Is AI a legit concern for current med students interested in rads?

[RadsvsAI]

This may sound naive but AI is literally the only thing keeping me from pursuing rads right now. I’m currently an MS3 with a strong interest in rads but extremely fearful of the future job outlook. I want to have a long, stable career (~30 years) but scared if I choose rads then AI will eventually take over and leave me jobless 10, 15, or 20 years from now. Obviously no one can predict the future and it’s impossible to say when/if AI will have a significant impact on rads but it’s still something I worry about right now. This is especially concerning because I haven’t even started residency yet.

Let’s say I do choose rads, it’s still going to be another ~2 years before I even start residency, which means it’s going to be about 8 years until I’m finished with rads residency + fellowship. Hypothetically speaking, let’s say AI is currently 20 years out from taking over rads, then I’d only have roughly a 10-12 year career as an attending. This scares me. I really want to choose rads but my concern is that I’ll be constantly stressed throughout residency and even after because of the fear of uncertainty regarding the potential AI takeover.

Is AI a legitimate concern that current med students should take into account when deciding to pursue rads? What evidence is there that shows it is or is not a concern for someone currently in med school who wants to have a 30+ year career in rads?

---

Members don't see this ad.

 

[GoPelicans]

If you actually do radiology for more than 5 minutes you realize AI won't replace radiologists' job in our lifetimes.

 

[Radology]

Is AI a legitimate concern that current med students should take into account when deciding to pursue rads?

No

What evidence is there that shows it is or is not a concern for someone currently in med school who wants to have a 30+ year career in rads?

That is a logical fallacy, asking me to prove non existence.

For instance:

“What evidence is there that shows you did or did not make love to an alpaca in 2017?”

 

[JoshSt]

Finalized AI products are going to be very expensive and probably only mildly embraced by radiology departments. I predict the majority of products with a $20k - $100k annual fee will have great opposition by radiology departments if hospital administration wants to purchase them. Useful products will only help radiologists and probably carry the expensive pricetag. Radiology will be fine. Someone still has to assume liability for a report and referring physicians will not want to assume that responsibility.

 

[JoshSt]

In addition, we're not ignoring the technology as a field, we're trying to control it. RSNA started the spinoff Radiology AI journal and the presence of AI research at the past couple of RSNA meetings has been tremendous.

 

[Gadofosveset]

Is AI a legit concern for med students interested in rads?

It depends on what you mean.

When picking up some supplies at Target this morning I managed go through the whole process without once needing a Target employee. There was one Target employee watching six automated registers. This was maybe envisioned thirty years ago in some strategic plan, but nowhere near being put in practice.

Diagnostic medicine is going to follow the same direction as the rest of business and industry. What we know now as radiology will not be the same in ten years or then in thirty years. A lot of the stuff that gets thrown around as medical AI hype now is half truths (and a lot of outright trash research as well), but it's going to evolve. There's a kernel of truth in the idea and that's going to grow over time.

So is this a concern for a medical student? It depends on what you want out of radiology. If you want to practice radiology in 2049 like it's 2019, then that absolutely will not happen... but it's ridiculous to think that in any specialty. Will we need a vast increase in highly paid specialists to handle the exponential growth of imaging? Not possible even given the vast resources in the U.S healthcare. Try to imagine what a weird 2049 that would be. The job is going to evolve with the technology as it always has.

Is rads a bad choice for a medical student in 2019? I'm biased, but I would say it's not a bad choice in that respect. You will have to be comfortable adopting AI augmentation as it comes out. You will have to be at least passingly comfortable in (or tolerant of) computer science. You cannot fight it and pretend that it's 2003 for the next thirty years - that will certainly lead to unhappiness. Although analogies to lab medicine in pathology seem ominous, the tasks involved are much more complex and are not going to suddenly appear and displace all radiologists. Of course, some clever health system will try that, it will backfire when it screws up, and then we'll get to a balance of human and machine.

