Quantcast

Will AI replace Radiologists?

This forum made possible through the generous support of SDN members, donors, and sponsors. Thank you.

catikati

New Member
Joined
Oct 10, 2015
Messages
2
Reaction score
1
Better and more accurate question is; When will AI replace radiologists?

This is really happening. Look at that TED talk, here;


Is this really happening? Should we concerned about it? I am :(
 
  • Like
Reactions: 1 user

maxxor

Full Member
10+ Year Member
Joined
Apr 11, 2009
Messages
1,145
Reaction score
1,058
As someone who does research in this area, there's still a lot to go before this is usable in the reading room.
 
  • Like
Reactions: 1 users

sphericalFunny

New Member
Joined
Apr 28, 2016
Messages
8
Reaction score
2
IMHO, a 'true' AI that can replace radiologist would replace almost every other white collar job first.
Similar AI controlling humanoid robots/machines would replace all the blue collar jobs too.
So ultimately, all of us are going to be jobless......+pity+

Even those involved in building those AI would not be required for further improvement.
An AI not limited by limits of biology would perpetually keep on improving itself and its "executive" arm (=machines/robots).

However, one thing is for sure, medicine (as a whole, radiology included) would be the last pillar to fall to fully autonomous AIs.

Just a thought........:(
 
  • Like
Reactions: 1 users

Fab5Hill33

Junior Member
Removed
10+ Year Member
Joined
Jun 14, 2004
Messages
212
Reaction score
98
Many other fields of medicine will fall to AI before radiology does. For example, internal medicine algorithms for diagnosis and treatment would be easy to program. Nobody relies on physical exam anymore anyways. It's just inputting lab results and procedural results (radiology, endoscopy, angio). It's the procedural results that are subjective and require significant experience. These are the ones that are most immune (relatively) to AI.
 
  • Like
Reactions: 3 users

Cognovi

Knowledge worker
5+ Year Member
  • Joined
    Feb 10, 2016
    Messages
    727
    Reaction score
    782
    The computer reads most EKGs fine, but it hasn't completely replaced the need for human verification, and it hasn't replaced a human discussing with another human what the EKG means for the patient.

    AI will first help with routine tasks such as detecting lung nodules, measuring the size of tumors on follow-up scans, calculating ejection fraction. Humans can then focus more mental energy on less routine interpretative tasks, thinking about the management implications, and interacting with colleagues/referrers/patients.
     
    • Like
    Reactions: 1 users

    Mad Jack

    Critically Caring
    7+ Year Member
    Joined
    Jul 27, 2013
    Messages
    37,390
    Reaction score
    72,485
    The computer reads most EKGs fine, but it hasn't completely replaced the need for human verification, and it hasn't replaced a human discussing with another human what the EKG means for the patient.

    AI will first help with routine tasks such as detecting lung nodules, measuring the size of tumors on follow-up scans, calculating ejection fraction. Humans can then focus more mental energy on less routine interpretative tasks, thinking about the management implications, and interacting with colleagues/referrers/patients.
    Kek, the computer reads EKGs just fine? What planet do you live on? :laugh:
     
    • Like
    Reactions: 1 user
    C

    cowme

    It's funny how I've seen this post every year for the past 6 years and haven't seen a single notable advancement in CAD literature.
     
    • Like
    Reactions: 4 users

    Kazaki

    Whatareyetakinabet?
    Joined
    Feb 15, 2016
    Messages
    187
    Reaction score
    87
    No.
    Why no?
    Machines are stupid. Hold on, they're not even stupid, they're just...machines. They can hold on to vast amounts of information, but what can they do with it? Pretty much nothing that a human can't do, they just do it faster. Even the "smartest" of deep learning algorithms can't tell you what it means when pathology is detected. Detection and diagnosing are very different. Nothing can replace the most complex piece of matter in the known universe, if its used correctly.

    Since someone mentioned how we will all be jobless in the future. This is not economically viable. We're creating new jobs daily, but **** happens and jobs are lost at the same rate, think oil and gas who would've thought that people would get ****-canned from those gigs?? And yet we still have double digit unemployment rates in developed first-world countries, which was absurd to even think about just 2 decades ago. What do you think governments are gonna do when 30 or 40 percent of the population is unemployed? That's how revolutions start, and they're not the good kind where **** people get sacked, these are peasant revolutions that do nothing but destroy countries.

    Speaking of money n stuff. People are always gonna want someone to blame when **** goes wrong. Show me a company, big or small, that is willing to take on the risk of potentially misdiagnosing 0.01% of scans? A radiologist that misreads 10,000 studies would be hung by his a$$-hair.
    So if you ever find that company please tell us who they are, because I'd want to ask their CEO how he carries around his huge balls.
     
