MD & DO Will AI and Machine Learning (ML) Doom Your Choice of a Radiology Career Before it Starts?

[RadsGuy]

There is a lot of hype around AI and ML right now and major hospitals, universities and tech companies have already started pairing off/partnering up with big industry players like IBM and GE to name a few. There’s also no shortage of articles about how AI has outperformed a pathologist or a dermatologist on diagnosing early cancer. Or how pulmonary nodules can be detected within seconds on Chest CTs using current ML techniques. Does that mean that the radiologists’ days are numbered? Will radiologists eventually be replaced by machines?

In short, NO. Here are my opinions on why not:

1. I know this is cynical, but you can’t sue a computer. The lawyers will need someone to blame when an incidental 2mm pulmonary nodule (which has a <1% risk of cancer) turns out to be cancer 5 years later. Who will take the liability? Certainly not IBM or GE. They wouldn’t be in this game if they opened themselves up to litigation like that. Moreover, if you got a CT done and the doctor told you “computer says it’s normal,” would you say ok or would you want to make sure that a human being looked at it? Even worse, “computer says it’s cancer, we’re going to admit you and prep for surgery,” wouldn’t you want some person to look at that?

2. While these algorithms are powerful, they are also very brittle. For example, the slightest amount of noise in an image can trick Google’s AI into thinking that 4 machine guns are a helicopter. There needs to be quite a few breakthroughs if we expect an AI to reliably call pathology on a radiology exam. I think AIs will probably get very good at calling a normal study, normal. But for it to differentiate pancreatitis, a pancreatic head mass, and focal atrophy of the pancreatic body (which gives the illusion of a panc head mass) is going to be tough to do. AIs will probably just flag studies for us, do some preliminary interpretation and we’ll put the final approval on it.

3. The amount of data needed to validate the algorithms is crazy. You need maybe 100,000 of examples minimum of appendicitis to create an AI that can call it well. Half of that is used to train the algorithm and the other half is used to validate/test it. Let’s say you use a technique called “transfer learning” to offset that big uphill of cases needed, you still need maybe 10,000 examples of appendicitis. That’s actually doable. But, try finding 10,000 cases of Rhombencephalosynapsis or some other rare disease. There’s just not that many cases of rare diseases for us to create and train algorithms on. That’s why you need a radiologist to find that Zebra sitting in a herd of horses.

4. Many of these “ground-breaking” algorithms and ML concepts existed in the 70s and 80s. It’s just the computing power that is the transformative factor today, but even that will have its saturation point. You can only fit so many transistors onto a microchip. Moore’s Law (the observation that the number of transistors in a dense integrated circuit doubles about every two years) is actually slowing down and at some point, we may saturate before achieving the computing power needed for a decent radiology AI.

I know it’s coming, but it’s not coming to replace us. It’s coming to transform our space and ultimately help us. AI is good at repetitive tasks in a very narrow range. If it can remove some of the mundane in my every day, I’m all for it.

---

Members don't see this ad.

 

[Lawpy]

@CherryRedDracul might agree with this thread

 

[CherryRedDracul]

@CherryRedDracul might agree with this thread

You know me like I know the back of my own hand.

 

[Matthew9Thirtyfive]

You know me like I know the back of my own hand.

You sit in the dark all day, so probably not very well.

 

[Rekt]

Still plenty of hospitals barely making the switch from paper to EMR. To think they'd ever switch anytime soon to AI is absurd, regardless of any other the other arguments surrounding it.

 

[libertyyne]

Still plenty of hospitals barely making the switch from paper to EMR. To think they'd ever switch anytime soon to AI is absurd, regardless of any other the other arguments surrounding it.

there isnt a whole lot of money to be saved by switching to EMRs. If hospitals could cut their cost of radiology services by 50% tomorrow by switching to AI , you would see them act fairly quickly.

 

[ACSurgeon]

There is a lot of hype around AI and ML right now and major hospitals, universities and tech companies have already started pairing off/partnering up with big industry players like IBM and GE to name a few. There’s also no shortage of articles about how AI has outperformed a pathologist or a dermatologist on diagnosing early cancer. Or how pulmonary nodules can be detected within seconds on Chest CTs using current ML techniques. Does that mean that the radiologists’ days are numbered? Will radiologists eventually be replaced by machines?

In short, NO. Here are my opinions on why not:

1. I know this is cynical, but you can’t sue a computer. The lawyers will need someone to blame when an incidental 2mm pulmonary nodule (which has a <1% risk of cancer) turns out to be cancer 5 years later. Who will take the liability? Certainly not IBM or GE. They wouldn’t be in this game if they opened themselves up to litigation like that. Moreover, if you got a CT done and the doctor told you “computer says it’s normal,” would you say ok or would you want to make sure that a human being looked at it? Even worse, “computer says it’s cancer, we’re going to admit you and prep for surgery,” wouldn’t you want some person to look at that?