Radiologists actually use a lot of technology right now to speed them up (PACS, etc.) which has come with its own costs, but it's unthinkable to work without them and the idea of preventing PACS to allow more radiologists in the job market to read CTs on film on alternators is ludicrous in 2019. Should we have discouraged medical students in 1989 from going into radiology?

There are some who think that a procedural specialty will give them more job security than radiology. Maybe. Those specialties have their own problems and are not immune from cuts, mid-level encroachment, or tech pressure either. Seems like a bad plan if you're not procedurally inclined.

My personal take is that AI is not going to decimate the radiology workforce in thirty years... but the job description is definitely going to start changing. Depending on how good the technology is, the workforce may start decreasing slowly, but I believe a sudden unexpected radiology jobs Armageddon is a fantasy. I would be way more worried about the political climate, nationwide health care models, and reimbursement cuts. That *would* decimate the future radiology workforce, and affect everything else in health care as well, including procedural specialties.

 

[Lubeckd]

Yes, it can diagnosis as well. Im into machine learning now, in the future the deep learning can have other machines code. The instruments wouldnt be too expensive either. Its all fields that will be affected,

Predictive Analytics, Artificial Intelligence, and Population Health Management News and Insights - HealthITAnalytics

Discover cutting-edge analytics in healthcare. Insights on big data, predictive analytics, and AI transforming healthcare decision-making.

healthitanalytics.com

Not only that but if it can operate instruments and tell you the disease, its over
I just finish Data Science right now

 

[clutch21]

Yes, it can diagnosis as well. Im into machine learning now, in the future the deep learning can have other machines code. The instruments wouldnt be too expensive either. Its all fields that will be affected,

Predictive Analytics, Artificial Intelligence, and Population Health Management News and Insights - HealthITAnalytics

Discover cutting-edge analytics in healthcare. Insights on big data, predictive analytics, and AI transforming healthcare decision-making.

healthitanalytics.com

Not only that but if it can operate instruments and tell you the disease, its over
I just finish Data Science right now

The pharmacist has spoken. The field of radiology must be doomed.

 

[Lubeckd]

I said all fields. But I actually have step into this. No worries pharmacist is just one of my hats . If you want a discussion we can continue

 

[JoshSt]

Yes, it can diagnosis as well. Im into machine learning now, in the future the deep learning can have other machines code. The instruments wouldnt be too expensive either.

There are many studies like this. AI performs well with medical imaging (radiology, pathology in the linked study, and others) using robust data sets. People do this research with low-cost software or freeware. I agree if a multi-billion dollar company aggressively bought IP, pursued and achieved FDA approval, and despite all the money they invested sold this to healthcare systems for a very low cost or gave it away for free... then radiology might be in trouble.

But do you think that's going to happen? I imagine studies with AI will get more robust - predict survival, predict disease development in patients with an imaging exam that was interpreted as normal but 5 years later developed a disease. So what? Great research, but is this patent-able - maybe/maybe not. Will it need FDA approval for clinical use - probably. That's not going to be free.

When these programs are commercialized, they're not going to be cheap. You're absolutely correct it's not expensive to do the research. But is all of it going to go through the process of FDA approval? Are the companies or research teams going to pursue such approval then make it available for cheap or free? Maybe some will. Given all the companies either involved with AI or dedicated AI start-ups I've seen at the past couple of RSNA meetings, I predict they're going to be a lot of products. Some may be lucrative and worth the 20k to 100k pricetag, maybe.

Point is, it'll be expensive and I don't think healthcare systems will pay 10 to 30 different license fees for the breast cancer AI, the pancreatic cancer AI, the IBD AI, the dementia AI, the GBM AI, the meningioma AI, the CXR AI, etc. Maybe I'm wrong, we'll see.

 

[Lubeckd]

There are many studies like this. AI performs well with medical imaging (radiology, pathology in the linked study, and others) using robust data sets. People do this research with low-cost software or freeware. I agree if a multi-billion dollar company aggressively bought IP, pursued and achieved FDA approval, and despite all the money they invested sold this to healthcare systems for a very low cost or gave it away for free... then radiology might be in trouble.

