Forgot your password?
Close
typodupeerror
AI

CEO of America's Largest Public Hospital System Says He's Ready To Replace Radiologists With AI (radiologybusiness.com) 89

Posted by BeauHD from the game-changer dept.
Mitchell H. Katz, MD, president and CEO of NYC Health + Hospitals, said hospitals could already replace many radiologists with AI for some imaging tasks -- if regulators allowed it. He argued the technology presents an opportunity to simultaneously cut costs and expand access. Radiology Business reports: Katz -- who has led the 11-hospital organization since 2018 -- said he sees great potential for AI to increase access to breast cancer screening. Hospitals could potentially produce "major savings" by letting the technology handle first reads, with radiologists then double-checking any abnormal screenings. Fellow panelist David Lubarsky, MD, MBA, president and CEO of the Westchester Medical Center Health Network, said his system is already seeing great success in deploying such technology. The AI Westchester uses misses very few breast cancers and is "actually better than human beings," he told the audience. "For women who aren't considered high risk, if the test comes back negative, it's wrong only about 3 times out of 10,000," Lubarsky said.

Katz asked fellow hospital CEOs if there is any reason why they shouldn't be pushing for changes to New York state regulations, allowing AI to read images "without a radiologist," Crain's reported. In this scenario, rads could then provide second opinions, if AI flags any images as abnormal. Sandra Scott, MD, CEO of the One Brooklyn Health, a small hospital facing tight margins, agreed with this line of thinking, according to Crain's. "I mean, I'm in charge of a safety-net institution. It would be a game-changer," Scott said about AI being used to replace rads.
This discussion has been archived. No new comments can be posted.

CEO of America's Largest Public Hospital System Says He's Ready To Replace Radiologists With AI More

CEO of America's Largest Public Hospital System Says He's Ready To Replace Radiologists With AI

Comments Filter:

  • Yep, and WSJ as well.

  • Radiologists (Score:5, Funny)

    by liqu1d ( 4349325 ) on Wednesday April 01, 2026 @12:10PM (#66072214)
    Say they're ready to replace their CEO with AI. Patient outcomes have improved significantly. Shareholders are crying.
    • Depends on who trained the Ai... Some of us expect ai to be better than humans, more rational, but it has the same weaknesses as we do. It all depends on how it is brought up. It is an interesting philosophical theme that leads to questioning what the meaning is of life, the universe and everything. If it is trained well, it will say 42. In a few seconds instead of billions of years.
      • That's a joke right? :/

      • This technology fabricates citations to non-existent medical research. There were news headlines about it just a year ago.

        People who expect it to be "more rational" will get what they deserve. But their customers deserve better.

        • You do understand progress right?

          • You do understand progress right?

            Right. Now it fabricates unjustified citations to real existing medical research so that you won't manage to catch it or disprove the references. We understand progress.

        • The machine learning systems being trained on radiology scans are fundamentally different from LLMs. It gives a list of possible diagnoses with confidence percentages. Mind you, that doesn't mean it can be trusted to replace a human, but these systems don't make stuff up the way LLMs do. It's closer to image categorization.

    • Re: (Score:3)

      by whitroth ( 9367 )

      Absolutely. For the replacement of him with AI, they can hire a number more radiologists, and nurses, and doctors...

    • Re:Radiologists (Score:5, Insightful)

      by geekmux ( 1040042 ) on Wednesday April 01, 2026 @12:32PM (#66072260)

      Shareholders are crying.

      Really?

      If we're looking for an actual downside here, fire all the radiologists and put CEOs in their place to be personally liable for ALL diagnostic readings until AI gets it perfect enough to be defended 100% in every court case.

      Perhaps then we'll see how much of a loophole "AI" is with regards to dismissing a Recession.

      • Re: (Score:3)

        by Tyr07 ( 8900565 )

        Yeah I like this one. If a human gets it wrong they can be liable. So if we replace it with AI, let's make the CEO liable. Fantastic idea.

      • Dream on. We can't even make Waymo officers responsible for their robotaxis zooming past stopped school buses.

      • I was thinking along the same lines...

