Can AI Replace Hospital Radiologists? (cnn.com) 112
An anonymous reader quotes CNN:
Radiologists, who receive years of training and are some of the highest paid doctors, are among the first physicians who will have to adapt as artificial intelligence expands into health care... Today radiologists face a deluge of data as they serve patients. When Jim Brink, radiologist in chief at Massachusetts General Hospital, entered the field in the late 1980s, radiologists had to examine 20 to 50 images for CT and PET scans. Now, there can be as many as 1,000 images for one scan. The work can be tedious, making it prone to error. The added imagery also makes it harder for radiologists to use their time efficiently... The remarkable power of today's computers has raised the question of whether humans should even act as radiologists. Geoffrey Hinton, a legend in the field of artificial intelligence, went so far as to suggest that schools should stop training radiologists.
X-rays, CT scans, MRIs, ultrasounds and PET scans do improve patient care -- but they also drive up costs. And now one medical imaging startup can read a heart MRI in 15 seconds, a procedure which takes a human 45 minutes. Massachusetts General Hospital is already assembling data to train algorithms to spot 25 common scenarios. But Brinks predicts that ultimately AI will become more of a sophisticated diagnostic aid, flagging images that humans should examine more closely, while leaving radiologists with more time for interacting with patients and medical staff.
X-rays, CT scans, MRIs, ultrasounds and PET scans do improve patient care -- but they also drive up costs. And now one medical imaging startup can read a heart MRI in 15 seconds, a procedure which takes a human 45 minutes. Massachusetts General Hospital is already assembling data to train algorithms to spot 25 common scenarios. But Brinks predicts that ultimately AI will become more of a sophisticated diagnostic aid, flagging images that humans should examine more closely, while leaving radiologists with more time for interacting with patients and medical staff.
Re: Can AI replace Slashdot editors? (Score:1)
Can we teach editors Betteridge's law of headlines?
Re: (Score:1)
Re:Can AI replace Slashdot editors?
How could you possibly train an AI to be that stupid?
Re: Can AI replace Slashdot editors? (Score:1)
A simple "Hello World" script could replace Slashdot editors.
Re: (Score:2)
i was thinking maybe something along the lines of if then else
An AI isn't smart by itself. (Score:4, Interesting)
An AI is only as good as the people that have taught it. Of course it can accumulate experience and never forget, but humans also have a thing called intuition to see things in a different view and capture things that are completely new.
Humans and AI will however supplement each other for improved accuracy.
Re:An AI isn't smart by itself. (Score:4, Interesting)
An AI is only as good as the people that have taught it.
Humans don't "teach" it. It learns directly from raw data.
humans also have a thing called intuition
That was also used to explain why computers couldn't play chess or Go as well as humans. Intuition is just pattern recognition.
Humans and AI will however supplement each other for improved accuracy.
It will start out that way. But later, humans will be removed from the process when it is clear that they just add cost, delay, and errors.
Re: (Score:2)
Not to mention, what happens when the computers crash (they always do)...you're gonna need Dr's to read films till the systems come back online....
Patients needing emergency reads won't simply go on hold till the computers come back online....
Re: (Score:2)
I think you're STILL going to want to have human Radiologist Dr's backing up what likely will be initial AI readings of the films....
Sure, for legal reasons, not technical reasons.
Not to mention, what happens when the computers crash
Computers can reboot in seconds, or at most a few minutes. According to TFA, humans take 45 minutes to read a scan. So the computer can crash and reboot a dozen times and still beat a human.
Re:An AI isn't smart by itself. (Score:4, Interesting)
Emergencies are even more reason to have a computer do the reading. The computer works 24/7, with at least 99% uptime. The human? Not so much. I work with radiologists. They're wonderfully trained. But they can only know so much, make mistakes, and want to go home to their families. They ignore or don't hear their pagers sometimes. Computers don't.
Re: (Score:2)
Re:An AI isn't smart by itself. (Score:4, Insightful)
No it learn from *annotated* data. That is data where a human, most probably a doctor, told the system what to look for. Good quality annotated data is scarce and expensive to get and that is the #1 problem in AI research.