But do you think that's going to happen? I imagine studies with AI will get more robust - predict survival, predict disease development in patients with an imaging exam that was interpreted as normal but 5 years later developed a disease. So what? Great research, but is this patent-able - maybe/maybe not. Will it need FDA approval for clinical use - probably. That's not going to be free.

When these programs are commercialized, they're not going to be cheap. You're absolutely correct it's not expensive to do the research. But is all of it going to go through the process of FDA approval? Are the companies or research teams going to pursue such approval then make it available for cheap or free? Maybe some will. Given all the companies either involved with AI or dedicated AI start-ups I've seen at the past couple of RSNA meetings, I predict they're going to be a lot of products. Some may be lucrative and worth the 20k to 100k pricetag, maybe.

Point is, it'll be expensive and I don't think healthcare systems will pay 10 to 30 different license fees for the breast cancer AI, the pancreatic cancer AI, the IBD AI, the dementia AI, the GBM AI, the meningioma AI, the CXR AI, etc. Maybe I'm wrong, we'll see.

the problem is even the people in the IT jobs are going to be affected. if you program a computer to code, its can code more programs. Software engineering and machine learning isn't safe either, that's why I said its going to affect all fields unless there is some kind of intervention

 

[Lubeckd]

There are many studies like this. AI performs well with medical imaging (radiology, pathology in the linked study, and others) using robust data sets. People do this research with low-cost software or freeware. I agree if a multi-billion dollar company aggressively bought IP, pursued and achieved FDA approval, and despite all the money they invested sold this to healthcare systems for a very low cost or gave it away for free... then radiology might be in trouble.

But do you think that's going to happen? I imagine studies with AI will get more robust - predict survival, predict disease development in patients with an imaging exam that was interpreted as normal but 5 years later developed a disease. So what? Great research, but is this patent-able - maybe/maybe not. Will it need FDA approval for clinical use - probably. That's not going to be free.

When these programs are commercialized, they're not going to be cheap. You're absolutely correct it's not expensive to do the research. But is all of it going to go through the process of FDA approval? Are the companies or research teams going to pursue such approval then make it available for cheap or free? Maybe some will. Given all the companies either involved with AI or dedicated AI start-ups I've seen at the past couple of RSNA meetings, I predict they're going to be a lot of products. Some may be lucrative and worth the 20k to 100k pricetag, maybe.

Point is, it'll be expensive and I don't think healthcare systems will pay 10 to 30 different license fees for the breast cancer AI, the pancreatic cancer AI, the IBD AI, the dementia AI, the GBM AI, the meningioma AI, the CXR AI, etc. Maybe I'm wrong, we'll see.

I think at this point its better to make your own tech startup as a MD. maybe basic income will come

 

[Lubeckd]

At the rate things are going if you can't beat them join them. MD/Heaklthcare salary is a lot, a lot of outsourcing on images on globalization has been affecting it but now diagosising. I code everyday and I plan to make the transition, healthcare isn't king anymore and they are looking to cut the salary on the cutting block. For some machine learning technique now makes your life easier but data is available to diagnosis. BIG DATA. and tech will make more advancement.

 

[Lubeckd]

well the tech changes every 2-3 year. So you have to keep up a CE along with your MD license...

 

[Lubeckd]

There are many studies like this. AI performs well with medical imaging (radiology, pathology in the linked study, and others) using robust data sets. People do this research with low-cost software or freeware. I agree if a multi-billion dollar company aggressively bought IP, pursued and achieved FDA approval, and despite all the money they invested sold this to healthcare systems for a very low cost or gave it away for free... then radiology might be in trouble.

But do you think that's going to happen? I imagine studies with AI will get more robust - predict survival, predict disease development in patients with an imaging exam that was interpreted as normal but 5 years later developed a disease. So what? Great research, but is this patent-able - maybe/maybe not. Will it need FDA approval for clinical use - probably. That's not going to be free.