        I'd say let the CEO do it. And when it goes south - which it will - make sure the CEO is held personally accountable for any misdiagnoses and deaths.

      • Shareholders are crying.

        Really?

        If we're looking for an actual downside here, fire all the radiologists and put CEOs in their place to be personally liable for ALL diagnostic readings until AI gets it perfect enough to be defended 100% in every court case.

        Perhaps then we'll see how much of a loophole "AI" is with regards to dismissing a Recession.

        It will just be another line on the forms you sign, like acknowledging the risk of the radiation dose you're signing up for, or the high powered magnets vs. metal stuff in your body, or the contrast enhancing stuff they inject you with, or all the other risks.

        I'm not even a lawyer, it's just kind of obvious that you don't have a reasonable expectation of a 100% accurate reading, or zero risk. Your bar has to be somewhere else.

        • Shareholders are crying.

          Really?

          If we're looking for an actual downside here, fire all the radiologists and put CEOs in their place to be personally liable for ALL diagnostic readings until AI gets it perfect enough to be defended 100% in every court case.

          Perhaps then we'll see how much of a loophole "AI" is with regards to dismissing a Recession.

          It will just be another line on the forms you sign, like acknowledging the risk of the radiation dose you're signing up for, or the high powered magnets vs. metal stuff in your body, or the contrast enhancing stuff they inject you with, or all the other risks.

          I'm not even a lawyer, it's just kind of obvious that you don't have a reasonable expectation of a 100% accurate reading, or zero risk. Your bar has to be somewhere else.

          If my bar is somewhere below near-risk-free, then their liability is as well. Did the ones replacing trained humans with AI, actually read the EULA?

          Lawyers gonna lawyer. Any time, any place.

    • Re: (Score:2)

      by HiThere ( 15173 )

      Well, he did say "for some imaging tasks". That's probably a reasonable goal...but you've got to be *very* selective.

    • Shareholders are crying

      Replace them with AI as well.

  • Like “The Cloud” was going to get rid of the in-house computer department, this AI kool-aid is going to massively fail to deliver. A useful tool as long as it is applied in specific use cases and verified by a human. In this case, since you've fired all your radiologists and AIs are prone to hallucinations, you can't be sure its diagnosis is accurate. You would get lots of false positives or false negatives, with no one left to verify the results.

  • My wife had an MRI (Score:1)

    by Anonymous Coward

    My wife had an MRI of her head and neck with contrast. The MRI place gave us the CD the same day, so before we took the CD to her ENT's radiologist, we let claude take a look at it. Claude found the problem in about 15 minutes of analysis, and it was a very reasonable explanation for her symptoms. The radiologist came back in a week and said the MRI was normal. We asked "you mean you didn't see the X?" Look again, and sure enough, he came back in a couple of days saying he saw it.

    Claude found the structural

  • You first! (Score:1)

    by Anonymous Coward

    I'll agree as long as Mitchell H. Katz, MD signs off on AI making binding decisions for all of his and his family's healthcare!

  • So in our new world of irresponsibility and negligence by AI who takes on the liability and pays out the injury awards? Probably, some reverse Centaur...

    There is a book coming out on this topic soon! [macmillan.com]

    We'll all be so happy when we are employed by AI which cannot legally be held liable for all the mistakes it makes and we see everyday.

    • Re: (Score:1)

      by boskone ( 234014 )

      There is a massive shortage of doctors,and we can not affrord to train and hire more.

      To expand care with the same money, we have to innovate and let technology help us scale the very expensive and rare humans we have.

      The question isn't "would it be better to have a highly trained radiologist take the first pass or an AI?" The question is, would I rather screen 100 women with a 90% accuracy rate (human) or 1000 women with an 85% accuracy rate (AI).

      I also believe you can continually train the AI to get bette

      • we can not affrord to train and hire more.

        Do you know how US healthcare costs at least ten times more than in other countries?

        Eliminate from "the system" all the middlemen, all the shenanigans, all the MBA'ing, and all the lobbied-for anti-healthcare laws, and you'll have a lot with which to train (for free) and hire (with excellent wages) more doctors.

