How are applicants dealing with the uncertainty of machine learning?

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

RadsFTW123

Full Member
7+ Year Member
Joined
Jun 19, 2016
Messages
57
Reaction score
26
Full disclosure, I'm a radiologist, and I think it's a fantastic field, would not have picked anything else. However, does anyone really think that we will need the same number of radiologists in 30 years from now? Ezekiel Emanuel is clearly full of it when he says rads will be replaced in 5 years (much more likely to be augmented), but 30 years is a very very long time (think how different the world was in 1986). Is it likely that we will be replaced, or more likely that much fewer of us will be needed as automation takes some of the low-hanging fruit of imaging studies? I'm not sure if this is greater than 50% change, but it's definitely much greater than 0 (as opposed to fields like ortho, neurosurgery who have 0 percent chance of being automated). I honestly don't see how a med student could choose radiology for a 35-40 year career with this huge uncertainty if you had other options.
How do applicants feel about this, and how are program directors still selling the field?
Members don't see this ad.
 
Maybe wait for a single viable piece of evidence before predicting the machine takeover of the field?
 
RadsFTW123 said:
I honestly don't see how a med student could choose radiology for a 35-40 year career with this huge uncertainty if you had other options.
How do applicants feel about this, and how are program directors still selling the field?
Click to expand...

As an applicant I can share my views. Technologically Radiology is not frozen. Back in the days, we were using actual films. When computers started to surface, one would have thought that radiologists would be much more efficient, therefore you would not need as many. Then we went from X-rays, to CT-Scans. I suppose with technological innovation, our depth of knowledge has increased, and this has also created a need for specialization. Thinking that automation will not happen is fallacy. Ultimately it will. I think as automation evolves and becomes reliable enough to be used in practice, the role of radiologists will change as well, no necessarily for the worse. For instance, while automation many not be able to diagnose, it may help with say "complexity" and say well this CT Scan will take X minutes to read while this one will take Y minutes to read. Or for these types of CTs Dr A seems to be better than Dr B, so we should route these to Dr A. There are many ways automation can help that in the end may prove more helpful that one may think. Right now thinking that computers can replace the radiologist is another fallacy, but they can surely help. For sure, the workflow can be improved and efficiency can be improved.
I am not even talking about all the possibilities for in-situ pathology (gross and perhaps microscopic with no biopsy- which would like become some sort of new sub-scpecialties), or other device advances (i.e. did you ever think of what could happen if we could uncouple X-ray production and delivery?)
I am not going into radiology "just to read" (like someone who would want to go to private practice for instance - before I get yelled at : nothing wrong with that if that's what floats your boat). I am more into it for all the opportunities it offers, all the technological advances that will come, and hope to even be part of some of them. Between data / image analysis and device development, I think we will more than enough for the radiologists to be busy. Time will tell.

Finally, I would say that despite all the uncertainty, there is really nothing else I would want to do. So for me it is a no brainer at this point. I am confident that I will be able to pay my loans, have a regular house and a car (that need not be German berline). I think I will be happy every morning when going to work. The question is "what type of work"? That I can't answer...
 
Last edited:
Members don't see this ad :)
It's my personal philosophy to never bet against technology, but yeah I could see some decline in the demand for radiologists in a timeline on the order of decades. For me, I also like IR, though I want to practice DR as well. DR is a hedge against back problems after years of wearing lead all day, and IR is my hedge against AI overtaking things in my lifetime. The fault in my plan, of course, is that my back is more likely to go out later on when AI would be taking over, but hopefully I will be financially independent at that point. If you are a radiologists today, the average incomes are such that you should be building multiple income streams. This is smart for any physician. My wife also has business ventures, small right now but will increase as I get some income to support them more. We also plan to have some rental properties, either let a prop manager run them or she will oversee them. Cash flow is king, so I think it would be unwise to rely solely on your radiology paycheck for the rest of your life, even if AI never makes it. It takes money to make money, and I foresee any average radiologist having plenty of capital to start ventures for the foreseeable future.


Sent from my iPhone using SDN mobile
 
1 in 5 doctors still use paper charts.

Think about that in the context of implenting and being overtaken by automation in our lifetimes....
 
This topic comes up a lot at my interviews (albeit, I bring it up). My personal vision is:

1. We will see radiology technologists trained to read images in the same way path techs do the first screen of slides. As evidence of this - In IR there are already nurse practitioners, PAs, and scrub techs gaining access and putting in lines. I think introductory pattern-recognition can be taught without the full MD / radiology residency training.
2. The radiology techs will wield machine learning to further verify their assumptions.
3. When 1 and 2 are in-place, machine learning algorithms will begin auto-generating notes for "normal" reads. Radiologists will not review normal reads unless the referring physician specifically requests a re-read.
4. Gradually the same systems will auto-generate notes for common pathology.

