Computers, Automation and Radiology Future- NYTimes

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

entrepreneurMD

Full Member
7+ Year Member
Joined
Feb 28, 2014
Messages
98
Reaction score
29
Hey guys,

I'm a second year student at a US MD school. I am very interested in Radiology as a specialty and was very optimistic about its future. But I woke up to this article this morning and it got me pretty rattled:

http://www.nytimes.com/2016/02/29/t...&gwh=337C59FE2EFEB1CC89D920F3A93C7CA4&gwt=pay


The company that they mentioned found breast cancer with 50% better accuracy with lower costs. Thats impressive and can only go up. How real is this threat of computers replacing radiologists and is it still wise for me to consider this field or should I look elsewhere 🙁
 
Nice try man. ECG machines cant even catch afib vs. flutter...
 
Hey guys,

I'm a second year student at a US MD school. I am very interested in Radiology as a specialty and was very optimistic about its future. But I woke up to this article this morning and it got me pretty rattled:

http://www.nytimes.com/2016/02/29/t...&gwh=337C59FE2EFEB1CC89D920F3A93C7CA4&gwt=pay


The company that they mentioned found breast cancer with 50% better accuracy with lower costs. Thats impressive and can only go up. How real is this threat of computers replacing radiologists and is it still wise for me to consider this field or should I look elsewhere 🙁

LOL. Hilarious first response to your article.

Where to begin?

Yes, eventually technology will supplant more basic algorithmic functions such as identifying certain specific lesions on various types of imaging in radiology. Should you be concerned? Yes and no.

Expect technology to create efficiencies and reduce costs across ALL specialties. Everything in this world trends towards it being done better, more quickly, and more cheaply... usually through technology and, in the case of healthcare, mid-level providers (but this is a separate issue). If you think about it, most of healthcare practice is algorithmic. What keeps you and I with job prospects after we graduate from med school is being able to think outside the algorithm and adjust for exceptions. Already computers can pretty much do most of what an internist does... all that's required is the input. It can be argued that with some demographic information along with CENTOR criteria, an algorithm could determine if someone should be treated for Strep and that if we did this for just about every other diagnosis we would leave internists out of jobs. Scary, right? But then there's the issue of appropriateness in even using something like CENTOR when a patient could have a hundred other things going on instead of Strep - something that arguably a computer can/cannot do and something human physicians can offer.

Lag times and # of diagnoses. Algorithms require clinical research and there is an inherent lag time in "reprogramming the algorithm" so to speak and implementing it to diagnose. With data mining, we can have numbers to perform some serious clinical research and find patterns/correlations we can reprogram our cute little algorithm with. Someone needs the medical knowledge to make sense out of the data. That's you and me.

Legal aspects. Radiologist MDs will still be required. Our legal system isn't designed to provide indemnity in healthcare. Someone has to be held accountable for medical decisions. This can't be a computer algorithm. Even if the algorithm is better than 50% of radiologists at detecting breast cancer (more of a statement to draw attention from readers than a reality), a radiologist will be required to make the final call and sign off. Technology in this sense may be a facilitator vs. a competitor of sorts to the radiologist.

Shifting radiologist roles. Look, the reality is humans are required for research, innovation, and entrepreneurship. What truly TRULY offers value isn't memorization and algorithmic thinking... computers can do that waaaaay better than us. The age of information technology in healthcare is here and we need to accept it. We should stop seeing it as a threat and instead embrace it as something that can help us to more accurately diagnose as radiologists and to reduce costs for our patients to make healthcare more affordable.

How realistic is it that technology will make you and I obsolete as radiologists in our lifetimes? Not very. Eventually however, current radiologist roles will be obsolete and will be more focused on research.

As for OBGYN, be my guest in going into a field plagued with litigation and a field that will also be subject to mid-tier provider takeover (already beginning to be the case for all primary care professions). It's a beautiful field, but don't expect to have it any easier.
 
Last edited:
Hey guys,

I'm a second year student at a US MD school. I am very interested in Radiology as a specialty and was very optimistic about its future. But I woke up to this article this morning and it got me pretty rattled:

http://www.nytimes.com/2016/02/29/t...&gwh=337C59FE2EFEB1CC89D920F3A93C7CA4&gwt=pay


The company that they mentioned found breast cancer with 50% better accuracy with lower costs. Thats impressive and can only go up. How real is this threat of computers replacing radiologists and is it still wise for me to consider this field or should I look elsewhere 🙁


If you don't feel comfortable about your choice of doing radiology, simply don't do it and choose another field.
 