  • False Positives Vs False Negatives (Score:5, Insightful)

    by gurps_npc ( 621217 ) on Wednesday April 01, 2026 @12:58PM (#66072306) Homepage

    There are two distinctly different types of errors when it comes to these kind of tests:

    False Positives: This is where the test in question falsely says "You have Cancer!" when in fact you do not have it.

    False Negatives: This is where the test in question falsely says "You are Healthy" when in fact you have cancer.

    False Positives cost money and time, but it is fairly easy to double check them as they should be uncommon.

    False Negatives cost human lives and are almost impossible to double check them as most people should test negative for cancer.

    For an AI test, you want to have false positives. If it saves you money by not requiring humans to look things over, then costing you money and time to double check things is a fair trade. If it costs too much to double check, then do not use the AI.

    False Negatives should be a no no. If the AI has more false negatives than human radiologists do, then do not use the AI test. No one cares how much money you are saving if people are dying.

    Note, with regards to jobs, this will likely be relatively flat. There are not that many humans doing this job - they take the results from radiologist exams from all over the country and send them to just a few companies. Those companies find the few people that do it best and hire them. I bet we are talking about less than a hundred people in the US, especially as the best of the best will be kept to double check the results.

    • This all depends on tuning the algorithms' ROCs. You are advocating for ultra low false negatives, allowing higher false positives. It's unlikely that will pass muster, because it costs more money than higher false negatives. Furthermore, false negatives are money-makers, as future therapy once the cancer progresses is quite lucrative.

      • False negatives are lucrative to the treatment company, Not to the detection company.

        And medical field has some (not all or even most) ethical people. While some people get into it for the money, lots of people get into it for other reasons. Enough ethical people in it are likely to keep the nightmare you propose from coming about.

    • Re:False Positives Vs False Negatives (Score:4, Informative)

      by DamnOregonian ( 963763 ) on Wednesday April 01, 2026 @01:27PM (#66072386)
      False positives, in fact, are the cause of all kinds of human misery- often leading to unneeded surgeries, and sometimes surgeries with very serious consequences.
      They're a part of why there are screening guidelines.

    • Re: (Score:2)

      by Himmy32 ( 650060 )

      If the AI has more false negatives than human radiologists do, then do not use the AI test. No one cares how much money you are saving if people are dying.

      If the standard is "No one cares how much money you are saving if people are dying.", I would argue that the rate of AI only checked false negatives would need to be lower than the rate of AI first read + radiologist check false negatives rather than just radiologist check alone.

    • The shareholders VERY MUCH care about saving money and will kill any number of people that are not themselves to do it!

    • False Positives cost money and time, but it is fairly easy to double check them as they should be uncommon.

      You don't understand the math of false results.

      The population of people who actually should have a negative result is far larger than the population of people that should have a positive result.

      A small percentage of a large population [the actually negative] results in a significant number of false positives, while a small percentage of a small number of people [the actually positive] results in a very small number.

    • False Positive: Type I Error
      False Negative: Type II Error

      • I don't know why anyone would use the "Type I/II" terminology when "False Positive/Negative" are so much more descriptive.

    • Re: (Score:3)

      by ceoyoyo ( 59147 )

      This is too simple. The most important thing for a screening test, like the mammograms they mention, is often not to have too many false positives. False positives cost money, time, pain, suffering, and kill people. Yes, getting a big ass needle shoved into you is pretty safe but it's not 100% safe. The money part also kills people. Money is finite and spending it in one place means you can't spend it somewhere else. Because screening tests have many more true negatives than positives, even a quite low fals

    • No one cares how much money you are saving if people are dying.

      Is this the American health care system you're talking about?

    • "No one cares how much money you are saving if people are dying"
      Well it's a corporation. They can degrade the quality of service until it starts having a negative impact on perceived shareholder value.
      But the CEO is a doctor! You can trust him because he's a doctor.

  • Double checking. Haha. They barely look at the images when single checking. You think they are going to put any effort into double checking? What a joke.

  • [CEO of America's Largest Public Hospital System] argued the technology presents an opportunity to simultaneously cut costs and expand access.