Re:An AI isn't smart by itself. (Score:4, Interesting)
Not so much anymore. Unsupervised pretraining lets the machine learn from unlabelled data. Most of the labels actually come from the followup. Did that patient have a brain tumor? You don't ask the radiologist, you ask the pathologist.
Re: (Score:2)
An AI is only as good as the people that have taught it.
Humans don't "teach" it. It learns directly from raw data.
And fail. For these application you can only do supervised learning, and that means exactly that humans "teach" it. Unsupervised learning is only useful for identifying clusters, but not for what they mean. Hence learning from "raw data" is completely worthless here.
Re: (Score:2)
For these application you can only do supervised learning
Not true. An ANN can learn a lot from unlabeled data. Enough to spot anomalies, although not enough to make a specific diagnosis. So you combine a lot of pretraining with unlabeled data, with fine tuning using human annotated images.
Re: (Score:2)
You seriously want to claim that after-the-fact labeling of the clusters by experts is "learning from the raw data" and not "learning from humans"?
Re: (Score:2)
That was also used to explain why computers couldn't play chess or Go as well as humans. Intuition is just pattern recognition.
Pattern recognition is great as long as there is a finite, quantifiable set of patterns that can be recognized. I'll truly be impressed when an AI can improvise a solution to a pattern that has never before been seen. Shortly thereafter, I expect to be in the unemployment line.
Re: (Score:1)
As soon as someone actually creates a true AI we can test your theories. Everyone claiming they have working AI's should be sued for fraud.
Time, or money? (Score:3)
And now one medical imaging startup can read a heart MRI in 15 seconds, a procedure which takes a human 45 minutes.
That's quite nice that you save 44.75 minutes, but isn't kind of more interesting how much money you save? Save for emergency situations, waiting an hour usually isn't a problem in US medicine. Even 45 minutes of CPU/GPU time is going to be cheaper than 45 minutes of human time.
Re: (Score:2)
Re: (Score:1)
Actually, I bet you the AI will be more expensive, at least in the short term. The reason scans are expensive isn't because the human doctors analysing the output is expensive (even though they are) - it's because the equipment is hugely expensive, because it has to amortise the cost of hugely expensive R&D into how to build it.
The same will be true of the AIs that analyse the output.
Re: (Score:2)
Re: (Score:2)
Re:Time, or money? (Score:5, Informative)
It costs about $500 for an hour long MRI, including the tech to run the scanner. We used to joke that the radiologists made $250 for sticking a film up on the screen (that was fifteen years ago or so). They're a little slower now, but they can still read most scans in closer to five minutes than half an hour.
* Figures may vary in crazy parts of the world, such as the USA.
Re: Time, or money? (Score:2, Informative)
I'm a radiologist.
Radiologists manage the electromagnetic & ultrasound spectra for the benefit of patients. We help make sure your test is both useful and safe. MRIs used fields 30-60,000x as powerful as the Earth's field and RF energy for your scan. If you've ever seen foil in a microwave or a rail gun, you could imagine what could go wrong. We decide what kind of test you should get (many options here, types of MRI sequences, radiotracers for nuclear medicine scans, using IV contrast) to best inform y
Re: Time, or money? (Score:3)
Re: (Score:2)
Supposedly this is what IBM/Merge is trying to do with Watson. I would guess it will eventually work, but still has a long way to go.
Re: Time, or money? (Score:2)
Computers will be better at reading images. They will also be better at deciding what tests to do.
Radiologists can easily be replaced.
Re: Time, or money? (Score:3)
Doctors and nurses account for less than 10% of healthcare costs put together. But I agree with you, I think YOUR doctors should be replaced with Google as quickly as possible.
Yes. (Score:3)
flagging images that humans should examine more closely
That's exactly how it's going to work. You could train a 5 year old to determine if an image showed scenario A or B. It's just fancy chicken sexing. Anything that is decisively A or B gets labeled as such. Images that are questionable get passed on to Level 2, the Human.
Then you re-train the network, rinse repeat.
Re: (Score:3)
Retraining is subject to pretty strongly diminishing returns. Also, the human experts spend only a small part of their time on the simple things, because they see at a glance what is going on. They do spend most of their time on the tricky stuff and that is not accessible to weak AI anytime soon, if ever. And strong AI is not even on the very distant horizon, and may never become available.