Radiologists will be the overseers of such infrastructure, and will always look over "complex, non-normal" findings (as identified by the algorithm + rad tech).

This paradigm will provide the radiologist more time to see patients and/or to pursue research interests.
 
That's a cute idea, but we would probably need about 5 percent as many radiologists as now. Competition will be fierce (and pay almost nothing) for these "research and complex scan review" jobs.

AI if anything will be like a super functional radiologist assistant, I can't see them training anymore people with that skill set.
Naijaba said:
This topic comes up a lot at my interviews (albeit, I bring it up). My personal vision is:

1. We will see radiology technologists trained to read images in the same way path techs do the first screen of slides. As evidence of this - In IR there are already nurse practitioners, PAs, and scrub techs gaining access and putting in lines. I think introductory pattern-recognition can be taught without the full MD / radiology residency training.
2. The radiology techs will wield machine learning to further verify their assumptions.
3. When 1 and 2 are in-place, machine learning algorithms will begin auto-generating notes for "normal" reads. Radiologists will not review normal reads unless the referring physician specifically requests a re-read.
4. Gradually the same systems will auto-generate notes for common pathology.

Radiologists will be the overseers of such infrastructure, and will always look over "complex, non-normal" findings (as identified by the algorithm + rad tech).

This paradigm will provide the radiologist more time to see patients and/or to pursue research interests.
Click to expand...
 
Naijaba said:
3. When 1 and 2 are in-place, machine learning algorithms will begin auto-generating notes for "normal" reads. Radiologists will not review normal reads unless the referring physician specifically requests a re-read.
Click to expand...

Wow? Who will pay the lawsuit settlement if something is missed? I doubt any company in its right mind will sell an algo and will assume responsibility for its predictions. They would likely be bankrupt within the first year of operation.
How many surgeons and other able physicians wait for the Radiologist read before proceeding. Do you really think they can't read some of their images?

I think algorithms will assist radiologists, will help with the workflow, will also help with the fee structure (no reason to be paid the same when a read takes you more time, or requires more effort/expertise). I think these innovation will increase productivity and maybe even income all in all. As for the need for radiologist, it is impossible to predict what it will be. Technology is moving fast, population aging which means more imaging (although I think for the first time in history life expectancy has gone down this year in the US by 0.1 year). Additionally a lot of retirements will happen. Too much of a gamble no to pursue radiology because of the ever changing landscape. It has changed in the past, keeps changing now and will keep changing. Radiologists will stick around. No doubt about it.
 
Naijaba said:
This topic comes up a lot at my interviews (albeit, I bring it up). My personal vision is:

1. We will see radiology technologists trained to read images in the same way path techs do the first screen of slides. As evidence of this - In IR there are already nurse practitioners, PAs, and scrub techs gaining access and putting in lines. I think introductory pattern-recognition can be taught without the full MD / radiology residency training.
2. The radiology techs will wield machine learning to further verify their assumptions.
3. When 1 and 2 are in-place, machine learning algorithms will begin auto-generating notes for "normal" reads. Radiologists will not review normal reads unless the referring physician specifically requests a re-read.
4. Gradually the same systems will auto-generate notes for common pathology.

Radiologists will be the overseers of such infrastructure, and will always look over "complex, non-normal" findings (as identified by the algorithm + rad tech).

This paradigm will provide the radiologist more time to see patients and/or to pursue research interests.
Click to expand...

My friend, you have a lot to learn about radiology. and for the love of god I hope you're not sharing this plan with your interviewers. Path techs don't sign off on studies as normal, they mainly assess adequacy of the tissue sample before it's interpreted. Rad Techs already make sure we get the best images possible, it's always been their main role.

Oh and regarding techs/AI signing off on normals? Come back to me after your first week of call, and tell me how many of the studies you called "normal" actually were negative. Most misses aren't misinterpreted, they're missed altogether, often being very subtle even in retrospect. Your plan is just a recipe for disaster, but it only puts me more at ease about the Silicon Valley takeover because those types are probably even more clueless about how radiology actually works.
 