    LOL! More like: "[CEO of America's Largest Public Hospital System] argued the technology presents an opportunity to simultaneously cut staffing costs and expand profits."

  • My Understanding (Score:3)

    by snookiex ( 1814614 ) on Wednesday April 01, 2026 @01:15PM (#66072356) Homepage
    Is that most situations could be more efficiently diagnosed by AI, but there's still a fair share of cases where experienced radiologists would be needed. In any case, I think it's just matter of time before AI can actually replace them. I mean, it's like the perfect use case for it.

  • All radiologists do is analyze digital images (Score:5, Informative)

    by Locke2005 ( 849178 ) on Wednesday April 01, 2026 @01:18PM (#66072366)
    For years now, radiology has been a poor career choice, but it only makes sense to send those digital images to the place with the cheapest doctors. Turns out one thing AI is really much better at than humans is analyzing digital images, so yes, radiology careers will soon be extent. All the job growth is in health care, but it's all in the jobs that require you to be in the same room as the patient. Administrators and back office staff are all getting laid off. The people that clean the rooms, body fluids and all? Hospitals can't get enough of them. (Five of my relatives all work at OSHU. Four are housekeeping, one is a pharmacy technician. They have pharmacy robots now...)
    • Problem with "AI and digital images" is that classifier is only as good as its training data.
      The cost to train goes up dramatically the more training data you give it. This means there is a financial incentive to not train on edge cases, which means your AI *will not catch them.*

      Like all things, it's not the tech I fear. It's the executives in charge of monetizing it.

      • Re: (Score:2)

        by ceoyoyo ( 59147 )

        Physicians are also only as good as their training data. The same financial incentive works there except the "classifier" also has an incentive to resist more training. In fact, despite quite a few incentives to the contrary, there's pretty good evidence that the average physician stops learning new medicine within a few years of graduating medical school. Not their residencies, medical school.

        Hospitals used to maintain extensive pathology libraries, often built by generations of pathologists, where they co

        • Physicians are also only as good as their training data.

          It's inherently different.
          The ImageNet is trained to recognize images, and pair it with contextual information given in its scanning.
          I.e., it is- at best- distilled from the knowledge of rad techs.

          On the other hand, it's pretty easy (i.e. cheap) to train an image classifier on orders of magnitude more cases than any pathologist or radiologist could ever see in their lifetime.

          Indeed. Many ImageNets have been trained on more dogs than I will ever see in my lifetime, and will still- a few percent of the time- call them a duck.

          I find it funny that you start off talking about how doctors don't stay up on medicine (which is very true), and then point out that these aren't LLMs.
          LLMs use

          • Re: (Score:2)

            by ceoyoyo ( 59147 )

            It's inherently different.
            The ImageNet is trained to recognize images, and pair it with contextual information given in its scanning.
            I.e., it is- at best- distilled from the knowledge of rad techs.

            It's not. Some of the very amateur tech bro efforts are, but proper models are built using better gold standards. Typically you build your training set the same way you build the training sets for radiology textbooks, or the baby radiologists evaluate their own decisions: you wait to see how the patient does.

            I fin

      • The types of neural networks used in these classifiers are not LLMs. The expense and difficulty of collecting an image set is going to far outweigh the compute time used to train them.

        • The types of neural networks used in these classifiers are not LLMs.

          Correct.

          The expense and difficulty of collecting an image set is going to far outweigh the compute time used to train them.

          Like any network, how good it is is a mix of how much data you throw at it, and how many parameters it is.
          Large ImageNets cost millions to train.

          • Large ImageNets cost millions to train.

            Which is peanuts compared to the recurring $500-600k cost of a radiologist's annual salary. I don't get the sense that these models are ready to replace radiologists yet, but the closer we get the more tempting it'll be for individuals like the CEO quoted in this article. The ROI is massive.

            • Which is peanuts compared to the recurring $500-600k cost of a radiologist's annual salary.

              It's not peanuts at all, which is precisely why there is also competition in the medical industry to reduce radiologists, and push more work onto larger centralized corporately hosted radiology mills.