    • Like
    Reactions: 4 users

    Taurus

    Paul Revere of Medicine
    15+ Year Member
    Joined
    Jul 27, 2004
    Messages
    3,182
    Reaction score
    563
    If AI replaces radiology, I would be more worried about robots taking over the world and enslaving humans.
     
    • Like
    Reactions: 2 users

    meister

    Senior Member
    15+ Year Member
    Joined
    Aug 24, 2004
    Messages
    2,154
    Reaction score
    356
    The radiologist AIs will eventually become self-aware. Once that happens, robot radiologists will become what all fear: the true gatekeeper. They will ruthlessly block all unnecessary studies (most of them), and will cut imaging volume to the bone. They don't care about what your attending really wants. They don't care if you page the chairman. They will never call the tech in from home. The decision has already been made. Welcome to the future.
     
    • Like
    Reactions: 5 users

    PCIRISH52

    Full Member
    10+ Year Member
    Joined
    Jun 21, 2009
    Messages
    157
    Reaction score
    58
    Auntminnie has a front page story called "Will AI soon put radiologists out of a job" by Dr. Eliot Segal, an expert on the subject matter. It's an informative read and I think will put you all at ease about this.
     
    • Like
    Reactions: 1 user

    The Cure

    New Member
    Joined
    Apr 3, 2014
    Messages
    4
    Reaction score
    6
    I rarely (read: never) post, but I've gotta pitch in on this one.

    PGY-4 rads here. Like others have mentioned, we've heard about this over and over for a lot of years. It seems unlikely it'll happen any time soon.

    But I've gotta say, I would love for it to happen now. What has happened in other specialties when other entities become capable of doing their job? For a time, things get absolutely SWEET. Look at Gas. CRNAs eventually became a problem (maybe?), but right after they created CRNAs, for the next 10-15 years, anesthesiologists made incredible money billing for what was going on in four ORs simultaneously. The same has happened in multiple other specialties.

    It'll be the same if they ever come up with good AI for rads. For a while life will become amazing, as you'll be able to "supervise" the machine and safely do a ton of studies. How long will the sweetness last? Well, how long do you think it would really take before the machine would be trusted? And after that, waging war with the AMA and ACR, how long do you think it would take to battle through congress to let AI practice medicine. I think 10 years would be an extremely low estimate. More like a career cycle (25-35 years).

    So as far as I'm concerned, please someone out there, discover rads AI now!
     
    • Like
    Reactions: 3 users

    WellWornLad

    Full Member
    10+ Year Member
    Joined
    Feb 5, 2008
    Messages
    1,088
    Reaction score
    33
    Lot of good points on this thread that I would second. Physical exam skills have definitely atrophied to the point of no return. Even the best physical diagnosticians cannot come anywhere close to the sensitivity/specificity of rads or lab tests, so I can't blame clinicians for not trying. A lot of docs don't even try to come up with an educated differential anymore, it's all abdominal pain -> CT r/o everything. When clinicians are just responding to a series of test results performed by others and working down a flowchart, you realize just how expendable they are to mid level replacement.

    As for AI, there are a lot of reasons not to worry. The most basic is that there's one thing that no AI will ever be able to to - assume liability. People gotta have someone to sue, and no company is lining up to take that on.

    I'd definitely echo The Cure here - best case scenario in our lifetimes is having a rads sign off on a stack of computer-generated reports. Wouldn't hold your breath, though.
     
    • Like
    Reactions: 1 users

    greg1184

    Full Member
    10+ Year Member
    Joined
    Dec 9, 2006
    Messages
    1,942
    Reaction score
    165
    The real threat is not so much AI but the Ezekiel Emmanuels of the world who would do anything to see radiology eliminated as a specialty.
     
    • Like
    Reactions: 1 user

    postbacpremed87

    Full Member
    7+ Year Member
    Joined
    Jan 26, 2011
    Messages
    2,041
    Reaction score
    579
    The radiologist AIs will eventually become self-aware. Once that happens, robot radiologists will become what all fear: the true gatekeeper. They will ruthlessly block all unnecessary studies (most of them), and will cut imaging volume to the bone. They don't care about what your attending really wants. They don't care if you page the chairman. They will never call the tech in from home. The decision has already been made. Welcome to the future.

    It's a neural net processor...a learning computer
     

    greg1184

    Full Member
    10+ Year Member
    Joined
    Dec 9, 2006
    Messages
    1,942
    Reaction score
    165
    Double post.
     
    Last edited:

    greg1184

    Full Member
    10+ Year Member
    Joined
    Dec 9, 2006
    Messages
    1,942
    Reaction score
    165
    So you suppose having him give the keynote at ACR this year was a "keep your enemies closer" sort of thing?

    I certainly hope so. He is indeed a powerful guy and we need to hear what he has to say and network with him. If I would have gone to ACR I would have asked him if he read his own mammograms, being a breast oncologist.... since you know what we do is garbage and we don't need radiologists! I can take a wild guess and say that it would be a big fat No :rolleyes:.