2. While these algorithms are powerful, they are also very brittle. For example, the slightest amount of noise in an image can trick Google’s AI into thinking that 4 machine guns are a helicopter. There needs to be quite a few breakthroughs if we expect an AI to reliably call pathology on a radiology exam. I think AIs will probably get very good at calling a normal study, normal. But for it to differentiate pancreatitis, a pancreatic head mass, and focal atrophy of the pancreatic body (which gives the illusion of a panc head mass) is going to be tough to do. AIs will probably just flag studies for us, do some preliminary interpretation and we’ll put the final approval on it.

3. The amount of data needed to validate the algorithms is crazy. You need maybe 100,000 of examples minimum of appendicitis to create an AI that can call it well. Half of that is used to train the algorithm and the other half is used to validate/test it. Let’s say you use a technique called “transfer learning” to offset that big uphill of cases needed, you still need maybe 10,000 examples of appendicitis. That’s actually doable. But, try finding 10,000 cases of Rhombencephalosynapsis or some other rare disease. There’s just not that many cases of rare diseases for us to create and train algorithms on. That’s why you need a radiologist to find that Zebra sitting in a herd of horses.

4. Many of these “ground-breaking” algorithms and ML concepts existed in the 70s and 80s. It’s just the computing power that is the transformative factor today, but even that will have its saturation point. You can only fit so many transistors onto a microchip. Moore’s Law (the observation that the number of transistors in a dense integrated circuit doubles about every two years) is actually slowing down and at some point, we may saturate before achieving the computing power needed for a decent radiology AI.

I know it’s coming, but it’s not coming to replace us. It’s coming to transform our space and ultimately help us. AI is good at repetitive tasks in a very narrow range. If it can remove some of the mundane in my every day, I’m all for it.

They’ll just program the machine to have a broad differential... how’s that any different then what you do?

Kidding of course. Lots of respect to radiology colleagues and I go talk to a radiologist in person very frequently.

 

[deleted938090]

Bookmarking this post to revisit in 30 years.

 

[deleted103644]

There is a lot of hype around AI and ML right now and major hospitals, universities and tech companies have already started pairing off/partnering up with big industry players like IBM and GE to name a few. There’s also no shortage of articles about how AI has outperformed a pathologist or a dermatologist on diagnosing early cancer. Or how pulmonary nodules can be detected within seconds on Chest CTs using current ML techniques. Does that mean that the radiologists’ days are numbered? Will radiologists eventually be replaced by machines?

In short, NO. Here are my opinions on why not:

1. I know this is cynical, but you can’t sue a computer. The lawyers will need someone to blame when an incidental 2mm pulmonary nodule (which has a <1% risk of cancer) turns out to be cancer 5 years later. Who will take the liability? Certainly not IBM or GE. They wouldn’t be in this game if they opened themselves up to litigation like that. Moreover, if you got a CT done and the doctor told you “computer says it’s normal,” would you say ok or would you want to make sure that a human being looked at it? Even worse, “computer says it’s cancer, we’re going to admit you and prep for surgery,” wouldn’t you want some person to look at that?

2. While these algorithms are powerful, they are also very brittle. For example, the slightest amount of noise in an image can trick Google’s AI into thinking that 4 machine guns are a helicopter. There needs to be quite a few breakthroughs if we expect an AI to reliably call pathology on a radiology exam. I think AIs will probably get very good at calling a normal study, normal. But for it to differentiate pancreatitis, a pancreatic head mass, and focal atrophy of the pancreatic body (which gives the illusion of a panc head mass) is going to be tough to do. AIs will probably just flag studies for us, do some preliminary interpretation and we’ll put the final approval on it.

3. The amount of data needed to validate the algorithms is crazy. You need maybe 100,000 of examples minimum of appendicitis to create an AI that can call it well. Half of that is used to train the algorithm and the other half is used to validate/test it. Let’s say you use a technique called “transfer learning” to offset that big uphill of cases needed, you still need maybe 10,000 examples of appendicitis. That’s actually doable. But, try finding 10,000 cases of Rhombencephalosynapsis or some other rare disease. There’s just not that many cases of rare diseases for us to create and train algorithms on. That’s why you need a radiologist to find that Zebra sitting in a herd of horses.

4. Many of these “ground-breaking” algorithms and ML concepts existed in the 70s and 80s. It’s just the computing power that is the transformative factor today, but even that will have its saturation point. You can only fit so many transistors onto a microchip. Moore’s Law (the observation that the number of transistors in a dense integrated circuit doubles about every two years) is actually slowing down and at some point, we may saturate before achieving the computing power needed for a decent radiology AI.

I know it’s coming, but it’s not coming to replace us. It’s coming to transform our space and ultimately help us. AI is good at repetitive tasks in a very narrow range. If it can remove some of the mundane in my every day, I’m all for it.

 

[libertyyne]

Bookmarking this post to revisit in 30 years.

Im gonna check back in 10.

 

[Evo14]

It's really hard to give an unbiased answer on this thread. I'd bet almost anyone who work w/ machine learning and A.I would give totally different answers than what's on this thread. So really, only time will tell what happens. My bet is that A.I will be used in conjunction w/ physicians. I'm not sure if that would decrease the demand of physicians or not.