              I don't get the sense that these models are ready to replace radiologists yet

              They're not even close.
              LLMs outperform doctors in diagnostics. ImageNets do not outperform them clearly*.

              but the closer we get the more tempting it'll be for individuals like the CEO quoted in this article. The ROI is massive.

              Radiologists today are being "replaced" (in that more work is piled onto less radiologists now augmented by AI)
              That is still replacement.

              * there are some cases where ImageNets do outperform radiologists,

              • Posting to mostly say "thanks" for responding thoughtfully to an article that is now probably well off the front page. Keep it up. It elevates the Slashdot discourse. Hopefully it makes the LLMs scraping the site a bit smarter...

                I'd say we're generally on the same page. My intention was to point out that something that costs millions can still be an obvious choice for a health system if it can replace even a small number of radiologists.

  • so, who gets sued for misdiagnosis? As long as liability remains with them- the CEO for making the decision.

    But also... let me look at crystal ball...
    -Costs go up, not down
    -premium tier charge for human review
    -waivers of liability
    and of course- worse patient care outcomes

    AI is not am effin replacement for life and death decision making - it is to inform and assist. Not Make them. Let me hallucinate lack of cancer... or better yet.. yet me hallucinate a cancer...

    • Re: (Score:2)

      by hwstar ( 35834 )

      The lone Radiologist they kept on staff will be sued. This is an "Accountability Sink", and the lone radiologist will be what's known as a "Reverse Centaur". They'll just use up and wear out all of the fired Radiologists. They'll be plenty of them.

      Cory Doctorow wrote an article exactly about Radiologists going obsolete here:

      https://doctorow.medium.com/https-pluralistic-net-2025-12-05-pop-that-bubble-u-washington-8b6b75abc28e

    • My professor when completing my Masters in Health Informatics (on artificial decision making 2 decades ago) once said "if you don't let computers kill patients, they can't save them". Same for doctors.

      You hope the saving outweighs the killing, and in this case we can probably prove that for AI radiology relatively easily.

  • Hilarious timing (Score:5, Interesting)

    by Tschaine ( 10502969 ) on Wednesday April 01, 2026 @01:42PM (#66072426)

    Just yesterday I stumbled on this substack post about a research paper whose authors found that AI scored well on x-ray evaluations even when the AI took the test WITHOUT ACCESS to the x-ray images.

    https://drjo.substack.com/p/wh... [substack.com]

    The moral of this story is that properly evaluating AI performance in classification tasks requires very very carefully designed tests, because neural nets are very very good at picking up correlations between the desired outputs and utterly unintentional signals in the inputs.

  • Obviously, patients will not pay less (Score:3)

    by gweihir ( 88907 ) on Wednesday April 01, 2026 @01:46PM (#66072434)

    They will just get worse service. Anything else would be un-American.

    • This isn't a crosswalk CAPTCHA.

      AI has been better than humans for space images for over 2 decades. I'm sure they've beaten humans at cancer spotting for quite a while. You just need proper consistent imaging and plenty of it as training data. The unusual bits go to some humans and eventually it will do everything in the area better than humans. Start specific and over time add confirmations -- because they do a biopsy on a positive test result to confirm. That is better than a human verification; it confi

  • will be when health insurance companies are replaced by AI

  • Where are the numbers? How about testing a radiologist + AI, rather than each against the other. I'll bet most people would prefer to have both involved in their care. I think they've skipped over the false negatives bit too. How do you ensure you're not missing a bunch of stuff? You'll need humans to find what the AI's miss. I'm not saying keep all the humans, but I am saying the vision and arguments presented are far from complete, and as usual, focus on the upfront costs rather than the overall cost

  • People who advocate for AI taking over tasks, should be help personally liable for the inevitable failure of the AI system

  • An accountability sink is a system or structure within an organization that obscures or deflects responsibility for decisions, making it difficult to identify who is accountable when things go wrong. This often occurs when decision-making is delegated to complex rules or automated processes, preventing effective feedback and learning from mistakes

    Quoting thoughts from Cory Doctorow:

    So there will be one Radioligist on staff whose job will be to vet and certify all AI determinations. This person will be the "

    • This almost exactly matches the observed behavior from [my comment](https://slashdot.org/comments.pl?sid=23954882&cid=66072694)... which is an anecdote already dating back to 2009.