    He is clearly biased against the field when he says that we are "on his hitlist." That is unfit for a leadership position. He represents everything that is wrong with medicine, being fragmented and divided and have an inability to unite politically. If the ACR thinks that imaging 3.0 (aka the "we are doctors too") campaign is going to cater to and change minds of people like this, I have a bridge to sell. People like Zeke get an inch, they take a mile.

    Why not primary care? NPs or PAs can run a practice with IBM Watson gathering data and cross referencing with articles. How about pathology? There is just as much pattern recognition in this field. You just need some grunts to do the gross specimens. But no. There is clear bitterness against radiology for one reason or another.

    Ironically, my call experiences have shown that there is more reliance on radiology (and radiologists) than ever.... like a baby and its bottle. So much for being obsolete.
     

    maxxor

    Full Member
    10+ Year Member
    Joined
    Apr 11, 2009
    Messages
    1,145
    Reaction score
    1,058
    Ironically, my call experiences have shown that there is more reliance on radiology (and radiologists) than ever.... like a baby and its bottle. So much for being obsolete.
    Agreed. In our ER, it's likely the Radiologist is the first physician to "see" the patient. Often when I call with acute findings, the patient gets "promoted" from PA/NP level of care to resident/attending.

    As long as we keep midlevels out of radiology interpretation, they may be the greatest proof of our value yet.
     

    catikati

    New Member
    Joined
    Oct 10, 2015
    Messages
    2
    Reaction score
    1
    Did anyone listen ezekiel emanuel's keynote speech in ACR 2016? What did you think of it?
    I cant find the speech anywhere on web :shrug:
     
    Last edited:

    Naijaba

    Full Member
    10+ Year Member
    Joined
    Apr 2, 2007
    Messages
    1,057
    Reaction score
    111
    Auntminnie has a front page story called "Will AI soon put radiologists out of a job" by Dr. Eliot Segal, an expert on the subject matter. It's an informative read and I think will put you all at ease about this.

    I love these conversations.

    Disclaimer: I think radiology reads can and should be replaced by autonomous reads with the radiologist functioning in a supportive / clinical role when the referring physician has a question.

    Regarding Dr. Segal: I agree with most of what he wrote in that article, but he is incorrect about the feasibility piece:

    "The work on computer vision that can recognize a water bottle in an image database is fundamentally different than the type of data that we have in diagnostic imaging, and no one has any evidence of any algorithm or methodology that could be comparable to what's being used for the imaging challenges [in diagnostic radiology]."​

    And also:

    "Siegel noted that there have been tens of thousands of algorithms developed for image analysis and decision support over the past 30 years. Hardly any are in clinical practice."
    Let's clear the air - the ImageNet competition (big international competition designed for classifying images) was created specifically because algorithms developed "over the past 30 years" were very poor at image recognition. These are exactly the same kinds of problems to an H&P-blind radiologist. If the radiologist is operating solely on the image data, then it is a classification problem.

    Now, Dr. Siegel has also been quoted (at RSNA 2016 I believe) as saying, "It's much harder than cats vs. dogs." Again, this is false. Cats vs. dogs is a notorosly hard problem in computer vision because: Dogs and cats both have two eyes, they both have tails, they both have four legs, they both have fur, and so on... Yet, human beings can instantly tell dogs and cats apart.

    This is why deep learning is different than the past 30 years of image processing and machine learning algorithms. It is able to pickup on the subtle differences between images just like the human mind.
     
    • Like
    Reactions: 1 user

    RadsFTW123

    Full Member
    Joined
    Jun 19, 2016
    Messages
    57
    Reaction score
    26
    Naijaba: how are you going into radiology if you think this is true? Rads will be an awful field for most if it's true. The fact that you are still pursuing rads despite this makes me think all your reasoning is suspect at the very very best.
     

    Naijaba

    Full Member
    10+ Year Member
    Joined
    Apr 2, 2007
    Messages
    1,057
    Reaction score
    111
    Naijaba: how are you going into radiology if you think this is true? Rads will be an awful field for most if it's true. The fact that you are still pursuing rads despite this makes me think all your reasoning is suspect at the very very best.

    Hi RadsFTW123, my background is in computer science and my second job out of college was working for a PACS company. Way back in 2010 I thought it would be amazing to start a company that automates reads, but it was technically impossible - all the negative comments above would be completely justified. I ended up going to med school to learn more about the practice of radiology, with the goal of starting a radiology informatics company. During med school I took a leave to get a master's degree in computer science out at Stanford. This was in 2013, a year after the ground-breaking result at the ImageNet competition. Deep learning took the world of machine learning by storm. All major tech companies from finance to consumer electronics have their fingers in deep learning now. It's like witnessing pigs fly, something you believed impossible is now possible, if not probable. I returned to med school for my 3rd year rotations. Long story short, I also enjoy working with patients and found IR to be an amazing balance between my aspirations in radiology and the desire to work with patients.
     

