine there will be mistakes and occasional tragedies but these are just a cost of business and with AI this cost is less.
=====
Those of us who have been around awhile have seen this happen in many areas. In the 1960s and 1970s acres of clerks in insurance offices and similar paper processing businesses (banks, wall street, and various back office operations) were reduced in stages over time where we now have an app and less than 1% of prior staffing. Same with machines replacing the factory workers. The UAW in auto manufacturing work has had a greater than 80% reduction in membership from the 70s.
And AI is accelerating the rate of change. The best news is that dramatically low cost medical testing can vastly expand the world wide market and the residual jobs for radiologists might be remote workers in America. That will help a bit.
Did Mousa make a prediction about the future of radiology? I missed that part of the article.
Timothy B. Lee: somehow my reply ended up as a regular entry and not as a reply. Please see my comment that begins "... If the most extreme predictions..." as my reply to your comment
https://www.youtube.com/watch?v=aYppKInSS4I&t=345s
This YouTube video starting at 5:45 has a couple of credible minutes on this topic and very supportive of my point of view.
Just here as an FYI.
"... If the most extreme predictions about the effect of AI on employment and wages were true, then radiology should be the canary in the coal mine.
But demand for human labor is higher than ever. In 2025, American diagnostic radiology residency programs offered a record 1,208 positions across all radiology specialties, a four percent increase from 2024, and the field’s vacancy rates are at all-time highs...."
This quote from the article and the general tone of the article seemed to me to be a very clear statement that Radiology was a safe career. Most of the article was positive on the tone for radilogy services and the inability for AI to fully grow into a replacement stack for human radiologists.
Radiology is essentially pattern recognition of computer friendly inputs with a high amount of statistical analysis (in the analysis sample and in the training material). Very much an early target for AI because there are no high hurdle interfaces to create or mechanical systems requiring displacement as part of the elimination process. Radiology is an early and relatively easy target that will have strong institutional support for quick adoption (relative) from medical corporations and insurance companies. The need for both insurance companies and medical corporations to lower costs to lead or follow competitors is a virtual guarantee of a quick execution cycle.
This article is very well researched and interesting to read. Thank you.
It follows from the “AI is Normal Technology” crowd. The AI itself is very powerful, but we’re just not seeing the real world impacts of it yet because the technology needs to flow through all of our social systems before it can truly be impactful.
I still believe that AI will take on a bigger role in diagnoses, but it will take much longer to get there.
"Out of the models in 2024 that reported the number of sites where they were tested, 38 percent were tested on data from a single hospital."
DNN need vast training data sets. This seems then to be one of the main problems.
The way healthcare is set up prevents us from having large, anonimized, data sets.
As in most European countries scans are mostly paid with public money, data sets (scan + doctor's comments) across the full population should be made available to companies in those countries.