I think this type of AI-aided diagnosis was coming sooner or later. If I'm not mistaken, this happens in pathology to some level as we speak. I am not sure why the company did not first try to analyze EKG's, which is a 1-D data and presumably an easier problem to attack. Maybe it has to do with the availability of digitized forms of data where radiology as a field has done an excellent job.

I think at least in the near future (10 years or so), if one or more of these AI projects actually gain traction, they will be used as support tools to aid radiologist in their image interpretation. I can imagine it will help reduce misses, shorten interpretation time (if it can get any faster than it already is) and maybe even help generate a differential. I think days of relying on physician memorization skills for things like differential diagnosis need to come to an end (this is not to say memorization has no place in medical education, but that's another topic). Also remember, they have shown success (however limited or inconclusive) with one type of diagnosis so far. There are hundreds (thousands? I don't know) of diagnoses that radiologists are trained to provide.

I don't claim to understand how deep-learning algorithms work (if someone can explain, that would be great), but with any kind of computer algorithms that generate important decisions like medical diagnoses, there have to be humans who check the work of the computer, and who is ultimately responsible for signing off on the diagnosis. Maybe the computers can be thought of as "robot residents".

If you ask me about what it's going to be like 30 years later, I don't know. Maybe the need for human radiologist work force will be cut to 10% of what it is today. Maybe it will stay about the same. A lot of jobs in and outside of medicine that exist today will have become obsolete by 2046.
 
Thanks so much for the thoughtful responses @radman321 & @LePeintre! I feel much better about my decision. You both are right. I feel that as long as I stay up to date and focus on research and furthering the field, I will still be of value to someone 🙂
 
I think this type of AI-aided diagnosis was coming sooner or later. If I'm not mistaken, this happens in pathology to some level as we speak. I am not sure why the company did not first try to analyze EKG's, which is a 1-D data and presumably an easier problem to attack. Maybe it has to do with the availability of digitized forms of data where radiology as a field has done an excellent job.

I think at least in the near future (10 years or so), if one or more of these AI projects actually gain traction, they will be used as support tools to aid radiologist in their image interpretation. I can imagine it will help reduce misses, shorten interpretation time (if it can get any faster than it already is) and maybe even help generate a differential. I think days of relying on physician memorization skills for things like differential diagnosis need to come to an end (this is not to say memorization has no place in medical education, but that's another topic). Also remember, they have shown success (however limited or inconclusive) with one type of diagnosis so far. There are hundreds (thousands? I don't know) of diagnoses that radiologists are trained to provide.

I don't claim to understand how deep-learning algorithms work (if someone can explain, that would be great), but with any kind of computer algorithms that generate important decisions like medical diagnoses, there have to be humans who check the work of the computer, and who is ultimately responsible for signing off on the diagnosis. Maybe the computers can be thought of as "robot residents".

If you ask me about what it's going to be like 30 years later, I don't know. Maybe the need for human radiologist work force will be cut to 10% of what it is today. Maybe it will stay about the same. A lot of jobs in and outside of medicine that exist today will have become obsolete by 2046.

Your info is about 10 years old. Such AI exists in clinical radiology and is called CAD in mammo. Almost all mammo workstations have CAD. It was FDA approved 10 years or more ago and has its own billing code.
 
Your info is about 10 years old. Such AI exists in clinical radiology and is called CAD in mammo. Almost all mammo workstations have CAD. It was FDA approved 10 years or more ago and has its own billing code.

I did see CAD in mammo during my rotations but not being a practicing radiologist, my knowledge is limited. I would be interested in hearing more about your opinions on this matter, specifically how much AI has changed your practice pattern and workload in the past 10 years and where you see this trend going in the next 10 years, not just in mammo but in all subfields.
 
My institution has CAD and it is an utter joke. It picks up things that are not, it does not pick up things that are.
 
I did see CAD in mammo during my rotations but not being a practicing radiologist, my knowledge is limited. I would be interested in hearing more about your opinions on this matter, specifically how much AI has changed your practice pattern and workload in the past 10 years and where you see this trend going in the next 10 years, not just in mammo but in all subfields.

So if you are not a practicing radiologist and your knowledge is limited (per your own statement above), it is better to mention it in front of your post. Your above post was so definite and overconfident that I thought it was from an expert in NIH who has been working on this topic for the last 15 years.

One problem with SDN is that there are too many people speculating about the topics that they don't have any clue about.
 
Almost all digital mammo in US have CAD with them.
 
So if you are not a practicing radiologist and your knowledge is limited (per your own statement above), it is better to mention it in front of your post. Your above post was so definite and overconfident that I thought it was from an expert in NIH who has been working on this topic for the last 15 years.