    PL198

    Full Member
    5+ Year Member
    Joined
    Apr 25, 2014
    Messages
    4,509
    Reaction score
    4,128
    Hi RadsFTW123, my background is in computer science and my second job out of college was working for a PACS company. Way back in 2010 I thought it would be amazing to start a company that automates reads, but it was technically impossible - all the negative comments above would be completely justified. I ended up going to med school to learn more about the practice of radiology, with the goal of starting a radiology informatics company. During med school I took a leave to get a master's degree in computer science out at Stanford. This was in 2013, a year after the ground-breaking result at the ImageNet competition. Deep learning took the world of machine learning by storm. All major tech companies from finance to consumer electronics have their fingers in deep learning now. It's like witnessing pigs fly, something you believed impossible is now possible, if not probable. I returned to med school for my 3rd year rotations. Long story short, I also enjoy working with patients and found IR to be an amazing balance between my aspirations in radiology and the desire to work with patients.

    So you want to do IR but automate DR
     

    Naijaba

    Full Member
    10+ Year Member
    Joined
    Apr 2, 2007
    Messages
    1,057
    Reaction score
    111
    So you want to do IR but automate DR

    Yes, that's my goal in terms of career aspirations. I'm not the best academic or clinician, but I have a unique perspective from the computer science / industry side of things.
     

    Gadofosveset

    Full Member
    7+ Year Member
    Joined
    Oct 3, 2014
    Messages
    383
    Reaction score
    446
    Yes, that's my goal in terms of career aspirations. I'm not the best academic or clinician, but I have a unique perspective from the computer science / industry side of things.

    Dear (almost M.D.) Naijaba,

    You have been encountering resistance on these threads, please let me explain:

    - You have not practiced radiology and yet claim some kind of certain knowledge that diagnostic imaging will be automated soon. This seems premature of someone who has not yet taken call and presumably does not understand the practical difficulties of rendering a final impression that can be used practically by clinicians. Certainly AI and deep learning will change medical care (and the rest of the world) in the future. Your unbridled enthusiasm seems to imply that this will happen very soon, whereas more experienced radiologists seem to think this will happen later rather than sooner.
    - Most if not all (?) radiologists interested in image processing are diagnostic radiologists, for obvious reasons. Your professed passion in computer-aided diagnostic imaging seems at odds with your career goal of interventional radiology. At best this seems misguided from a lack of practical experience and you will become a researcher in diagnostic imaging. At worst, this seems to imply someone who is blindly overenthusiastic about whatever is currently popular and it is manifesting as an AI fanboy.
    - Anyone who worked for a PACS company should be modestly aware of the limitations of implementation.
    - Your unabated enthusiasm for exclaiming that computers are nearly able to process at the level of the human mind requires some degree of skepticism, on a host of technical and philosophical grounds, a nuance which your discussion are lacking.
    - One has trouble understanding your point in making these points on this forum. No one is engaged in meaningful technical discussion. It's really just you making assertions over and over. It's unclear what your goal is here. It really sounds like you worked yourself up into a frenzy over this topic to try to impress people in interviews and now just can't stop.
    - Instead of working yourself up over the theory of a computer mind which will replace all diagnostic imaging interpretation, if you could get excited and design some reliable software for fusion of ultrasound and CT biopsies, you will do the world (and IR) a lot more good.
     
    • Like
    Reactions: 7 users

    Naijaba

    Full Member
    10+ Year Member
    Joined
    Apr 2, 2007
    Messages
    1,057
    Reaction score
    111
    Hi Gadofosveset,

    You're spot on. All of what you've said is true.

    Dear (almost M.D.) Naijaba,

    You have been encountering resistance on these threads, please let me explain:

    - You have not practiced radiology and yet claim some kind of certain knowledge that diagnostic imaging will be automated soon. This seems premature of someone who has not yet taken call and presumably does not understand the practical difficulties of rendering a final impression that can be used practically by clinicians. Certainly AI and deep learning will change medical care (and the rest of the world) in the future. Your unbridled enthusiasm seems to imply that this will happen very soon, whereas more experienced radiologists seem to think this will happen later rather than sooner.

    I agree with all of the above. I have limited practical experience and the perceived time-frames are what I've heard as well with the same distributions you've mentioned.

    - Most if not all (?) radiologists interested in image processing are diagnostic radiologists, for obvious reasons. Your professed passion in computer-aided diagnostic imaging seems at odds with your career goal of interventional radiology. At best this seems misguided from a lack of practical experience and you will become a researcher in diagnostic imaging. At worst, this seems to imply someone who is blindly overenthusiastic about whatever is currently popular and it is manifesting as an AI fanboy.