Naijaba said:
This topic comes up a lot at my interviews (albeit, I bring it up). My personal vision is:

1. We will see radiology technologists trained to read images in the same way path techs do the first screen of slides. As evidence of this - In IR there are already nurse practitioners, PAs, and scrub techs gaining access and putting in lines. I think introductory pattern-recognition can be taught without the full MD / radiology residency training.
2. The radiology techs will wield machine learning to further verify their assumptions.
3. When 1 and 2 are in-place, machine learning algorithms will begin auto-generating notes for "normal" reads. Radiologists will not review normal reads unless the referring physician specifically requests a re-read.
4. Gradually the same systems will auto-generate notes for common pathology.

Radiologists will be the overseers of such infrastructure, and will always look over "complex, non-normal" findings (as identified by the algorithm + rad tech).

This paradigm will provide the radiologist more time to see patients and/or to pursue research interests.
Click to expand...

.....Yes this is not the case. You're pre-medical and really haven't the slightest clue as to what you are talking about... Where did you come up with this delusion? Machine learning has been around for some time and has been in use by pathology for years now and has, like many have stated, have augmented their work to make them more efficient at their job and are doing the menial tasks such as cell counting etc. You know there are also EKG machines that are used in the ED that "diagnose" the pathology seem on the rhythm strip and know that ED physicians hardly pay more than a half a second of attention to the machines interpretation? They often don't agree with its findings and move on with their day.

Rad techs reading imaging.... LOL
 
Naijaba said:
As evidence of this - In IR there are already nurse practitioners, PAs, and scrub techs gaining access and putting in lines.
Click to expand...

This is logic is deeply flawed. NPs and PAs gaining ground in their scope of practice when it comes to clinical procedures is not analogous to rad techs being able to interpret an image. Mid-levels receive graduate level education and have clinical procedural skills incorporated into their curriculum. Interpretation of imaging is not remotely within the purview of rad tech training. Radiologist have no overlapping roles or skill sets with radiology techs. They are essentially two disparate occupations within the same field.
 
clutch21 said:
This is logic is deeply flawed. NPs and PAs gaining ground in their scope of practice when it comes to clinical procedures is not analogous to rad techs being able to interpret an image. Mid-levels receive graduate level education and have clinical procedural skills incorporated into their curriculum. Interpretation of imaging is not remotely within the purview of rad tech training. Radiologist have no overlapping roles or skill sets with radiology techs. They are essentially two disparate occupations within the same field.
Click to expand...
Yeah, there are some good rad-techs who will call you up when the patient is in the scanner and tell you they see a SAH or a huge PE or lung white-out or something like that, but at most it just saves you a few minutes because you'd be looking at those stat cases quickly anyway. In no way, shape, or form are they looking for (or have the time or training) to look for anything subtle.
 
So much contention about my post! Let me respond as best I can...

RadsFTW123 said:
That's a cute idea, but we would probably need about 5 percent as many radiologists as now. Competition will be fierce (and pay almost nothing) for these "research and complex scan review" jobs.

AI if anything will be like a super functional radiologist assistant, I can't see them training anymore people with that skill set.
Click to expand...

Do you mean five percent less (100% -> 95%)? Or radiology jobs will go down to 5% of what they are now (100% -> 5%)? I think the impact of machine learning + rad tech readers will be low initially, but will ramp up over time.

redoitall said:
Wow? Who will pay the lawsuit settlement if something is missed? I doubt any company in its right mind will sell an algo and will assume responsibility for its predictions. They would likely be bankrupt within the first year of operation.
How many surgeons and other able physicians wait for the Radiologist read before proceeding. Do you really think they can't read some of their images?

I think algorithms will assist radiologists, will help with the workflow, will also help with the fee structure (no reason to be paid the same when a read takes you more time, or requires more effort/expertise). I think these innovation will increase productivity and maybe even income all in all. As for the need for radiologist, it is impossible to predict what it will be. Technology is moving fast, population aging which means more imaging (although I think for the first time in history life expectancy has gone down this year in the US by 0.1 year). Additionally a lot of retirements will happen. Too much of a gamble no to pursue radiology because of the ever changing landscape. It has changed in the past, keeps changing now and will keep changing. Radiologists will stick around. No doubt about it.
Click to expand...

I don't think radiologists will be out of a job, I just think they will be overseeing a cohort of workers and a computational infrastructure. As to liability, when a PA/NP/Nurse Anesthetist performs a procedure they are covered by insurance. This is why a pure-computational solution is a long way off, but a hybrid rad tech + machine learning infrastructure could be feasible. I see credentialed workers doing initial reads and full reads for normal images. My belief is that if reimbursement keeps decreasing, hospitals and radiology departments will have to employ some techniques to keep volume up, this is just one approach.

cowme said:
My friend, you have a lot to learn about radiology. and for the love of god I hope you're not sharing this plan with your interviewers. Path techs don't sign off on studies as normal, they mainly assess adequacy of the tissue sample before it's interpreted. Rad Techs already make sure we get the best images possible, it's always been their main role.