One problem with SDN is that there are too many people speculating about the topics that they don't have any clue about.

Point well taken. Perhaps I should preface all of my posts from now on saying that "I am not a practicing radiologist".

I thought by saying "I think" and "I don't claim to understand" etc., I was making it pretty clear that my posts are one man's opinion and I never meant to say anything as a definitive authority.

I understand your frustration with people on SDN saying all kinds of things without bases, but such is the nature of anonymous forums. SDN specifically is frequented a lot more by medical students than residents or attendings as I am sure you are well aware.

Your input on this board is very much appreciated but I sometimes feel that your comments are not very constructive and even stifling. You can correct misinformed opinions rather than dismissing them with what amounts to "you don't know sxxx". If it weren't for medical students posting their baseless opinions, this forum would be very much dead and boring. I would think that well-educated adults (which med students and even college students are) would know to take information posted on anonymous forums with a grain of salt. If they don't know how to do that, bless their hearts. A forum where well-intentioned but misguided/outdated opinions are allowed, respected and corrected is more constructive than one where only expert opinions are allowed. If I wanted only authoritative information, I would not come to this board.
 
Point well taken. Perhaps I should preface all of my posts from now on saying that "I am not a practicing radiologist".

I thought by saying "I think" and "I don't claim to understand" etc., I was making it pretty clear that my posts are one man's opinion and I never meant to say anything as a definitive authority.

I understand your frustration with people on SDN saying all kinds of things without bases, but such is the nature of anonymous forums. SDN specifically is frequented a lot more by medical students than residents or attendings as I am sure you are well aware.

Your input on this board is very much appreciated but I sometimes feel that your comments are not very constructive and even stifling. You can correct misinformed opinions rather than dismissing them with what amounts to "you don't know sxxx". If it weren't for medical students posting their baseless opinions, this forum would be very much dead and boring. I would think that well-educated adults (which med students and even college students are) would know to take information posted on anonymous forums with a grain of salt. If they don't know how to do that, bless their hearts. A forum where well-intentioned but misguided/outdated opinions are allowed, respected and corrected is more constructive than one where only expert opinions are allowed. If I wanted only authoritative information, I would not come to this board.

Nope. Misinformation is much worse than no information.

You don't need to mention your position at the beginning of all your posts. But the above post was your FIRST POST on this forum. As a first time poster, I think you have to mention your position.

In my opinion this forum used to be a better place to discuss different topics. Unfortunately, recently I see a lot more speculations and nonsense posts. This problem has become more common recently with the new "share account". What do you expect me to post in response to someone who posts almost a page of BS about the workflow at one of the top academic centers (BWH on the other forum)?

When you talk or post crap about a topic, don't expect to receive compliment from people who may know at least a little more about that topic.

Anyway, I am going to post less on this forum since the number of BS posts are getting almost out of control and it seems people get offended if someone criticizes their BS.
 
CAD in mammography is will not be replacing radiologists any time soon. I have used CAD for almost two decades and still have a job. In the mean time we get to bill $3/case extra for the extra time it takes to look at the CAD findings. CAD will be most useful when you have to look at large amount of data for relatively simple findings like lung nodules. No CAD system that I am aware of can find an appendix on CT abdomen or pelvis as well first year radiology resident, let alone an experienced radiologist. Machine learning will change the way radiology is practiced, but IMO radiology will be one of the last medical specialties to me replaced by machines. Primary care midlevel providers used in conjunction with computers are more likely to replace primary care doctors.
 
Computers are going to replace primary care long before they replace Radiology. As has been proven, current processors cannot evaluate differences in basic rhythms on EKG. However, most PCPs work in a very "if, then" manner.
 
There's a newer thread where we've been discussing this topic:

http://forums.studentdoctor.net/threads/machine-learning-and-radiology.1224279/

I think it's important to note that EKG-reading algorithms that are built into the EKG machines haven't been updated in years. I agree that those are inaccurate often/most of the time. However, if deep learning were applied to EKG reads....well it'd be a solved problem. The issue is that physicians do not stand to gain anything by automating EKG readings. It takes only a few seconds to read an EKG and there's no report that has to be written. Compare this to radiology - It's much harder to automate diagnoses and note-writing for a radiology image, but I think feasible within the next 20 years.
 
There's a newer thread where we've been discussing this topic:

http://forums.studentdoctor.net/threads/machine-learning-and-radiology.1224279/

I think it's important to note that EKG-reading algorithms that are built into the EKG machines haven't been updated in years. I agree that those are inaccurate often/most of the time. However, if deep learning were applied to EKG reads....well it'd be a solved problem. The issue is that physicians do not stand to gain anything by automating EKG readings. It takes only a few seconds to read an EKG and there's no report that has to be written. Compare this to radiology - It's much harder to automate diagnoses and note-writing for a radiology image, but I think feasible within the next 20 years.