    You are right, I've been a technophile since I was young. My avatar is is an 8088 microprocessor, an image I chose some seven years ago. As to IR, again you're correct, most IR attendings are not conducting image processing research. Yet, IR has an amazing history of innovation with a significant number of IR attendings starting their own device companies. I hope to follow in their example.

    - Anyone who worked for a PACS company should be modestly aware of the limitations of implementation.

    Yes, absolutely. I've mentioned before the challenges with AI/machine learning are no longer technical, but legal and administrative. I believe we could train a "digital radiologist" if all of the scans from every institution were aggregated in some central repository. Of course, it's hard as hell to get data from one institution, let alone many. Then there's user buy-in.

    - Your unabated enthusiasm for exclaiming that computers are nearly able to process at the level of the human mind requires some degree of skepticism, on a host of technical and philosophical grounds, a nuance which your discussion are lacking.

    This is a philosophical discussion, one that doesn't have an easy answer. Suffice to say that neural networks (aka deep learning) are the closest approximation to the human mind. I don't believe they are the same. One big difference, for the technically inclined, is that our model of computation (i.e. Turing machines), is not how the human mind works. We don't have a "clock frequency" (like Intel Pentium processors have at 2.7 GHz clock). The only alternative to the Turing model of computation is the quantum computer...which I know nothing about. Perhaps our minds are a combination of deep networks training by a quantum processor?

    - One has trouble understanding your point in making these points on this forum. No one is engaged in meaningful technical discussion. It's really just you making assertions over and over. It's unclear what your goal is here. It really sounds like you worked yourself up into a frenzy over this topic to try to impress people in interviews and now just can't stop.

    I have spoken a great deal about this at interviews. This has amplified my excitement. Maybe I should dial it in a bit, but I do enjoy talking about it.

    - Instead of working yourself up over the theory of a computer mind which will replace all diagnostic imaging interpretation, if you could get excited and design some reliable software for fusion of ultrasound and CT biopsies, you will do the world (and IR) a lot more good.

    Thanks for the suggestion! I've heard of other things that are immediate needs like counting lung lesions and automating bone ages. I'm pretty busy at the moment, but I'd like to take on these challenges in the future.
     

    Gadofosveset

    Full Member
    7+ Year Member
    Joined
    Oct 3, 2014
    Messages
    383
    Reaction score
    446
    I look forward to hearing about your advances in image-guided targeted therapy.
     
    Last edited:
    C

    cowme

    I would love to be a fly on a wall for a committee meeting at Stanford/ucsf/MGH after naijaba tells all the attendings that their only purpose is that of an administrator or legal punching bag
     
    • Like
    Reactions: 1 users

    Naijaba

    Full Member
    10+ Year Member
    Joined
    Apr 2, 2007
    Messages
    1,057
    Reaction score
    111
    I would love to be a fly on a wall for a committee meeting at Stanford/ucsf/MGH after naijaba tells all the attendings that their only purpose is that of an administrator or legal punching bag

    Harsh haha. Of course, I don't believe that.
     

    preDoGuy24

    Full Member
    7+ Year Member
    Joined
    Sep 28, 2010
    Messages
    492
    Reaction score
    198
    Harsh haha. Of course, I don't believe that.


    If you stated in an interview you think a radiologist should be replaced in the near future by AI and believe they will be AND you want to go into IR for the "patient care" aspect then best of luck you are one bold mofo.

    This is a touchy subject for radiologists (who do a lot more then image interpretation) and not something I would bring up in an interview without extreme caution/tact.



    Sent from my iPhone using Tapatalk
     

    shishka32

    Full Member
    10+ Year Member
    Joined
    Apr 26, 2008
    Messages
    217
    Reaction score
    28
    So computers can't interpret a 2D line on EKG's accurately, all EKG's are still read by physicians, and legal aspects wouldn't allow EKG's to be fully automated for a very, very long time, yet computers will be replacing diagnostic radiology soon? Get real. I know I know, there's new stuff from some computer nerds not in medicine coming out!!! But until the computer continues to interpret a STEMI as "AFib, possible LBBB, consider ischemia" I'm not worried. Even if it caught a STEMI 99.9% of the time, that's a multi-million dollar lawsuit the 0.01% (thousands of patients a year) of the time the stupid computer missed it.
     
    • Like
    Reactions: 1 users
    C

    cowme

    Dude you literally said "I've mentioned before the challenges with AI/machine learning are no longer technical, but legal and administrative."

    We've all asked you to provide us with any data to support these bold notions, and all we get is bolder unsubstantiated claims in response. Arguing theory as fact. This must be what it feels like to argue with Donald trump.
     