When these programs are commercialized, they're not going to be cheap. You're absolutely correct it's not expensive to do the research. But is all of it going to go through the process of FDA approval? Are the companies or research teams going to pursue such approval then make it available for cheap or free? Maybe some will. Given all the companies either involved with AI or dedicated AI start-ups I've seen at the past couple of RSNA meetings, I predict they're going to be a lot of products. Some may be lucrative and worth the 20k to 100k pricetag, maybe.

Point is, it'll be expensive and I don't think healthcare systems will pay 10 to 30 different license fees for the breast cancer AI, the pancreatic cancer AI, the IBD AI, the dementia AI, the GBM AI, the meningioma AI, the CXR AI, etc. Maybe I'm wrong, we'll see.

You just need to find the right vcs to fund you...

 

[CharlieBillings]

.

 

[Lubeckd]

It seems like the biggest threat to radiology jobs/earnings is changes in reimbursement. Although I do wonder if AI could be introduced that does not replace radiologists, but simply makes them more efficient (and thus you would need less radiologists).

well the jobs the machine does can free you up to other work. Yes the reimbursements are getting less and less. Thanks for the input. This is why people are voting for andrew yang

 

[trying_to_be_rad]

Lubeckd -

I say this with no intention to offend, but you are speaking in mainstream generalizations common in the popular science/futurism crowd ("machines will teach other machines to code", "find the right VCs"). Worse, you say things like "people are voting for andrew yang" because "reimbursements are getting less and less"... when the current declines in reimbursements for radiology, and for medicine in general (including surgical subpsecialties like ortho), have nothing to do with AI. I don't want to emphasize the fact that you are not a physician. However, because you are not one, I don't think you are fully aware of the unique opportunities and challenges that AI poses for medicine, just as I wouldn't presume to know the issues you face as a pharmacist.

I am a radiology resident, but I also know how to code and work with higher-level data (i.e. not just plugging big data into a stats software to churn out P-values). In fact, that seems to be an increasingly common trend in radiology - three other residents in my program were CS majors and hope to work in imaging informatics. I will tell you that all of us are embracing the opportunity for AI to facilitate disease diagnosis, because that improves patient care, However, we don't see how currently crude deep learning models can replace radiologists any time soon. For instance, let's take an abdominal CT for cancer restaging after lung adenocarcinoma to monitor metastasis to other systems. Because this is a cross-sectional modality, with at least 200 sections per plane (coronal, axial, sagittal), in addition to images produced during different phases of contrast, it would be incredibly difficult to gather a robust dataset to train a binary classifier (think about it - you would need thousands of images for EACH section in EACH plane to train the model). In another instance, let's take an ICU patient with a multitude of lines (Swan-Ganz catheters, NG tube, etc.). In this scenario, it is difficult to train a classifier to check whether the NG tube was correctly placed and did not slip into the lung, mainly due to the large number of artifacts introduced by the numerous other lines in the patient. This is why the majority of current models are trained under a carefully curated, controlled environment (e.g. detecting pneumonia on a chest xray) - otherwise, it is just too difficult. In addition, a radiologist is not just someone who looks for abnormalities in images - the job of a radiologist is to MAKE the diagnosis, which requires a very confident grasp of medicine and pathophysiology, and the ability to generate a clinical picture from the patient's medical history (for instance, a hepatic lesion on CT abdomen/pelvis concerning for mets vs artifact in a patient with pancreatic adenocarcinoma s/p Whipple - oh wait, the Whipple was 20 years ago? Okay that's not a met, nothing to see here, folks.)

Out of curiosity, what software packages have you used and what data science experience have you had? Is it limited to scikit-learn and TensorFlow on python? Have you actually trained a model and translated it to production? I am curious to hear what experience you have had to say such things with confidence.