 

[nimbus]

there isnt a whole lot of money to be saved by switching to EMRs. If hospitals could cut their cost of radiology services by 50% tomorrow by switching to AI , you would see them act fairly quickly.

The promise of EMR for hospitals is that it’s supposed to capture every single charge. It’s purely economic, that’s why a hospital system will spends hundreds of millions of dollars for an EMR.

Hospitals don’t pay for radiology services. For the most part, radiologists get paid by insurance companies.

 

[aprilfools]

It's really hard to give an unbiased answer on this thread. I'd bet almost anyone who work w/ machine learning and A.I would give totally different answers than what's on this thread. So really, only time will tell what happens. My bet is that A.I will be used in conjunction w/ physicians. I'm not sure if that would decrease the demand of physicians or not.

I worked in machine learning/AI during my gap year before medical school (ML/AI in a healthcare setting at that) and I couldn't be more in agreeance with OP.

As many people in ML/AI who AREN'T in healthcare are finding out, "Nobody knew healthcare could be so complicated"

 

[PapaGuava]

It's really hard to give an unbiased answer on this thread. I'd bet almost anyone who work w/ machine learning and A.I would give totally different answers than what's on this thread. So really, only time will tell what happens. My bet is that A.I will be used in conjunction w/ physicians. I'm not sure if that would decrease the demand of physicians or not.

Until those A.I working in conjunction with physicians become the administrators, then fire human staff

 

[NotAProgDirector]

AIs will probably just flag studies for us, do some preliminary interpretation and we’ll put the final approval on it.

This is exactly what is likely to happen, IMHO. The effect will be less need for radiologists, since each radiologist will be able to review more studies in a given time (since most of the work is done by the computer). We see the same tech in the lab -- used to be that a lab tech would need to look at a blood smear and count cells to get a diff. Now, the machine does all the counting, then shows the cells on a screen that it can't tell what they are, and have a tech label them. Anything that is wildly outside norms (i.e. circulating blasts) gets kicked out for a manual review. Turnaround time is now much faster, with less people needed.

So I don't think radiologists will be replaced. But I do worry that there will be less jobs.

 

[libertyyne]

The promise of EMR for hospitals is that it’s supposed to capture every single charge. It’s purely economic, that’s why a hospital system will spends hundreds of millions of dollars for an EMR.

Hospitals don’t pay for radiology services. For the most part, radiologists get paid by insurance companies.

The only reason hospitals switched to EMRs was the medicare billing requirement. Medicare would have stopped reimbursing hospitals which would have caused them to shut down.
If the rads group is owned by the hospital , its ths hospitals collecting the charges.
If the hospital contracts with the rads group , and pays them a cut of collections , it is the hospitals bottom line.
If the rad group is privately owned and now suddenly there is a chance to implement AI and reduce staff costs, they are going to do it.

 

[nimbus]

The only reason hospitals switched to EMRs was the medicare billing requirement. Medicare would have stopped reimbursing hospitals which would have caused them to shut down.
If the rads group is owned by the hospital , its ths hospitals collecting the charges.
If the hospital contracts with the rads group , and pays them a cut of collections , it is the hospitals bottom line.
If the rad group is privately owned and now suddenly there is a chance to implement AI and reduce staff costs, they are going to do it.

Every radiologist I know work in single specialty multi site radiology groups or as part of a multi specialty group and bill their own patients. The hospital is completely out of the picture.

 

[libertyyne]

Every radiologist I know work in single specialty multi site radiology groups or as part of a multi specialty group and bill their own patients. The hospital is completely out of the picture.

two of the 5 that I know of are joint ventures with hospitals with hospital ownership.

 

[nimbus]

two of the 5 that I know of are joint ventures with hospitals with hospital ownership.

Interesting. Perhaps regional differences. I know a couple dozen radiologists and not a single one is hospital employed.

 

[orainformatics]

I have been a radiologist for 8 years and been following the ML trend pretty closely to watch the 130 worldwide ML vendors grow. The quote is: radiologists won't be replaced by ML, but radiologists who fail to use ML may be replaced by those who do. I have a 10 part series on LinkedIn, article 10 to post tomorrow - pretty easy to find me on linkedin: tyvachon

 

[NicMouse64]

There is a lot of hype around AI and ML right now and major hospitals, universities and tech companies have already started pairing off/partnering up with big industry players like IBM and GE to name a few. There’s also no shortage of articles about how AI has outperformed a pathologist or a dermatologist on diagnosing early cancer. Or how pulmonary nodules can be detected within seconds on Chest CTs using current ML techniques. Does that mean that the radiologists’ days are numbered? Will radiologists eventually be replaced by machines?