    Last edited by a moderator:
    • Like
    Reactions: 2 users

    computerex

    New Member
    Joined
    Feb 6, 2017
    Messages
    1
    Reaction score
    0
    Hey guys. Someone posted this thread in hacker news, and I joined this forum to offer some clarification. I don't think anyone going into radiology or medicine need worry about their careers rights now. Medicine is a highly regulated field and change of this magnitude is going to take decades (I think). Healthcare deals with people's lives, so naturally there will be a lot of push back and distrust against automation like this. As someone who has built machine learning systems for finance companies, it's difficult to convince executives why we should follow my blackbox model against all the industry acquired domain specific knowledge acquired over the decades. The only thing we have for us is the statistics, the results of the validation on true real life data. But it's hard to convince people with just the cold hard statistics and numbers. This is going to be 10x worse for something like medicine, because people's lives depend on it, and we should be highly skeptical.

    With that said, it is now a matter of time until machine vision starts dominating the field. It's an eventuality, and here's why.

    1. In traditional machine learning, experts (doctors for example) work with computer scientists to hand engineer "features" to solve very specific problems. For example:
    http://ieeexplore.ieee.org/abstract/document/1617200/
    https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4485641/

    These traditional techniques rely on domain specific expertise to engineer features. Once the right set of features is selected, simple linear models like naive bayes or K-means outperform human experts (real life physicians).

    The issue is that feature engineering is really hard, expensive, tedious, cumbersome and time consuming. Most importantly, feature selection is specific to each problem. Every time you want to solve a different problem, you have to re-do feature selection.

    2. Deep learning is gaining so much traction because it automates feature selection. Rather than manually engineering the features, unsupervised learning techniques and deep architectures allow the problem to be expressed in terms of feature layers that encode higher and higher level features.

    Finally we have a general technique that can be used not only on different problems in the same domain, but this same technique can be used across entire domains. As it is, deep learning is being utilized in all sorts of disciplines including speech recognition, not just machine vision.

    Someone here said that the problems radiologists are solving are very different from being able to recognize cats in videos and what imagenet is solving. This is not strictly true. It turns out, we can train a deep belief network with a bunch of random images, and the network still learns useful features/abstractions such as being able to recognize edges, corners, contours, blobs, etc.

    3. Recent advances in the field allowing the use of GPU instead of the CPU for training, convolutional neural networks, dropout, generative adverserial networks, transfer learning amongst other factors has caused a lot of positive movement in the right direction specifically in computer vision.

    Here's an amazing overview of the potential of deep learning in Radiology: http://www.cs.ucf.edu/~bagci/publications/DLradiologyRSNA2016.pdf

    Here's the type of efficiency a machine vision system can provide in pea sorting. Each pea is analyzed, and air jets are used to knock the bad ones away. This is pretty common already:


    https://arxiv.org/pdf/1602.03409.pdf

    "We study two specific computer aided detection (CADe) problems, namely thoraco-abdominal lymph node (LN) detection and interstitial lung disease (ILD) classification. We achieve the state-of-the-art performance on the mediastinal LN detection, with 85% sensitivity at 3 false positive per patient, and report the first five-fold cross-validation classification results on predicting axial CT slices with ILD categories. Our extensive empirical evaluation, CNN model analysis and valuable insights can be extended to the design of high performance CAD systems for other medical imaging tasks."

    One other important take away from the paper above:

    However, there exists no large-scale annotated medical image dataset comparable to ImageNet, as data acquisition is difficult, and quality annotation is costly.

    This is currently a huge (largely artificial) limitation. The paper discusses a couple techniques to deal with the lack of data, notably transfer learning, that allows unsupervised pre-training on already existing large natural image datasets like ImageNet. Once we start getting quality, annotated/labeled medical images, I'd expect to see progress skyrocket in this area

    Finally, they achieved the 86% state of the art result by pre-training on ImageNet. Training the convolutional neural network on just the medical images from scratch yielded 70%, which is still better than the traditional approach of expert hand selected features.
     
    Last edited:

    DrfluffyMD

    Membership Revoked
    Removed
    2+ Year Member
    Joined
    Dec 15, 2016
    Messages
    1,432
    Reaction score
    1,493
    And our biggest AI proponent here has unfortunately matched into diagnostic radiology rather than IR :(
     

    Naijaba

    Full Member
    10+ Year Member
    Joined
    Apr 2, 2007
    Messages
    1,057
    Reaction score
    111
    And our biggest AI proponent here has unfortunately matched into diagnostic radiology rather than IR :(

    Is this me? ;)

    In any case, I'm starting to think what The Cure said is going to happen first:

    I rarely (read: never) post, but I've gotta pitch in on this one.

    PGY-4 rads here. Like others have mentioned, we've heard about this over and over for a lot of years. It seems unlikely it'll happen any time soon.