Oh and regarding techs/AI signing off on normals? Come back to me after your first week of call, and tell me how many of the studies you called "normal" actually were negative. Most misses aren't misinterpreted, they're missed altogether, often being very subtle even in retrospect. Your plan is just a recipe for disaster, but it only puts me more at ease about the Silicon Valley takeover because those types are probably even more clueless about how radiology actually works.
Click to expand...

I've mentioned the plan to nearly every interviewer, and most (all?) were supportive. One IR stated the exact same belief before I had a chance to give my little speech. I admit that "rad tech" was the wrong title because it causes confusion with current role of rad techs. What I'm envisioning is a physician-assistant like role with additional training in radiology. Maybe 2 years of physician assistant training and 2 years of reading room - a "radiologist assistant." It's not just Silicon Valley, nearly ever academic department is devoting resources to machine learning research.

vlaka90 said:
.....Yes this is not the case. You're pre-medical and really haven't the slightest clue as to what you are talking about... Where did you come up with this delusion? Machine learning has been around for some time and has been in use by pathology for years now and has, like many have stated, have augmented their work to make them more efficient at their job and are doing the menial tasks such as cell counting etc. You know there are also EKG machines that are used in the ED that "diagnose" the pathology seem on the rhythm strip and know that ED physicians hardly pay more than a half a second of attention to the machines interpretation? They often don't agree with its findings and move on with their day.

Rad techs reading imaging.... LOL
Click to expand...

I'm not a pre-med...I see that my SDN tag is out of date. I've written about this before, but machine learning of the past was based on fundamentally different principles that emerging techniques such as deep learning. Yes, mammography and EKG readers have failed to realize the vision of machine learning that works. But, these operate on historical algorithms. I'm happy to discuss more, and even go into the mathematics. Maybe it'd be useful to have a separate topic on the mathematics of machine learning for radiologists. There's a lot of misconceptions about what is and is not possible. Since you mention pathology, here's a really great article about how these new methods are changing that field, beyond the innovations introduces years ago: http://www.nature.com/articles/ncomms12474

Oh, and for kicks, here's another article where the same algorithms are offering new solutions for quadriplegics in neurosurgery: http://www.nature.com/nature/journal/v533/n7602/full/nature17435.html

clutch21 said:
This is logic is deeply flawed. NPs and PAs gaining ground in their scope of practice when it comes to clinical procedures is not analogous to rad techs being able to interpret an image. Mid-levels receive graduate level education and have clinical procedural skills incorporated into their curriculum. Interpretation of imaging is not remotely within the purview of rad tech training. Radiologist have no overlapping roles or skill sets with radiology techs. They are essentially two disparate occupations within the same field.
Click to expand...

I agree that my nomenclature is correct. My vision is a new training pathway (e.g. a "radiologist assistant") where there is extensive (2+ years) of reading room preparation in addition to a year or two of basic sciences / medicine.
 
Members don't see this ad :)
Sorry, but this is akin to saying mid levels should be doing appys and choles and saving the complex cases for the surgeons.

It is a profoundly flawed plan that's frankly a little insulting and it just shows your ignorance with the field as a whole.

But hey, you're the Med student and im the one with experience in the field so who am I to question you
 
Last edited by a moderator:
Yeah, just a 4th year here, but my perspective is that when a radiologist makes a read, he/she is making a judgment based on the entirety of his or her medical knowledge. You never know what you are going to need to know and when, thus you have to learn a ton through med school and residency etc. no real way to short change that transfer of knowledge. Very different than training a mid level to recognize a small or even moderate number of patterns.