In short, NO. Here are my opinions on why not:

1. I know this is cynical, but you can’t sue a computer. The lawyers will need someone to blame when an incidental 2mm pulmonary nodule (which has a <1% risk of cancer) turns out to be cancer 5 years later. Who will take the liability? Certainly not IBM or GE. They wouldn’t be in this game if they opened themselves up to litigation like that. Moreover, if you got a CT done and the doctor told you “computer says it’s normal,” would you say ok or would you want to make sure that a human being looked at it? Even worse, “computer says it’s cancer, we’re going to admit you and prep for surgery,” wouldn’t you want some person to look at that?

2. While these algorithms are powerful, they are also very brittle. For example, the slightest amount of noise in an image can trick Google’s AI into thinking that 4 machine guns are a helicopter. There needs to be quite a few breakthroughs if we expect an AI to reliably call pathology on a radiology exam. I think AIs will probably get very good at calling a normal study, normal. But for it to differentiate pancreatitis, a pancreatic head mass, and focal atrophy of the pancreatic body (which gives the illusion of a panc head mass) is going to be tough to do. AIs will probably just flag studies for us, do some preliminary interpretation and we’ll put the final approval on it.

3. The amount of data needed to validate the algorithms is crazy. You need maybe 100,000 of examples minimum of appendicitis to create an AI that can call it well. Half of that is used to train the algorithm and the other half is used to validate/test it. Let’s say you use a technique called “transfer learning” to offset that big uphill of cases needed, you still need maybe 10,000 examples of appendicitis. That’s actually doable. But, try finding 10,000 cases of Rhombencephalosynapsis or some other rare disease. There’s just not that many cases of rare diseases for us to create and train algorithms on. That’s why you need a radiologist to find that Zebra sitting in a herd of horses.

4. Many of these “ground-breaking” algorithms and ML concepts existed in the 70s and 80s. It’s just the computing power that is the transformative factor today, but even that will have its saturation point. You can only fit so many transistors onto a microchip. Moore’s Law (the observation that the number of transistors in a dense integrated circuit doubles about every two years) is actually slowing down and at some point, we may saturate before achieving the computing power needed for a decent radiology AI.

I know it’s coming, but it’s not coming to replace us. It’s coming to transform our space and ultimately help us. AI is good at repetitive tasks in a very narrow range. If it can remove some of the mundane in my every day, I’m all for it.

On point 4, while you are correct, I've been reading about quantum computing in Scientific American. There is a near infinite amount of computing that could fit onto a quantum computer as the binary system would become obsolete.

 

[Treebeard]

We won't get to ZERO radiologists any time soon, but the demand is definitely going to plummet and the job market will suffer. It probably only makes to go rads if you are competitive for top tier institutions that will be driving this research, or if you want to do IR.

 

[MD-2022]

I am an expert in the field with relatives in biotech. Radiology will be an extinct field in 10 years. Already, in China, AI and biotech have already automated many of the tasks of radiologists. Include machine learning and Watson, the field will be dead soon. Healthcare will always invest in new tech that can save them money in the long run. Will patients trust an AI with error rate of .001% of a human. This will be decided by society. However, interventional radiology still is safe. Gluck.

 

[deleted938090]

Who has the liability?

 

[DrWhozits]

Aren’t each of these ML algorithms created to analyze a specific case? I.e. this one does lung nodules, that one does pneumonia.

If this is the case, the FDA would theoretically have to approve each individual algorithm as they are developed. For every. Single. Pathology.

It seems to me that it would take time to not only develop and perfect these, it will also take forever just to approve them. In this scenario radiologists will have plenty of time to bail and use their medical degrees for something else.

 

[CherryRedDracul]

I am an expert in the field with relatives in biotech. Radiology will be an extinct field in 10 years. Already, in China, AI and biotech have already automated many of the tasks of radiologists. Include machine learning and Watson, the field will be dead soon. Healthcare will always invest in new tech that can save them money in the long run. Will patients trust an AI with error rate of .001% of a human. This will be decided by society. However, interventional radiology still is safe. Gluck.

o rly

As someone who isn't intimately familiar with radiology in China (except that they've put out a ton of papers on contrast-enhanced ultrasound), what tasks have they automated so far? And who will be making the diagnoses for zebra conditions for which one would expectedly have tiny training sets?

 

[Angus Avagadro]

There is a lot of hype around AI and ML right now and major hospitals, universities and tech companies have already started pairing off/partnering up with big industry players like IBM and GE to name a few. There’s also no shortage of articles about how AI has outperformed a pathologist or a dermatologist on diagnosing early cancer. Or how pulmonary nodules can be detected within seconds on Chest CTs using current ML techniques. Does that mean that the radiologists’ days are numbered? Will radiologists eventually be replaced by machines?

In short, NO. Here are my opinions on why not:

1. I know this is cynical, but you can’t sue a computer. The lawyers will need someone to blame when an incidental 2mm pulmonary nodule (which has a <1% risk of cancer) turns out to be cancer 5 years later. Who will take the liability? Certainly not IBM or GE. They wouldn’t be in this game if they opened themselves up to litigation like that. Moreover, if you got a CT done and the doctor told you “computer says it’s normal,” would you say ok or would you want to make sure that a human being looked at it? Even worse, “computer says it’s cancer, we’re going to admit you and prep for surgery,” wouldn’t you want some person to look at that?