    But I've gotta say, I would love for it to happen now. What has happened in other specialties when other entities become capable of doing their job? For a time, things get absolutely SWEET. Look at Gas. CRNAs eventually became a problem (maybe?), but right after they created CRNAs, for the next 10-15 years, anesthesiologists made incredible money billing for what was going on in four ORs simultaneously. The same has happened in multiple other specialties.

    It'll be the same if they ever come up with good AI for rads. For a while life will become amazing, as you'll be able to "supervise" the machine and safely do a ton of studies. How long will the sweetness last? Well, how long do you think it would really take before the machine would be trusted? And after that, waging war with the AMA and ACR, how long do you think it would take to battle through congress to let AI practice medicine. I think 10 years would be an extremely low estimate. More like a career cycle (25-35 years).

    So as far as I'm concerned, please someone out there, discover rads AI now!

    The companies at RSNA are pitching AI/ML as a tool to make radiologists more efficient, while few (no?) businesses are taking the angle that it should replace radiologist. Efficiency improvement is a marketing technique used for years in radiology. Private practices want to squeeze more volume into the same amount of time, and those practices that do will be obtain better contracts. This is how it will enter our workflow, and I expect there will be a surge in salaries (just like The Cure suggested), followed by a decline in reimbursement. It's hard to project where it goes from there, but a the notion of a "Certified Registered Nurse Radiologist" is not hard to imagine if the ML becomes sufficiently advanced.
     
    D

    deleted103644

    AI will not directly replace radiologists anytime soon, but it might make us much more efficient. Whether that's good or bad depends where you stand.

    Efficiency is great if you have a job and a large volume of cases coming in to your practice, but maybe not if you're one of 1000+ new graduates looking for a job.

    AI will let us read more studies faster and safer, but that might mean fewer radiologists are required overall.
     

    harambe4ever

    Full Member
    2+ Year Member
    Joined
    Feb 14, 2017
    Messages
    84
    Reaction score
    69
    ... "Certified Registered Nurse Radiologist" is not hard to imagine if the ML becomes sufficiently advanced.

    Dude. Come on. Your are entitled to your opinions about AI but your naivety is really showing.

    By the way, you still haven't answered my question. If computers and "CRNR's" are going to replace you, why are you going into radiology? Makes no sense.
     
    • Like
    Reactions: 1 user

    Rads312

    Full Member
    2+ Year Member
    Joined
    Apr 23, 2016
    Messages
    130
    Reaction score
    152
    Dude. Come on. Your are entitled to your opinions about AI but your naivety is really showing.

    By the way, you still haven't answered my question. If computers and "CRNR's" are going to replace you, why are you going into radiology? Makes no sense.

    One thing I've learned - never say never. We HAVE to make sure midlevel practitioners don't come anywhere near being compensated for interpretation ANY kind of imaging, wether under a radiologist or not.

    So far, the complexity and wide breadth of the field has been enough to keep midlevel interest at a minimum..but if AI makes radiology 'dummy proof' - you better believe you'll see nurses and PA's trying to start up pilot programs to train midlevel radiology readers...After which will come some BS studies proving equivalent error levels / accuracy..after which will come independent practice.

    Let's all learn a lesson from Anesthesiology here and stay alert.
     
    • Like
    Reactions: 3 users

    DrfluffyMD

    Membership Revoked
    Removed
    2+ Year Member
    Joined
    Dec 15, 2016
    Messages
    1,432
    Reaction score
    1,493
    One thing I've learned - never say never. We HAVE to make sure midlevel practitioners don't come anywhere near being compensated for interpretation ANY kind of imaging, wether under a radiologist or not.

    So far, the complexity and wide breadth of the field has been enough to keep midlevel interest at a minimum..but if AI makes radiology 'dummy proof' - you better believe you'll see nurses and PA's trying to start up pilot programs to train midlevel radiology readers...After which will come some BS studies proving equivalent error levels / accuracy..after which will come independent practice.

    Let's all learn a lesson from Anesthesiology here and stay alert.

    Totally agree. One of my buddy at an upstate NY institution told me his IR department hired a NP but did not specify what can that NP do evidently.

    You would think that NP would try to get consents, go to the clinic, etc. Nope. All they wanted to do is to learn how to do procedures and pester trainee to teach them. It wasn't until my buddy went to his awesome PD did this become resolved.

    Remember, nurses want the best of your procedures. My buddy told me that senior residents were getting consents done for nurses to do procedures. Disgusting.
     
    • Like
    Reactions: 4 users

    Rads312

    Full Member
    2+ Year Member
    Joined
    Apr 23, 2016
    Messages
    130
    Reaction score
    152
    Totally agree. One of my buddy at an upstate NY institution told me his IR department hired a NP but did not specify what can that NP do evidently.

    You would think that NP would try to get consents, go to the clinic, etc. Nope. All they wanted to do is to learn how to do procedures and pester trainee to teach them. It wasn't until my buddy went to his awesome PD did this become resolved.