Sent from my iPhone using SDN mobile
 
Naijaba, I appreciate your ideas, but honestly I don't understand how this could possibly work out well. First, I don't see how hiring essentially radiology PAs would be a logical approach to keep volumes up as you suggest. I am not a radiology resident yet, but it is my understanding that a significant portion of making a read is determining what is normal and what is not. Wouldn't you want the expert radiologist to be the one making the nuanced call on "normal" studies rather than a radiology mid-level who could easily miss something (even if we are putting a lot of trust in AI)? How would the training differ between a radiologist and a radiology mid-level other than time duration (like others have mentioned you need the entirety of your medical knowledge to make accurate reads)? As far as liability is concerned for PAs, CRNAs, etc. in other fields, MD's are still liable legally for their decisions and patient outcomes. Thus, I doubt a radiologist would blindly trust AI+mid-level radiology provider on the vast majority of scans and would just end up over reading them anyway. And lastly, why would you even entertain the idea of proposing mid-levels to do some of our work in the first place (even if they could do it well which I don't think they could)? In almost every other field of medicine (EM, Anesthesia, Primary Care, etc.) mid levels were introduced to "help" with the easy monotonous stuff of the respective field only to grow into a significant threat later down the line when they feel they are just as qualified and demand more autonomy. Just look at CRNAs and anesthesia and what a mess that has turned that field into. Mid levels make some sense in procedural specialties, but not in diagnostic radiology.
 
Making a call of "normal" is not quite the same as "absence of all pathological patterns." Normal is not infrequently a tough (but necessary) call to make.

I don't really see the point of diluting the workforce with half-trained readers.
 
I should be clear that the machine-learning + mid-level provider idea is just one model for AI integration. I've had several discussions about machine learning and how it will be incorporated into radiology practice. Here's another one worth thinking about. Existing PACS vendors could incorporate an "App Store for algorithms" into their PACS. I know of at least one major vendor who is doing this. The idea is that developers can come up with specific algorithms such as RECIST and sell them on the PACS. A radiologist could select two images, download the "RECIST" algorithm, and run it. I like this model a lot because it reduces time-to-market and specific algorithms could be made for specific studies. For example, hemorrhage maps in neuro or LI-RADS calculations for HCC patients. The radiologist could then configure the system to automatically run these algorithms on specific studies.
 
cowme said:
Sorry, but this is akin to saying mid levels should be doing appys and choles and saving the complex cases for the surgeons.

It is a profoundly flawed plan that's frankly a little insulting and it just shows your ignorance with the field as a whole.

But hey, you're the Med student and im the one with experience in the field so who am I to question you
Click to expand...
Thank you for saying what everyone is thinking...
 
To address the original question from the OP. Admittedly when I chose radiology as an MS3/4 "AI" and "Machine learning" were not part of my lexicon and I had no idea that this new technology even existed. However, even though I am not an applicant this year but rather an intern, I would still pick radiology again despite the uncertainty of AI. For one, I enjoy the work and people in radiology by far the most compared to any other specialty. With each passing rotation of my intern year, I feel more comfortable and happy with my choice of radiology. Needless to say, I can't wait to finish up this year.

As far as the future of radiology is concerned. I believe AI will have a significant impact on the field. In the short term (next 3-10 years) I think the AI onboarding process into the radiology work flow will initially create MORE work for radiologists as we figure out how to most appropriately use the technology. It will be a redundant "second opinion" so to speak initially. This could actually create a shortage of radiologists for some time. However, in the medium term (10-20 years), radiologist will become more savvy with the the tech and learn to trust it more, thus making them more efficient correcting the job market as less radiologists will be able to to more work. In the long term (20 years plus), AI will continue to get better (but never completely replace, largely for legal reasons) and the need/ability for radiologists to produce reads will become much more efficient and less important. The job market at this point could substantially turn south if all we do is produce reads, however, I would hope that by then our field as a whole will evolve to produce value in other ways than just making imaging reads. Some of the things I hope to see are radiologist becoming more of a diagnostician consultant, utilizing big data and AI to help clinicians make more informed diagnostic decisions (somewhat similar to that recent JAMA article). Essentially, Imaging 3.0 becoming fully realized with radiologists coming out of the reading room more often and engaging with other physicians and patients. Ultimately, the future radiology job market will depend on how we evolve as a field with this technology. Also, anyone trying to make bold steadfast predictions 20 years out is bound to look like a fool. Nobody can truly predict the future so it's not worth getting completely worked up about what the field will look like that far out. I, for one, am cautiously optimistic about radiology's future.
 
I wouldn't even say it's that close. Usually promising research and multi center trials precede the pricey vendor specific software packages by about 5-10 years. Much Longer to be standard of care. And everything is still purely theory at this point.
 
cowme said:
Sorry, but this is akin to saying mid levels should be doing appys and choles and saving the complex cases for the surgeons.

It is a profoundly flawed plan that's frankly a little insulting and it just shows your ignorance with the field as a whole.

But hey, you're the Med student and im the one with experience in the field so who am I to question you
Click to expand...

vlaka90 said:
Thank you for saying what everyone is thinking...
Click to expand...