 

[Lubeckd]

Lubeckd -

I say this with no intention to offend, but you are speaking in mainstream generalizations common in the popular science/futurism crowd ("machines will teach other machines to code", "find the right VCs"). Worse, you say things like "people are voting for andrew yang" because "reimbursements are getting less and less"... when the current declines in reimbursements for radiology, and for medicine in general (including surgical subpsecialties like ortho), have nothing to do with AI. I don't want to emphasize the fact that you are not a physician. However, because you are not one, I don't think you are fully aware of the unique opportunities and challenges that AI poses for medicine, just as I wouldn't presume to know the issues you face as a pharmacist.

I am a radiology resident, but I also know how to code and work with higher-level data (i.e. not just plugging big data into a stats software to churn out P-values). In fact, that seems to be an increasingly common trend in radiology - three other residents in my program were CS majors and hope to work in imaging informatics. I will tell you that all of us are embracing the opportunity for AI to facilitate disease diagnosis, because that improves patient care, However, we don't see how currently crude deep learning models can replace radiologists any time soon. For instance, let's take an abdominal CT for cancer restaging after lung adenocarcinoma to monitor metastasis to other systems. Because this is a cross-sectional modality, with at least 200 sections per plane (coronal, axial, sagittal), in addition to images produced during different phases of contrast, it would be incredibly difficult to gather a robust dataset to train a binary classifier (think about it - you would need thousands of images for EACH section in EACH plane to train the model). In another instance, let's take an ICU patient with a multitude of lines (Swan-Ganz catheters, NG tube, etc.). In this scenario, it is difficult to train a classifier to check whether the NG tube was correctly placed and did not slip into the lung, mainly due to the large number of artifacts introduced by the numerous other lines in the patient. This is why the majority of current models are trained under a carefully curated, controlled environment (e.g. detecting pneumonia on a chest xray) - otherwise, it is just too difficult. In addition, a radiologist is not just someone who looks for abnormalities in images - the job of a radiologist is to MAKE the diagnosis, which requires a very confident grasp of medicine and pathophysiology, and the ability to generate a clinical picture from the patient's medical history (for instance, a hepatic lesion on CT abdomen/pelvis concerning for mets vs artifact in a patient with pancreatic adenocarcinoma s/p Whipple - oh wait, the Whipple was 20 years ago? Okay that's not a met, nothing to see here, folks.)

Out of curiosity, what software packages have you used and what data science experience have you had? Is it limited to scikit-learn and TensorFlow on python? Have you actually trained a model and translated it to production? I am curious to hear what experience you have had to say such things with confidence. I dont know what sets of data you are using in the hospital. lol, its not just mean, mode and p value. Its skewness and so many other topics.

Yes I have with scikit-learn and Tensorflow through datacamp, its the deep learning that is a bit hard, Im working with people in the field. I know R and Python and I have also finished Quant finance. I finish the complete Data Science course, if you are curious give me a shoot Ill send you my cerf. Because I work with people in the industry this is the same thing they worry about. I havent trained one into production but I will show you my trading algos soon. No offend taken, Im not a doctor and dont assume I know. Im just putting out the general challenges out there and I love these conversation but dont think because you are a "Doctor" you can be high and mighty.

Oh I said vote for Andrew Yang because robots may take over. all fields are going to be affected. Its something you learn example of the moment you take supervised, unsupervised learning. You seem to know more about healthcare tech, are you using machine learning there? There is no higher data, its all just data.

 

[Lubeckd]

I can also tell you in pharmacy informatics none of what you and I just said applies. Im not sure, but I dont think the healthcare field works with Data at all. And machine learning is just Stats on data. Data science is multidisciplinary , not Computer science. I also have to deal with Skewness and Kurtosis. And deep learning applies a lot of stats on the nodes.