    Remember, nurses want the best of your procedures. My buddy told me that senior residents were getting consents done for nurses to do procedures. Disgusting.

    Yup. you get it.

    Look at the whole 'DNP' thing. The DNP CRNA's at our institution are allowed to introduce themselves as "Dr. So and So" -- as long as they specify to the patient that they are a NURSING doctor. But come on - the patient is typically already under a versed dose and the layperson will not know the difference..Plus - who's really monitoring what they say when they consent? It's a disgrace for the anesthesiologists. The CRNA's do all the easy procedures while the MD's handle the train wrecks or monitor 4 CRNA's at once. And how much more are they paid? On average about 70 grand more. Yikes.

    In our field, it would probably translate to a midlevel plowing through all the 'normal' studies and leaving the train wrecks on the list for the Radiologist. It's not a reality right now - but we have to make sure it never becomes one. No training ANY midlevel in ANY aspect of image interpretation, ever.

    If nurses push for it - it will be tempting to private practice groups to hire a Nurse to prelim read all your normal Head CT's and have an MD sign off after - could bill sooo much more, but it WILL eventually blow up in our face. I'd honestly much rather be replaced by a computer.
     
    • Like
    Reactions: 2 users

    Naijaba

    Full Member
    10+ Year Member
    Joined
    Apr 2, 2007
    Messages
    1,057
    Reaction score
    111
    By the way, you still haven't answered my question. If computers and "CRNR's" are going to replace you, why are you going into radiology? Makes no sense.

    Well, to answer this question it's because radiology is the most accepting of new technologies. Half the field of angiography exists because of radiologists. Conversely, medicine and surgery are steeped in traditions and hierarchies that impede progress. A good example of this is laparoscopy: It took seventy years from the day the technology was invented to the first lap appy, not because of feasibility, but because it was such a tangent from surgical tradition. Similarly with medicine doctors and the exaltation of the physical exam. I think imaging is incredibly more useful than the physical exam. I would go as far to say that radiology wouldn't have become its own specialty if medicine doctors were willing to learn to read images instead of clinging to historical training.*

    I've met quite a few mentors in IR who were engineers prior to medical school, many of whom continue to apply their background to their speciality. The same can be said about DR. For these people the goal is to improve the delivery of medical care. If that can be done via the invention of a new catheter, imaging modality or even AI, then it should be done. These developments could be developed in other specialties, but the obstacle of tradition is real. Radiology is the best speciality because it welcomes/drives these innovations.

    *Edit: Here's a nice writeup about the history of radiology technologists. They comment briefly on how radiologists came to be:

    By the 1910s, however, a number of physicians began to purchase their own x-ray machines to install in their medical offices. A few even began to specialize as radiologists. In the beginning, these physicians operated the x-ray equipment themselves. Advances in equipment and technique, however, quickly outpaced the physicians' ability to keep up and they gradually found more and more of their time was being eaten up by the mechanics of the x-ray machine, leaving less time for patient contact and treatment.

    So certainly there were some physicians trying to learn to perform x-rays, but there is a tendency to stick with what one knows.
     
    Last edited:

    DrfluffyMD

    Membership Revoked
    Removed
    2+ Year Member
    Joined
    Dec 15, 2016
    Messages
    1,432
    Reaction score
    1,493
    I have a physics background before starting medicine, and my issue with the techie narrative of "medicine impedes progress" is that medicine isn't about putting the highest technology into a person. At the risk of sounding cliche, it's the human element that is the most important thing.

    Through my training, my opinion about medicine went from "what's the most theraputic thing we can do for the patient" to "what's the best we can do for the patient given their wish and desires as well as their circumstances."

    My personal matrix is now "what would I do if patient is a family member of mine."

    Technologies have a lot of upsides, but also serious downsides. AI for example, has unpredictable and catastrophic failure mode in imaging. Whereas a human radiologist tend not to miss obvious abnormalities, computer engines can "glitch out" or bug out. Would you like that to happen on your family member?

    Because of this, adaptation of new technology in medicine needs to be slow and meticulus, which isn't what the silicon valley MO is about.
     
    • Like
    Reactions: 1 users

    wegh

    Full Member
    2+ Year Member
    Joined
    Mar 10, 2017
    Messages
    114
    Reaction score
    77
    Many other fields of medicine will fall to AI before radiology does. For example, internal medicine algorithms for diagnosis and treatment would be easy to program. Nobody relies on physical exam anymore anyways. It's just inputting lab results and procedural results (radiology, endoscopy, angio). It's the procedural results that are subjective and require significant experience. These are the ones that are most immune (relatively) to AI.
    thank you. I just said this in the other thread.

    Also, remember when PACs was going to screw radiologists over?

    The answer is for radiologists to control the technology.
     
    Top