2. While these algorithms are powerful, they are also very brittle. For example, the slightest amount of noise in an image can trick Google’s AI into thinking that 4 machine guns are a helicopter. There needs to be quite a few breakthroughs if we expect an AI to reliably call pathology on a radiology exam. I think AIs will probably get very good at calling a normal study, normal. But for it to differentiate pancreatitis, a pancreatic head mass, and focal atrophy of the pancreatic body (which gives the illusion of a panc head mass) is going to be tough to do. AIs will probably just flag studies for us, do some preliminary interpretation and we’ll put the final approval on it.

3. The amount of data needed to validate the algorithms is crazy. You need maybe 100,000 of examples minimum of appendicitis to create an AI that can call it well. Half of that is used to train the algorithm and the other half is used to validate/test it. Let’s say you use a technique called “transfer learning” to offset that big uphill of cases needed, you still need maybe 10,000 examples of appendicitis. That’s actually doable. But, try finding 10,000 cases of Rhombencephalosynapsis or some other rare disease. There’s just not that many cases of rare diseases for us to create and train algorithms on. That’s why you need a radiologist to find that Zebra sitting in a herd of horses.

4. Many of these “ground-breaking” algorithms and ML concepts existed in the 70s and 80s. It’s just the computing power that is the transformative factor today, but even that will have its saturation point. You can only fit so many transistors onto a microchip. Moore’s Law (the observation that the number of transistors in a dense integrated circuit doubles about every two years) is actually slowing down and at some point, we may saturate before achieving the computing power needed for a decent radiology AI.

I know it’s coming, but it’s not coming to replace us. It’s coming to transform our space and ultimately help us. AI is good at repetitive tasks in a very narrow range. If it can remove some of the mundane in my every day, I’m all for it.

With respect to Moores law, I read that Intel recently announced it had about 4 yrs until it could no longer double the processing speed of the chip
I have commented on this before. A radiology career is 30 to 40 years. 30 years ago, cell phones were the size of laptops, and they were only phones. Now, smart phones are computers, Skypers,connected to the web, texters, emailers, movie watchers, users of voice recognition, cameras, video recorders..and phones. Imagine what technology will be like in 30 years. Replace radiologists? No, reduce the number needed to diagnose zebras? I'm certain, as there are only so many zebras amongst the horses to diagnose. I believe AI will be reading all the Normals, reducing the workload, reimbursement and the need for radiologists in 10 to 20 yrs. One man's opinion, wife is radiologist and disagrees. And yes, she IS the smart one

 

[RadsGuy]

I see your points, but anyone want to tackle who gets sued if a robot misses something?

 

[Lawpy]

And

 

[Angus Avagadro]

I see your points, but anyone want to tackle who gets sued if a robot misses something?

Hospital or imaging center who owns the machine and software I would imagine.

 

[deleted103644]

I am an expert in the field with relatives in biotech. Radiology will be an extinct field in 10 years. Already, in China, AI and biotech have already automated many of the tasks of radiologists. Include machine learning and Watson, the field will be dead soon. Healthcare will always invest in new tech that can save them money in the long run. Will patients trust an AI with error rate of .001% of a human. This will be decided by society. However, interventional radiology still is safe. Gluck.

You made a reference to Watson, so you are clearly not an expert in the field.

https://gizmodo.com/why-everyone-is-hating-on-watson-including-the-people-w-1797510888

Watson won jeopardy and has a great (or just aggressive) PR department. It has nothing on Google, Facebook, Amazon, or even Uber.

AI will come for us all eventually, but all things take time. Radiology is far from the only field in jeopardy.

 

[MD-2022]

You made a reference to Watson, so you are clearly not an expert in the field.

https://gizmodo.com/why-everyone-is-hating-on-watson-including-the-people-w-1797510888

Watson won jeopardy and has a great (or just aggressive) PR department. It has nothing on Google, Facebook, Amazon, or even Uber.

AI will come for us all eventually, but all things take time. Radiology is far from the only field in jeopardy.

Please do your due diligence. I said 10 years. This is Watson's current process.

IBM Watson Health - Cognitive Healthcare Solutions

Playing Doctor with Watson: Medical Applications Expose Current Limits of AI - SPIEGEL ONLINE - International

 

[deleted103644]

Please do your due diligence. I said 10 years. This is Watson's current process.

IBM Watson Health - Cognitive Healthcare Solutions

Playing Doctor with Watson: Medical Applications Expose Current Limits of AI - SPIEGEL ONLINE - International

Please do your due diligence.

Watson is not cutting edge AI, and referencing it reveals your ignorance.

 

[Kuratz]

I am an expert in the field with relatives in biotech. Radiology will be an extinct field in 10 years. Already, in China, AI and biotech have already automated many of the tasks of radiologists. Include machine learning and Watson, the field will be dead soon. Healthcare will always invest in new tech that can save them money in the long run. Will patients trust an AI with error rate of .001% of a human. This will be decided by society. However, interventional radiology still is safe. Gluck.