Here are some topics
Variance, t test p test, hypothesis testing, standard dev, kurtosis, skewness, prob density function
Maximum likelihood estimator MLE

Also I happen to be very involved with the startup people and tech founders, as well as VC. I say it because I know you just need to start and find the funding. Which is why I said a MD can do it too

 

[Lubeckd]

https://thumbor.forbes.com/thumbor/...018/09/Jobs-Upskill-Route-150DPI-1200x856.jpg

You can laugh at data all you want. But its already been said its the future

 

[trying_to_be_rad]

I apologize if I came off as "high and mighty." That was not my intention. I was simply concerned that you were speaking in generalizations, many of which border on science fiction and won't become relevant for a very long time. This was a thread started by a medical student looking for well-informed advice on his/her career choice. Such advice should be based on experiences from health care experts in the field. For pharmacy career advice, wouldn't you turn to senior pharmacists? And wouldn't you try to correct "high and mighty" physicians who claim that we should no longer train pharmacists because Epic automatically does all our dosing anyway (NOT true, by the way - pharmacists have saved my ass so many times while I was an intern)? No one is questioning that AI will have a big part to play in our future. Current predictions, however, seem to be overblown by hype.

I think we are venturing a bit off topic - but sure, checking normality in your data is important for generating a reliable model. Healthcare field doesn't work with data at all? That is an interesting observation, as too much literature in medicine has been generated using big data in the past 10 years. Look at the National Inpatient Sample, NSQIP, National Cancer Databse, SEER... all datasets containing millions of data points.

 

[Lubeckd]

I apologize if I came off as "high and mighty." That was not my intention. I was simply concerned that you were speaking in generalizations, many of which border on science fiction and won't become relevant for a very long time. This was a thread started by a medical student looking for well-informed advice on his/her career choice. Such advice should be based on experiences from health care experts in the field. For pharmacy career advice, wouldn't you turn to senior pharmacists? And wouldn't you try to correct "high and mighty" physicians who claim that we should no longer train pharmacists because Epic automatically does all our dosing anyway (NOT true, by the way - pharmacists have saved my ass so many times while I was an intern)? No one is questioning that AI will have a big part to play in our future. Current predictions, however, seem to be overblown by hype.

I think we are venturing a bit off topic - but sure, checking normality in your data is important for generating a reliable model. Healthcare field doesn't work with data at all? That is an interesting observation, as too much literature in medicine has been generated using big data in the past 10 years. Look at the National Inpatient Sample, NSQIP, National Cancer Databse, SEER... all datasets containing millions of data points.

Im actually looking into pharmacy informatics and it applies non of the data analytics and its more project management. I say my lifespan in pharmacy is another 3-5 years. Im trying to leave very bad. Pharmacy has their own problems, as I said earlier it can help reduce your workload (machines) its like automation for pills. it reduces workload so you can focus on other things. I say within 6 months I would leave pharmacy for Tech/Finance very soon. Even fintech like blockchains are coming into the hospital to protect patient information. Ill try to make a healthcare data with machine learning soon and send to you via github if you are interested. Im very interested in these topics.

 

[Lubeckd]

I apologize if I came off as "high and mighty." That was not my intention. I was simply concerned that you were speaking in generalizations, many of which border on science fiction and won't become relevant for a very long time. This was a thread started by a medical student looking for well-informed advice on his/her career choice. Such advice should be based on experiences from health care experts in the field. For pharmacy career advice, wouldn't you turn to senior pharmacists? And wouldn't you try to correct "high and mighty" physicians who claim that we should no longer train pharmacists because Epic automatically does all our dosing anyway (NOT true, by the way - pharmacists have saved my ass so many times while I was an intern)? No one is questioning that AI will have a big part to play in our future. Current predictions, however, seem to be overblown by hype.

I think we are venturing a bit off topic - but sure, checking normality in your data is important for generating a reliable model. Healthcare field doesn't work with data at all? That is an interesting observation, as too much literature in medicine has been generated using big data in the past 10 years. Look at the National Inpatient Sample, NSQIP, National Cancer Databse, SEER... all datasets containing millions of data points.

Im actually looking for a site that has healthcare data I can play with for machine learning. It would help my project.
Well I say Pharmacy doesnt work with data at all, I keep asking IT pharmacist to give me data and where to find it and they all say the field is more in project management
I dont plan to stay in Data though. Ill only apply machine learning to my trading algorithms as a Quant. Software engineering is a much better route for me, even Im sick of the data and math

 

[wandering star]

I currently work at the intersection of AI, radiology, and physics. Here are my observations on this thread:

Is AI a legitimate concern that current med students should take into account when deciding to pursue rads? What evidence is there that shows it is or is not a concern for someone currently in med school who wants to have a 30+ year career in rads?