I didn't mean to insult, and you're right - mid-levels should not be doing appendectomies or cholecystectomies, even if they are the "easier" surgeries.
 
I think there's probably three phases of thinking about AI and deep learning in radiology:

1) Med student: Eager to talk about it because it sounds cutting edge and sounds knowledgable in interviews. No real idea of what it entails.

2) Resident: Dismissive of a technology that's going to render years of hard work cheap. No incentive to love something that endangers her or his job, but also appropriately skeptical of handwaving a technology that claims to accurately duplicate a higher-level brain function.

Then the path diverges:

3a) The PP attending: Willing to adopt any technology that increases his or her throughput and bottom line. No other thought to the matter.

3b) The academic attending: Willing to hype any technology that generates publications. Some kind of weak celebrity for being a cutting edge pundit is a plus.
 
And for the OP, the answer is: Who knows.

It's like wondering if it's a bad idea to open a driving school since they're developing self-driving cars. Who knows. Maybe. Maybe not.
 
I wish deep learning could scan the charts and find me pertinent medical history for the study at hand. That would be the best possible use of current technology to a Radiologist IMO.
 
I think our field will be in jeopardy from AI when other we no longer have cab-drivers, pilots, postal workers etc...also many private practices such as my own, maintain relevance bc we are there to do light IR/breast imaging-hard to see how AI can replace us doing this in the next 30-40 years of so (will become concerned when sex-bots completely replace human prostitutes)
 
tco said:
I wish deep learning could scan the charts and find me pertinent medical history for the study at hand. That would be the best possible use of current technology to a Radiologist IMO.
Click to expand...

you mean "pain" isn't enough info for you?
 
vlaka90 said:
.....Yes this is not the case. You're pre-medical and really haven't the slightest clue as to what you are talking about... Where did you come up with this delusion? Machine learning has been around for some time and has been in use by pathology for years now and has, like many have stated, have augmented their work to make them more efficient at their job and are doing the menial tasks such as cell counting etc. You know there are also EKG machines that are used in the ED that "diagnose" the pathology seem on the rhythm strip and know that ED physicians hardly pay more than a half a second of attention to the machines interpretation? They often don't agree with its findings and move on with their day.
Click to expand...

cowme said:
My friend, you have a lot to learn about radiology. and for the love of god I hope you're not sharing this plan with your interviewers. Path techs don't sign off on studies as normal, they mainly assess adequacy of the tissue sample before it's interpreted. Rad Techs already make sure we get the best images possible, it's always been their main role.
Click to expand...

I want to chime in one two particular uses of machine learning in pathology.
1) Gyn cytology (Pap smears): Pap smears are primary read by techs. Then machine learning with neural networks started with the now-defunct Papnet in the 1990s and has evolved to include Thinprep, Surepath, etc. In Papnet, 64 images per slide were scanned and various classifiers were used. Then the same 64 images were displayed for manual evaluation. My understanding was that this system was designed to speed up review of Pap smears and reduce interobserver variability. With Thinprep and Surepath, the systems screen for potential abnormalities, and then forces a manual review of a minimum number of areas, which would include the potentially abnormal areas.
2) Cell Counting: FDA approved algorithms are available on slide scanners for certain immuno stains, but the antibodies from the specified manufactures must be used in order to use the software within the guidelines of the FDA approval. With that being said, a pathologist would still have to fairly computer-savvy in order to use these algorithms. I don't mean being able to code, but at least being extremely comfortable with computers.

I can see a role for machine learning in radiology, but it would not replace radiologists. Perhaps machine learning could be used in automated segmentation of hemorrhage, e.g. based on Hounsfield units, for measurement and comparison of hemorrhage dimensions. Another use for neuroradiology could be in monitoring of multiple sclerosis -- counting plaques and comparing with older MRI scans. In these cases, a radiologist would still have to decide if the CT hyperdensity represents hemorrhage, or if the oval shaped MRI hyperintensities could be MS plaques.
 
kchan99 said:
I want to chime in one two particular uses of machine learning in pathology.
1) Gyn cytology (Pap smears): Pap smears are primary read by techs. Then machine learning with neural networks started with the now-defunct Papnet in the 1990s and has evolved to include Thinprep, Surepath, etc. In Papnet, 64 images per slide were scanned and various classifiers were used. Then the same 64 images were displayed for manual evaluation. My understanding was that this system was designed to speed up review of Pap smears and reduce interobserver variability. With Thinprep and Surepath, the systems screen for potential abnormalities, and then forces a manual review of a minimum number of areas, which would include the potentially abnormal areas.
2) Cell Counting: FDA approved algorithms are available on slide scanners for certain immuno stains, but the antibodies from the specified manufactures must be used in order to use the software within the guidelines of the FDA approval. With that being said, a pathologist would still have to fairly computer-savvy in order to use these algorithms. I don't mean being able to code, but at least being extremely comfortable with computers.
Click to expand...