If you are saying that machines will take over most jobs in 20 years than maybe. But you are simplifying radiology way too much if you think that radiology will be automated in 20 years. In 20 years we will have great computer assistance for mundane things like lung nodules, breast lesions on mammo, and maybe somthing to help find the appendix or peritoneal mets. But honestly if you truly think radiology will be automated, then you must also beleive primary care will be extinct as will be most jobs. Talk to me after you take call at a tertiary care level 1 trauma center overnight making big calls in which someone goes to surgery/angio or not, or whether you ignore a finding or make a huge deal about it. That will give you computer scientists some real perspective on the high complexity of radiology. This is not chess or jeopardy or cook book algorithmic medicine at all.
 
Computers are going to replace primary care long before they replace Radiology. As has been proven, current processors cannot evaluate differences in basic rhythms on EKG. However, most PCPs work in a very "if, then" manner.

Sorry but this isn't true. PCPs have something that radiologists often don't have (while interpreting images). That is a patient in front of them. You think grandma wants to see R2D2? Is that smoker going to get counseling from an automated voice? Does that patient want to hear they have cancer from a machine?
 
Sorry but this isn't true. PCPs have something that radiologists often don't have (while interpreting images). That is a patient in front of them. You think grandma wants to see R2D2? Is that smoker going to get counseling from an automated voice? Does that patient want to hear they have cancer from a machine?

You're totally missing the point, which isn't unexpected because I didn't explain it well. If we're talking about computers, innovation, and cost savings, replacing PCPs is going to happen far before Radiologists. The diagnostic aspect of their job is much easier to do with an algorithm than interpreting subtleties in imaging. The extra cost and morbidity related to the additional imaging/procedures from mistakes with CAD (yeah, it already exists) would be astronomical. On the other hand, diagnostic evaluations are, quite literally, "if, then" functions. We're talking average clinicians here, not excellent clinicians. For example, if persistent cough, then CXR. If negative or chest pain and SOB, then CT. I see it all the time in the ER. It applies to numerous other decision trees, too. Just check out Step Up to Medicine for more examples.

It also doesn't matter what Grandma wants. What matters is what the government and insurance is going to pay for. Just because a patient does or doesn't want something to happen, doesn't mean it will or won't happen. Medicine is shifting to efficiency, and that's what this thread is discussing.
 
Last edited:
It also doesn't matter what Grandma wants. What matters is what the government and insurance is going to pay for. Just because a patient does or doesn't want something to happen, doesn't mean it will or won't happen. Medicine is shifting to efficiency, and that's what this thread is discussing.

Actually it does matter want Grandma wants. She and everyone else vote which influences who is elected and our policies. Patients are also consumers, not just of healthcare, but also health insurance.
 
Sorry but this isn't true. PCPs have something that radiologists often don't have (while interpreting images). That is a patient in front of them. You think grandma wants to see R2D2? Is that smoker going to get counseling from an automated voice? Does that patient want to hear they have cancer from a machine?
Grandma may not want R2D2, but she may be fine with mid level provider (using computer decision support) who has a lot of time to chat with her. You don't need MD and residency to tell someone they have cancer or get them to quit smoking, lower paid mid levels can do that just as well as MDs.
 
Actually it does matter want Grandma wants. She and everyone else vote which influences who is elected and our policies. Patients are also consumers, not just of healthcare, but also health insurance.

That's a very naive outlook. It's cute, though.
 
Grandma may not want R2D2, but she may be fine with mid level provider (using computer decision support) who has a lot of time to chat with her. You don't need MD and residency to tell someone they have cancer or get them to quit smoking, lower paid mid levels can do that just as well as MDs.

And diagnostic algorithms can help them in that path.

I'm not knocking their job, only saying that if there were a field to be eliminated by a program, it's going to be primary care far before radiology.
 
I feel like it needs to be reiterated that systems like Watson are never going to replace doctors. There's a special on 60 minutes about how Watson was able to find something that a clinician missed. The problem is Watson is a rules based system, it is not a deep learning system. It's the difference between learning on the raw pixels vs. learning on derived features that were already labeled (ribs, organs, shadows, etc). Watson requires data to be somewhat structured as in the latter case. IBM has done a wonderful job hyping (maybe even lying) about Watson being a state of the art artificial intelligence. Here's an excellent article about why this is not the case:

http://www.rogerschank.com/fraudulent-claims-made-by-IBM-about-Watson-and-AI
 
Top