Quoting for posterity. Claiming that radiology will be an extinct field in 10 years is completely ridiculous because even if we had perfectly optimal AI today it wouldn't be implemented in it's entirety by 2028 for reasons totally unrelated to the technology. If such an AI were to exist then Pathology, Derm, Retina, IM, FM etc etc would all be facing a similar fate, albeit a more insidious one with merging of AI and midlevels. The technology just isn't there yet and the wisest leaders in the field don't make such certain conjectures about the future as you have. (Also lol for citing freaking Waston). I know AI is always the hottest topic but you "experts" have been hailing the end of radiology since my father was in radiology residency.

 

[Black Coffee 24/7]

I see Diagnostic Radiology job market to be the same as Pathology in 10 years. Radiologists won’t be replaced but the efficiency of a radiologist will be 10x of current level, which decreases the demand for # of radiologists.

 

[nimbus]

26 years ago when I went into anesthesia they told me it would be a dead field because of emerging CRNAs. In fact the year after I finished training, most anesthesia programs went unfilled. A few years ago we saw sedasys come and go. In the mean time I’ve enjoyed a continually evolving, engaging, rewarding career.

Give me a radiologist or pathologist who’s read thousands upon thousands of studies, who can explain their line of reasoning any time. Someone I can talk to and someone who’s wisdom, judgement and experience I trust (and I don’t trust every radiologist). A real doctor.

 

[nimbus]

Please do your due diligence. I said 10 years. This is Watson's current process.

IBM Watson Health - Cognitive Healthcare Solutions

Playing Doctor with Watson: Medical Applications Expose Current Limits of AI - SPIEGEL ONLINE - International

Did you even read these?

 

[CherryRedDracul]

Did you even read these?

Of course not, just like he's ignoring my post.

Ignore the troll.

 

[sovereign0]

1. I know this is cynical, but you can’t sue a computer. The lawyers will need someone to blame when an incidental 2mm pulmonary nodule (which has a <1% risk of cancer) turns out to be cancer 5 years later. Who will take the liability? Certainly not IBM or GE. They wouldn’t be in this game if they opened themselves up to litigation like that. Moreover, if you got a CT done and the doctor told you “computer says it’s normal,” would you say ok or would you want to make sure that a human being looked at it? Even worse, “computer says it’s cancer, we’re going to admit you and prep for surgery,” wouldn’t you want some person to look at that?

Sure, that's a valid point. But ultimately a lot of people with this concern are medical students who are wary of entering the field. Why would a medical student want to go into a field that exists partially because the legal system needs a punching bag? Three lefts make a right - while this point does support that Rads won't completely be eliminated, it still pushes people away from the field.

A similar comparison could be drawn regarding teleradiology (personally I think of it to be less of an issue). Do you really want to go into a career where part of the reason you have a job is because patients need someone to sue when things screw up?

2. While these algorithms are powerful, they are also very brittle. For example, the slightest amount of noise in an image can trick Google’s AI into thinking that 4 machine guns are a helicopter. There needs to be quite a few breakthroughs if we expect an AI to reliably call pathology on a radiology exam. I think AIs will probably get very good at calling a normal study, normal. But for it to differentiate pancreatitis, a pancreatic head mass, and focal atrophy of the pancreatic body (which gives the illusion of a panc head mass) is going to be tough to do. AIs will probably just flag studies for us, do some preliminary interpretation and we’ll put the final approval on it.

Well, I think that both radiologists and AI can be expected to be pretty good at recognizing normal and grossly abnormal. The real art is in that gray area. And even when it comes to that, I'm not sure the data is too convincing. It's a common joke at my institution that "if you show the same study to 10 different radiologists, you'll get 10 different answers". Hyperbole certainly, but there is some literature suggesting that interobserver reliability for many radiology studies isn't that great.

I'm just a med student so I'm sure I'm behind the times and the attendings and residents here will cut me apart, but here's what I was able to find:
Contrast-enhanced ultrasound in the diagnosis of pediatric focal nodular hyperplasia and hepatic adenoma: interobserver reliability. - PubMed - NCBI
Inter- and intraobserver reliability for angiographic leptomeningeal collateral flow assessment by the American Society of Interventional and Thera... - PubMed - NCBI
Inter- and intra-observer reliability of contrast-enhanced magnetic resonance imaging parameters in children with suspected juvenile idiopathic art... - PubMed - NCBI
Interobserver reliability and intraobserver reproducibility of three radiological classification systems for intra-articular calcaneal fractures. - PubMed - NCBI
Interobserver reliability of radiologists' interpretations of mobile chest radiographs for nursing home-acquired pneumonia. - PubMed - NCBI
Interobserver agreement on MRI evaluation of patients with cervical radiculopathy. - PubMed - NCBI

My point with this is that we often consider the radiology report as "the truth" when sometimes we forget that it's really an interpretation of shades of gray.