No, AI can not (Score:5, Insightful)
Can AI Replace Hospital Radiologists?
No, AI can not. What AI can do is be an extremely valuable tool for radiologists and doctors, one that makes analyzing all the various forms of synthetic medical imagery more accurate, and most likely increasing their productivity. It can reduce oversights and errors, but it won't be able to fully replace expert human analysis for quite some time. Like most AI solutions it will most likely take far longer than AI experts predict. Perhaps we need an AI to predict timeframe for AI solutions since people seem to do that poorly. :-)
Does this perhaps lead to reducing the number of radiologists due to increased productivity? Probably not, more likely radiologists will be bombarded with more imagery to analyze as technology improves and costs lower and is more frequently used.
Re: (Score:1)
If new tools allow a person to handle twice or even ten times the case load, you've effectively replaced people with A.I.
Only if the number of tasks to perform remains constant. Currently radiologists are a bottleneck that constrains the use of medical imagery. With the easing of this bottleneck through productivity, and the lowering costs of the technology, medical imagery is likely to be used far more than it currently is. Therefore the need for radiologists remains.
Re: (Score:2)
Obviously AI can replace radiologists. If new tools allow a person to handle twice or even ten times the case load, you've effectively replaced people with A.I.
Radiology can turn into a call center job as well, where a small team of people handle departments at multiple hospitals and medical centers. Imagine a world where a staff of 3-4 handles every case for 100 locations. Easily means 1000s of cases per day, but with A.I. it is conceivable.
Idiots think A.I. replacing a job means the original job no longer exists. That is a naÃve postion to take.
It is already a call-center job. You still have a tech who wheels you into the machine,
administers contrast dyes or whatever, and pushes the buttons. But the scan (e.g. X-Ray)
is read by some radiologist sitting in a lab or call-center or their home. They are being
fed many, many images over the Internet to their laptop. They are given a certain amount
of time to read the image and return a diagnosis. They are overworked and treated like
a machine already, generally feeling that the allotted time is too sm
Never say never =) (Score:1)
No, AI can not.
Most of the world didn't think that the automobile would replace horse + carriage when first introduced.
Re: (Score:2)
No, AI can not.
Most of the world didn't think that the automobile would replace horse + carriage when first introduced.
Horses are still more common than AI driven cars. The human driven cars don't count. :-)
Re: (Score:2)
It was loud and prone to breakdown. It's not that people were against a working automobile. People today for example are against 3D TV, VR, and fusion energy .. but not because they think seeing things in 3D is bad or that clean energy is useless. They think the technology can't do it properly or will turn out to be impossible to implement. Etienne Lenoir's automobile of 1862 was very noisy and traveled at a whopping 2 miles per hour. That's slower than a 90 year old man with a cane. A human normally walks
Re: (Score:1)
What has happened instead is that radiology techs now do a lot of the jobs; like ultrasound scanning and reporting, protocolling and reporting CT scans, protocolling MRI, performing minor interventional procedures like central lin
Re: No, AI can not (Score:2)
Agreed. Say instead that such technology would decrease the number of things radiologists miss, allowing them to spend more time on comparisons and differential diagnosis... the things people actually hire them to do.
Radiologists won't be the only ones. (Score:2)
Dermatologists are looking towards extinction as well.
BTW, as a physician who spends half his day in the hospital and half in the office, I average speaking to a radiologist about once a year.
You could replace all the radiologists (excepting interventional radiologists) in the U.S. with IBM's Watson and a dozen humans for over-reads. You don't go into radiology if you're a people person.
Re: Radiologists won't be the only ones. (Score:1)
As a radiologist in training, I literally talk to clinicians every day. We discuss what study to order (or shouldn't order as it's a waste of time and resources) with the least amount of harm to patients. We also can communicate with clinicians about probable diagnoses given the context and the limitations of a poor quality image which I don't see AI being able to handle well any time soon.
Radiologists guide a lot of medical management and what may be read as a volume averaging artifact by a computer may
Re: (Score:3)
Don't worry. I recently had a chance to talk to some Watson people from IBM, and while they see excellent potential for supporting experts in diverse fields, the answer as to replacing experts was "not in the next 50 years". That is pretty extreme, as they really know what the state-of-the-art is.