Neural networks weren't successful in the past, except on trivial tasks such as numeric digit recognition. The recent emergence (or resurgence?) of machine learning in radiology is largely due to the incredible breakthrough that happened at the 2012 ImageNet Classification Challenge. Alex Krizhevsky, et. al built a deep convolutional neural network that outperformed every other classifier by a huge margin. For those unfamiliar, the ImageNet Challenge involves classifying some 1 million+ images into 1,000 different classes. The classes are so specific that humans have trouble telling them apart (e.g. frog vs. toad, alligator vs crocodile, white husky vs white wolf). The error rate before AlexNet was a whopping 26.1%, with AlexNet bringing the error rate down to 15.3%. A 10% reduction was unprecedented and caused a paradigm shift in machine learning applied to imaging. It showed that deep convolutional neural networks could potentially outperform humans at image recognition tasks. Humans beings are capable of labeling ImageNet items with an error rate of 5.1%. In 2014, Google won the ImageNet Challenge with GoogleNet, a massive deep learning classifier that brought the error rate down to 6.67%. Then, in 2015 super-human performance was achieved by He, et. al with an error rate of 4.94%. It's important to note that going from 6.67% --> 4.94% is nearly as significant as going from 26.1% --> 15.3%, because those last few percentage points are the most difficult. Similar results are being achieved in fields like speech recognition.

The last point I want to make is that neural networks are not walled gardens. You can follow this tutorial and achieve near-state of the art results on your desktop. You only need a Nvidia GPU and knowledge of Linux to get going. This is what is meant by the machine learning renaissance. The tools are available to anyone and we're going to see some very interesting startups once people with computer science backgrounds enter the medical field.


AlexNet: https://papers.nips.cc/paper/4824-i...n-with-deep-convolutional-neural-networks.pdf

GoogleNet: https://www.cs.unc.edu/~wliu/papers/GoogLeNet.pdf

Super-Human Performance: https://arxiv.org/abs/1502.01852
 
Last edited:
I'm aware that the machine learning world has expanded. I was hoping to give a couple of examples where machine learning has either improved quality of care or speeded up mundane tasks. Anyway, physician judgement is important because the patient's clinical history must be incorporated.
 
PCIRISH52 said:
https://medium.com/the-mission/deep...nges-and-opportunities-d2eee7e2545#.q81b48u0m

This article is one of the best I've read on the challenges and potential of AI in radiology.
Click to expand...

Yes this is a very good piece. Dr. Channin has a similar background to my own. I also worked as a software developer for a PACS vendor before attending medical school. I think he hits the nail on the head with the "data in, data out, regulation" problem. There are challenges with getting good training data, problems with making the data provide useful results, and enormous challenges getting FDA approval / connecting a device to any given hospitals network.
 
I am shocked candidates are still picking radiology. While its true that we won't be replaced immediately. We could be replaced half way through your career & never fully achieve your earning potential.

Also what will happen when one "Augmented radiologist" can do the job of several radiologists? The job market would suffer again. We hear stories of old PP docs who won't quit their job. What happens when machine learning makes it even easier for them to work & stay in their jobs because all they have to do is agree with prelim reports generated by computers?
 
I wasn't aware of the advances in machine learning in radiology when I applied 4 years ago. There was no active discussion at that time. Now its just a ticking time bomb.

Its actually in society's best interest that we replace expensive radiologist's with cheap machines. Its probably the best sector in medicine for easy cost cutting utilizing AI since a large portion of our work is pattern recognition & data interpretation. Theres alot more funding going into AI reading medical imaging than there is in AI taking over other kinds of medical jobs.

Again, I don't think it will replace all radiologists over night. But what happens when ONE radiologist with AI can do the job of two radiologists? The job market is cut in half.
 
Last edited:
You are oversimplifying things. AI introduction into medicine will likely be messy and result in even more data that someone (radiologists) will need to interpret. Just because a job is done on a computer does not mean it is easy for a computer to do it.
 
meister said:
You are oversimplifying things. AI introduction into medicine will likely be messy and result in even more data that someone (radiologists) will need to interpret. Just because a job is done on a computer does not mean it is easy for a computer to do it.
Click to expand...