Even if we accept your scenario that AI or machine learning will just flag studies and do preliminary interpretation, that removes a huge workload burden for radiologists, which means hospitals will need fewer of them.

3. The amount of data needed to validate the algorithms is crazy. You need maybe 100,000 of examples minimum of appendicitis to create an AI that can call it well. Half of that is used to train the algorithm and the other half is used to validate/test it. Let’s say you use a technique called “transfer learning” to offset that big uphill of cases needed, you still need maybe 10,000 examples of appendicitis. That’s actually doable. But, try finding 10,000 cases of Rhombencephalosynapsis or some other rare disease. There’s just not that many cases of rare diseases for us to create and train algorithms on. That’s why you need a radiologist to find that Zebra sitting in a herd of horses.

Most medical students are looking at 20-40 year long careers (if not more, considering how student debt is evolving). Even if AI hasn't seen enough cases of rare diseases, it will certainly see enough cases of the bread and butter. How many unremarkable AP chests, CT non-con head, plain lumbar or cervical films, plain abdominal films, or extremity films do you think can be run through an algorithm in a year? What about focal lobar consolidations, post-intubation films, cholelithiasis, pneumothorax, post line placement, or post tube placement films? I won't presume to know what your actual day is like, but that's got to encompass like 90% of studies that a private radiologist at a non-academic institution reads every day.

Sure, even if AI doesn't get a bead on the rare diseases (which I think is actually doable, but I don't have solid evidence to support that), AI will definitely get the bead on the bread and butter. When you consider how much of an average day is spent on bread and butter, how many radiologists will the industry keep around to catch rhombencephalosynapsis (I don't even know what that is) at hospitals outside of academic centers?

4. Many of these “ground-breaking” algorithms and ML concepts existed in the 70s and 80s. It’s just the computing power that is the transformative factor today, but even that will have its saturation point. You can only fit so many transistors onto a microchip. Moore’s Law (the observation that the number of transistors in a dense integrated circuit doubles about every two years) is actually slowing down and at some point, we may saturate before achieving the computing power needed for a decent radiology AI.

I'll admit I'm not super well-versed in computer science, but I do think the societal emphasis we place on computer technology has skyrocketed since the 70s and 80s, when these concepts were being drawn up by undergrads in their basements between classes. Those undergrads are now sitting on their silicon towers out west, leading some of the largest companies in the history of our economy. Not to mention the emphasis in healthcare on simultaneously cutting costs and making as much money as possible, both of which AI would facilitate.

You definitely raise a good point about technology reaching a saturating point, but there's a lot of maybe's and possibly's in there about when that will actually happen. AI in radiology is one of the optimal places where I'm sure the industry is going to find that limit. I'm not willing to stake my career on that kind of gamble.

I know it’s coming, but it’s not coming to replace us. It’s coming to transform our space and ultimately help us. AI is good at repetitive tasks in a very narrow range. If it can remove some of the mundane in my every day, I’m all for it.

A lot of radiology is composed of repetitive tasks in a narrow range, with significant subtleties on occasion. The question is, are those subtleties frequent and significant enough to single-handedly define and sustain a field of medicine? I agree, it's never going to replace a skilled radiologist and erase the field, but in my opinion it's definitely going to constrict the field in a way that we haven't yet seen happen in medicine.

Drawing a comparison to hospitalist medicine - if all of the CHF/COPD exacerbations, sepsis, cellutilitis, acute coronary syndromes, pulmonary emboli and social issues were able to be managed by someone else (be it midlevels or computers), how much space would there really be left for hospitalists? It's hard to convince a fifth of medical students to go into a field saying "yeah, all that stuff is gone, but there's always the occasional cytokine storm, tropical infectious disease, or other rare condition!". There just won't be enough room left in the field.

 

[deleted938090]

The issue with evaluating Watson in the context of Jeopardy is the issue of buzzing in. When you're on the Jeopardy podium, there will be lights that turn on when the question is over: if you buzz earlier than its over, you get locked out, and if you buzz late, someone will buzz before you. (This is from my own experience during auditions). A computer that can time itself to buzz in precisely will be able to answer more questions than the other contestants. Notice how many times Brad and Ken were trying to buzz in but were outbuzzed by Watson. Also note how the critical final Jeopardy question was answered correctly by both Ken and Brad, while Watson didn't. If all three contestants were able to answer questions without being blocked by the buzzer, then the actual score could be very different.

 

[nimbus]

Judging by history and reading these posts, the likelihood of a shortage of radiologists seems equal to or maybe even greater than the likelihood of an oversupply. But maybe I’m just clueless.

 

[qwerty89]

With respect to Moores law, I read that Intel recently announced it had about 4 yrs until it could no longer double the processing speed of the chip
I have commented on this before. A radiology career is 30 to 40 years. 30 years ago, cell phones were the size of laptops, and they were only phones. Now, smart phones are computers, Skypers,connected to the web, texters, emailers, movie watchers, users of voice recognition, cameras, video recorders..and phones. Imagine what technology will be like in 30 years. Replace radiologists? No, reduce the number needed to diagnose zebras? I'm certain, as there are only so many zebras amongst the horses to diagnose. I believe AI will be reading all the Normals, reducing the workload, reimbursement and the need for radiologists in 10 to 20 yrs. One man's opinion, wife is radiologist and disagrees. And yes, she IS the smart one

Every time I hear something like this I always remember this Peter Thiel interview.