Re: (Score:2)
Watson can be Pubmed for doctors who have to deal with obscure stuff they aren't a specialist in. It's not going to replace the specialists that actually deal with that obscure stuff on a daily basis.
Re: (Score:2)
One of the things they are wanting to do with Watson is ingest images and reports, develop some computer aided diagnostic information based on that to to make suggestions on pathologies. It will bring up a window with images of examples of diagnosed pathologies that are similar to what is on the rads workstation. Supposedly will even do boilerplate for the report based on that.
Re: (Score:2)
IBM doesn't want to piss off their customers (radiologists buy the radiology-related equipment).
IBM came to the hospital where my lab is to talk about their genomics platgorm, but threw in a demo of Watson assessing a patient in the ER, ordering a CT, evaluating the CT and prescribing treatment. They're playing it smart: tell the radiologists you have a great new tool to help them. Then let nature, economics, and hospital administrators take their course.
Re: (Score:2)
Not credible and does not match published research. Oh, sure, they are using the limitations of what is possible to do marketing by giving it a positive spin, but the limitations are real.
Re: (Score:1)
"Not credible and does not match published research."
Cite.
Re: (Score:2)
No. And for obvious reasons. That "comeback" just means you have run out of arguments.
Re: Radiologists won't be the only ones. (Score:2)
AI will do everything you can do... And do it better
Re: (Score:1)
The radiology nurses that were studying at my university were definitely people persons, if you catch my drift. They were infamous for how people person-y they were.
Re: (Score:1)
Dermatologists are looking towards extinction as well.
BTW, as a physician who spends half his day in the hospital and half in the office, I average speaking to a radiologist about once a year.
You could replace all the radiologists (excepting interventional radiologists) in the U.S. with IBM's Watson and a dozen humans for over-reads. You don't go into radiology if you're a people person.
From the summary:
while leaving radiologists with more time for interacting with patients and medical staff.
So far I have 14 years in hospital IT, and when I read "while leaving radiologists with more time for interacting with patients and medical staff", I thought well that's something no one wishes for.
Re: (Score:2)
Where I used to live, back in the 1990s radiologists were charging $250 for a quick look at a simple chest x-ray. Who knows what they are charging now. I would love to see those people replaced with something less greedy.
Re: (Score:1)
I'm a radiologist. I spent 13 years in school after finishing college to obtain my degrees. I am paid $6 by Medicare to read a two view chest x-ray. Please don't post misinformation. Reimbursement rates are public and published by CMS, and your information is incorrect.
If you know of a job where chest x-rays are reimbursed at $250, please respond to this message so I can contact you and immediately switch jobs.
Re: (Score:2)
This is not misinformation. I don't know what the radiologist got paid, but $250 is what I was billed.
Re: Radiologists won't be the only ones. (Score:2)
As an intensivist at a top 10 hospital, I'll tell you that either you are a) trolling or b) yourself, your radiologists or both are incompetent. I speak to radiologists on a daily basis... and value their opinions.
more time (Score:1)
.. while leaving radiologists with more time for interacting with patients and medical staff.
It will leave more time for radiologists to review additional flagged scans since that will probably make them the most money.
Would you trust an AI? (Score:2)
Second Opinion... (Score:3)
Siri, do you concur with Watson?
Re: (Score:2)
Actually difficult image-based problems is where AI will excel, give time and good training data. However a kind of medical "expert system" is not exactly in sight yet. So I basically agree with you: AI will be an excellent tool, boosting productivity and driving costs down. I am not sure it will drive the number of doctors down.
AI (Score:3)
Re: (Score:2)
It is. Sort of. But it is weak AI (the "AI" without actual intelligence) and far, far removed from what a radiologist does.
Re: (Score:2)
"Image recognition is not AI. Sorry."
No existing software can do image recognition yet. Some can identify a face, a dog or a common item in a picture. Most pictures are far more complex.
The more accurate term would be 'pattern recognition'. A program may be able to quickly compare features of an image with those of millions of others and recognize common patterns. Unfortunately some images- CT scans, sonograms etc have such poor resolution that certainty is elusive.