Perhaps. Integration will likely be messy. But spending 6 years in residency and fellowship is alot of time to sacrifice based on an a likely assumption. If you honestly disk-like every other field in medicine than stick with it as we will probably still need a few radiologists in the future. But a warning to those who are on the fence between other fields and those considering going to weaker radiology programs, you could be on the chopping block. Don't assume that because you invest 6 years of your life learning that you are owed a life long job by society. The moment your expensive salary can be replaced, it will be. Clinicians are harder to replace.
 
iCY said:
Perhaps. Integration will likely be messy. But spending 6 years in residency and fellowship is alot of time to sacrifice based on an a likely assumption. If you honestly disk-like every other field in medicine than stick with it as we will probably still need a few radiologists in the future. But a warning to those who are on the fence between other fields and those considering going to weaker radiology programs, you could be on the chopping block. Don't assume that because you invest 6 years of your life learning that you are owed a life long job by society. The moment your expensive salary can be replaced, it will be. Clinicians are harder to replace.
Click to expand...
I agree, radiology is a very risky career choice (and that's the whole supposed upside of medicine is that it's low risk). I'm pretty far along in my training and am doing IR, but I definitely would not choose radiology today given AI concerns. I don't think the chance of a significant change in radiology is really that high, but a 35 year time horizon is an eternity and any risk of unemployment after 10 years of training and very few other marketable skills is an unacceptable risk when you could do medicine or surgery instead.
 
Every field has a disaster scenario, such as midlevel encroachment (optometrists, crnas), to blood tests replacing your fields highest volume procedures (colon cancer dna testing vs colonoscopy, immunotherapy vs XRT).

The only difference with these and radiologys doomsday scenario is that radiologys doomsday scenario remains theoretical
 
cowme said:
Every field has a disaster scenario, such as midlevel encroachment (optometrists, crnas), to blood tests replacing your fields highest volume procedures (colon cancer dna testing vs colonoscopy, immunotherapy vs XRT).

The only difference with these and radiologys doomsday scenario is that radiologys doomsday scenario remains theoretical
Click to expand...

True it is theory at this point. But at hte rate the technology is developing it is on a trajectory to replace a large portion of our workload. If you speed up one radiologist 50%, another radiologist will get fired. The job market may turn into what it is for pathology. Limited jobs because you only really need a handfull of pathologists to run an entire department.

Again it is theory but future candidates need to know there is a chance they are signing up for a costly 7 year mistake. Everyone needs to decide if its worth it for them to take this theoretical risk. Personally the theory (& proof so far) would have been enough to push me to match a different residency for job security. But! Its too late for us haha. I'd rather finish now & take my chances than leave half way to scramble into some malignant open spot.

Also what do you think will happen when teleradiology gets a hold of something like this?
 
iCY said:
True it is theory at this point. But at hte rate the technology is developing it is on a trajectory to replace a large portion of our workload. If you speed up one radiologist 50%, another radiologist will get fired. The job market may turn into what it is for pathology. Limited jobs because you only really need a handfull of pathologists to run an entire department.

Again it is theory but future candidates need to know there is a chance they are signing up for a costly 7 year mistake. Everyone needs to decide if its worth it for them to take this theoretical risk. Personally the theory (& proof so far) would have been enough to push me to match a different residency for job security. But! Its too late for us haha. I'd rather finish now & take my chances than leave half way to scramble into some malignant open spot.

Also what do you think will happen when teleradiology gets a hold of something like this?
Click to expand...

" If you speed up one radiologist 50%, anotherradiologist will get fired."

Check your math there bro...
 
Yea bro that just means when you get fired I'm gonna get all your monneeeeyyyy
 
futureradres12345 said:
" If you speed up one radiologist 50%, anotherradiologist will get fired."

Check your math there bro...
Click to expand...

Haha I love how you think the math needs to be 100%. In my department when one person leaves the others just work more hours to compensate. One radiologist works at 150%... Why keep the guy working 50 but hey let's pretend I'm wrong and business people will be kind to you
 
You must log in or register to reply here.

Similar threads

R
Now that the dust has settled, how often are you reading remotely from home?
Replies
3
Views
1K
GadRads
G
M
Learning PAD in program without a lot of it
Replies
3
Views
983
irwarrior
I
Steve_Zissou
The use of Midlevels in image interpretation
Replies
1
Views
1K
GadRads
G
O
Are there any rads jobs that force daytime workers to work deep nights?
Replies
2
Views
647
MadRadLad
M
R
Breast fellowship applicants
Replies
8
Views
3K
Rubix8
R
Top