Anything that's highly regulated, such as healthcare, moves at a slow pace. Airplanes? Your typical Boeing 737 was orginally released in the late 60s and is still being used (yes I know there have been incremental upgrades). Nuclear engineering? Forget it. Biotech? Takes hundreds of millions/billions to get through the FDA.

 

[Black Coffee 24/7]

Every time I hear something like this I always remember this Peter Thiel interview.



Anything that's highly regulated, such as healthcare, moves at a slow pace. Airplanes? Your typical Boeing 737 was orginally released in the late 60s and is still being used (yes I know there have been incremental upgrades). Nuclear engineering? Forget it. Biotech? Takes hundreds of millions/billions to get through the FDA.

The reason why the mentioned technologies take forever for mass adoption is due to the low price of the traditional methods.

The problem for at least Dx Radiology is that AI will lower cost of radiology service in the hospital by making current radiologists more efficient at reading bread and butter cases. In contrast, there are actual incentives as in cheaper costs to push for AI in radiology, especially a field that had zero human touch.

I personally believe that the mass adoption of AI in radiology will happen in 10 years. As someone who won’t be an attending for another 6-8 years, there’s no way in hell am I going into Dx Rad with that belief.

 

[Matthew9Thirtyfive]

The reason why the mentioned technologies take forever for mass adoption is due to the low price of the traditional methods.

The problem for at least Dx Radiology is that AI will lower cost of radiology service in the hospital by making current radiologists more efficient at reading bread and butter cases. In contrast, there are actual incentives as in cheaper costs to push for AI in radiology, especially a field that had zero human touch.

I personally believe that the mass adoption of AI in radiology will happen in 10 years. As someone who won’t be an attending for another 6-8 years, there’s no way in hell am I going into Dx Rad with that belief.

Please bookmark this thread and come back in 10 years.

 

[hamstergang]

There is a near infinite amount of computing that could fit onto a quantum computer

That's quite hyperbolic. Quantum computing can certainly be an improvement, but it's nowhere near infinitely powerful.

 

[badasshairday]

Y'all should book mark this thread. 10 years from now you will be wishing you did rads. Brb back to the darkroom to drink coffee, listen to music, and look at pictures. Brb back to the angio suite to save another life. Brb radiology rules.

@meister

 

[RNthenDoc]

I have a question that stems from ignorance:

I see a LOT of discussion about AI replacing rads (not just posts here but in articles featured in other places), but it has always felt to me that there are other specialties that would be replaced first, yet I never see them mentioned.

In my mind it seems that the specialties with the least procedures and most algorithmic protocols would be the first to utilize a piece of software. My academic center already uses a very, very regimented flowchart for things like Abx selection (and you cannot deviate from the flowchart without SIGNIFICANT annoyances), so it seemed to me like things of that nature could be farmed out to a computer.

What am I missing?

 

[shouldvestudied]

2. While these algorithms are powerful, they are also very brittle. For example, the slightest amount of noise in an image can trick Google’s AI into thinking that 4 machine guns are a helicopter.

The claims of AI taking over Radiology have been greatly exaggerated, but you have a fundamental misunderstanding of what those researchers did if you think that the typical noise present in an imaging study is going to trick an AI into thinking someone has a machine gun in their abdomen.

 

[managedcarefin]

Still plenty of hospitals barely making the switch from paper to EMR. To think they'd ever switch anytime soon to AI is absurd, regardless of any other the other arguments surrounding it.

This is true...Hospitals spend millions each year on consulting firms to help them fix basic operations and many of those firms aren't advising on using MI in a major way yet...It will be a while before hospitals and healthcare as industry move to such an innovative tech. Leaders of these major organizations that can afford to invest in this would need to see a high return on their investment before they'd make such a dramatic switch.

 

[hamstergang]

I see a LOT of discussion about AI replacing rads (not just posts here but in articles featured in other places), but it has always felt to me that there are other specialties that would be replaced first, yet I never see them mentioned.

In my mind it seems that the specialties with the least procedures and most algorithmic protocols would be the first to utilize a piece of software. My academic center already uses a very, very regimented flowchart for things like Abx selection (and you cannot deviate from the flowchart without SIGNIFICANT annoyances), so it seemed to me like things of that nature could be farmed out to a computer.

I think the issue is one of data gathering and input. With a patient presenting with a possible infection, who's going to ask the questions to determine the right antibiotics or tests to order (also not forgetting that this patient might not have an infection, I'm not sure in what setting you imagine such an already differentiated patient arrives)? If you've interviewed patients then you know there's a lot of noise to shift through. And who's going to do the physical exam?

With radiology, we're just focusing on the image which is already in the computer anyway.