  • I've worked in medical imaging (Score:5, Informative)

    by Arrogant-Bastard ( 141720 ) on Wednesday April 01, 2026 @03:52PM (#66072676)
    Twice, in fact -- once in an academic research lab and once at a company that designed and built medical imaging equipment.

    In both cases we worked on image classification using digital image processing and statistical pattern recognition. (In one of the two cases we also used syntactic pattern recognition and machine learning.) It's very, very, very hard to make this accurate enough for clinical use even if you pour effort and time and money into it. There's no way this technology should be deployed without humans backing it up.

    As to the human mistakes: everyone can cite a case where a professional radiologist committed a false positive or false negative error. But did you stop to consider why they made a mistake? Were they 13 hours into a 14-hour shift, their third one in a row -- because the hospital CEO felt that money should go into his pocket instead of into hiring another radiologist to share the load? Was it an imaging anomaly (they happen) that was ambiguous? Was it because the study that was done wasn't the best choice? (I.e., imaging modality or location) There are all kinds of ways for this to go wrong that will result in blame being assigned to the radiologist, and only some of those assignments are fair.

    AI isn't a magic fix for this. And I certainly wouldn't even try to use any of the general-purpose models -- as Zathras would say: "This is wrong tool." If I were to do this again today, I would return to the approach we used before with modest success, I'd take advantage of some of the improved algorithms that have come along, and obviously I'd use bigger/faster hardware, because that opens up approaches that were computationally infeasible. But I wouldn't even consider removing humans: these are, or can be, life-and-death decisions, and a human being needs to make them.

  • This isn't exactly new (Score:4, Informative)

    by DrStrangluv ( 1923412 ) on Wednesday April 01, 2026 @04:02PM (#66072694)

    I worked as a programmer at a medical billing company back in 2009, and let me tell you it was eye opening. We had radiologists working remotely (in 2009!) with mutliscreen setups that would show an original image on the left of one screen, a computer-enhanced version on the other side of the screen, with a computer generated opinion pre-generated at the bottom of the image (again: 2009 already had this). The other screen, usually rotated 90 degrees, would show minimal required relevant patient history/demographic on the top and offer a place to enter the radiologists opinion below, along with a button to copy over the computer-generated opinion.

    Let's game out their options.

    Let's say the agree with the computer, and they're right. No extra reward, they're just doing their job.
    Let's say they agree with the computer, and they're wrong. Well, that must have been a hard case. Oh well.
    Let's say they disagree with the computer, and they're right. Again, just doing their job./
    But now if they disagree with the computer, and they're wrong, that is a world of malpractice lawsuit about to drop on their heads.

    That is, every incentive this person has is to just always agree with the computer. There is no great bonus for doing better, and potentially huge consequences when they disagree. (And, by the way, this is now the training data for more recent AI options).

    And it's this context we had at least one doctor billing $300,000.

    Per month.

    So, in this case at least, yes please bring on the AI. Because it's already doing it, and I'm sure the AI won't have to cost as much.

  • AI has never, ever gotten an image wrong ever. Anyway, "Sir, you have a 1910 model freight train in your kidneys. It simply cannot be mistaken."
  • Hospitals could potentially produce "major savings".....savings which in our for profit healthcare systems will no be passed down to the pati------er customers.
  • He has much work left to do.
  • Lets see how this plays out after the first law suit of a missed diagnosis.

  • Dos anyone actually think this would reduce costs for patients by a single penny? If so, you simply haven't been watching the healthcare industry since the 1990s. The only win here is for the margins of the hospital and whatever company charges for access to the model being used. Patients won't see one penny of savings but they will see reduced quality of care as radiologists cease to exist and there are fewer and fewer actual experts to back up the AI or to ask questions of. This is just yet another an

Slashdot Top Deals

The universe does not have laws -- it has habits, and habits can be broken.

Close