Re: (Score:2)
For the case of cancer screening with an X-ray, they look for things like a high-density mass that shouldn't be there - a fuzzy white patch. Maybe it has gained a better blood supply and there are fuzzy white lines around it - enlarged arteries/veins. Maybe there is a cluster of small tumors nearby. But then again, those could just be fat cells. Each of those needs to be checked up.
Re: (Score:2)
Then that is another AI goalpost moved. Eventually, it won't be AI until it walks into the room, discusses philosophy, and declares its eternal love or hate for you.
Can any high-skill profession be replaced? (Score:3)
I think most routine radiology could end up at least being assisted by AI, given that the entire practice revolves around using imaging techniques that return incomplete data and making a judgement call. This is the sort of thing machine learning is good at -- reading billions of images and determining what something looks like. Real radiologists in training do the same thing -- except they have a much smaller data set to fall back on.
The real question is how we're going to deal with the sudden flip in what is considered a highly-skilled job:
- Doctors in general are a perfect example - because the supply of medical school slots is kept low, only the people with perfect grades and photographic memories, _and_ who can ace the MCAT get into med school (in the US.) If machine learning becomes a thing, then having a photographic memory is not going to be as important as it once was...it's already less important.
- The Bar Association didn't limit the amount of law school slots the way the AMA does, and the result today is that law school grads can't find work. Just 20 years ago, having a law degree would definitely get you a job, and having a Top 14 law degree would set you up for life permanently. Law is a profession that relies on interpreting vast amounts of data, and computers are really good at the routine parts of the job that junior associates used to make $180K a year doing.
- From the non-professions, another example is air traffic controllers. Even with computers aiding them, humans who have the unique ability to think natively in 3 dimensions and keep an entire sector of airspace's inhabitants in their brains along with their speed, altitude and heading have been doing it for ages. It takes years of training to understand and definitely qualifies in my book as a highly skilled job. They make a lot of money because few people have the ability to do it and keep their stress levels non-lethal. But, it also sounds like something computers could take over eventually.
No, obviously (Score:2)
And nobody honest claims that they can. For example, the Watson people from IBM say "will not replace experts in the next 50 years". The state of the art in AI gets massively overestimated all the time. Actual fact is that the only thing available is "weak AI", and that one does not have any actual intelligence and is restricted to library look-up and statistical classificators. Very useful, but not even remotely doing what a smart and experienced human being can do. These algorithms do not have insight or
What Geoff Hinton said (Score:2)
Geoff Hinton said that machine learning and particularly deep learning is so good now that anything that amounts to classification will soon be done better by machines than humans. The except is here [creativede...ionlab.com]. So logically medical schools should stop churning out radiologists now.
I would like to agree with him, but note that there is still a large difference between even large synthetic tests and the real world. At the end of the day a trained medical specialist will have to make the diagnosis, machines are not ther
Malpractice?? (Score:1)
Hell no. (Score:2)
There will always be things that the AI will miss.
Re: (Score:2)
And there will always be things that humans miss. The key to this is what proportion will machine-learning trained systems miss vs. human radiologists? Given that the machine systems can handle feedback from a near-infinite number of case (including that from human experts), thus meaning it can, theoretically, improve faster than any human can.
social responsibility (Score:2)
Programmers should have a duty of social responsibility
to not develop AI which will eventually replace any worker.
The easiest person to replace is an AI developer.
Re: (Score:1)
OK, let's stop work on Internet hardware (replaced digital telephone exchanges, Strowger electro-mechanical systems, telephone operators and telegram delivery people), or how about elevators (replaced the elevator operator, a butler like person who wore white glove and moved the up/down/stop lever), or Photoshop and digital looms (replaced four artisans operating one weaving loom), highway traffic lights (replaced a police officer at each intersection), word processing software and laser printers (replaced
For now no, but CS folks always want to shine (Score:1)
In medicine Radiology appears to be the hi tech thing everyone except the patient or radiologist want to automate/eliminate. Sure radiologists are expensive, and hospital admins hate them for their costs, but it's still going to be awhile before they are replaced. It's not about identifying the problems, but preventing false positives. AI algos are good at the former, not the latter.
But of course, radiology is heavy image based, so all our computer vision and machine learning CS folks want to completely sol
1 in 8 radiologists spotted a problem I have (Score:1)