Deep learning effectively "started" in 2013 when a neural network (AlexNet, primitive by today's standards) outclassed all existing approaches in a giant challenge. The techniques have improved considerably in the past 6 years and, more importantly, they've changed in surprising ways. Around 2015, generative adversarial networks, the ones that make fake faces, appeared. They've given us fundamentally new capabilities; soon we won't be able to tell whether the average online video is real or fake.

The real problem for trying to extrapolate out 30 years, then, is that we don't know what we don't know. A really new network might appear in 2021, and another one in 2025. It's hard to say now whether we'll plateau or continue to make new progress. For example, while networks today have *some* capacity to generalize and transfer knowledge, there might be a design in the future that does so much better, thus requiring much less data to train.

To go a completely different direction, a consortium of academic centers might get together and provide each patient a waiver for them to release their data publicly with annotations. If this happens and a quarter of patients sign the form for science, we'd get access to much "bigger" data overnight.

However, we don't see how currently crude deep learning models can replace radiologists any time soon. For instance, let's take an abdominal CT for cancer restaging after lung adenocarcinoma to monitor metastasis to other systems. Because this is a cross-sectional modality, with at least 200 sections per plane (coronal, axial, sagittal), in addition to images produced during different phases of contrast, it would be incredibly difficult to gather a robust dataset to train a binary classifier (think about it - you would need thousands of images for EACH section in EACH plane to train the model).

I don't actually think this is a huge problem. You wouldn't want a binary classifier - you'd just have to provide enough annotations so that all the mets would be marked. A binary classifier ("does this dataset have mets or not") would be hard to train, but if you tell the network *where* the mets are and provide one click in 3D per met, it can automatically be registered in the sections (axial vs sagittal) and with a little more creativity it can be registered across the different contrast phases too.

Finalized AI products are going to be very expensive and probably only mildly embraced by radiology departments. I predict the majority of products with a $20k - $100k annual fee will have great opposition by radiology departments if hospital administration wants to purchase them.

Of all the translational issues, I am least worried about cost. Companies would sell the products in a way that you would want to buy them.

In the intermediate term, I would predict that AI products would be designed to improve efficiency. Suppose as you read the study the AI includes a list of possible annotations, and supports each annotation with e.g. automatic measurements ("lesion is 22 mm by 15 mm"), highlights relevant information in the EMR, etc. AI would not be designed to replace the entire radiologist but would make parts of your job faster. If, on average, it saves you 10 minutes an hour, and AI is priced per scan, and the cost is a fraction of the time savings converted to dollars, we're looking at a potential win-win-win: findings are caught that might have been missed; hospital saves money; radiologist becomes more efficient. The number of radiologists that society needs would be reduced, but not dramatically so.

 

[YoungDoc25]

Kind of related, but are there books or podcasts y’all would recommend to have a better basic understanding of AI? Planning on going into radiology and feel like a basic understanding could benefit me down the road when the fields become more intertwined

 

[Tiger100]

To OP:
I wouldn't be worried about it.

If we reach a point that AI replaces Radiologists, then most jobs including most medical specialties will be heavily affected.

However, if you really lose sleep over it, don't do it and choose another field.

Honestly, I am not sure whether going into medical school these days worth it or not.
I've had a very good career in medicine and have made a $hitload of money. But my best guess is that the future of medicine won't be bright. In my generation, people can retire after 15-20 years of work but it won't happen in the current generation.

 

[trying_to_be_rad]

I don't actually think this is a huge problem. You wouldn't want a binary classifier - you'd just have to provide enough annotations so that all the mets would be marked. A binary classifier ("does this dataset have mets or not") would be hard to train, but if you tell the network *where* the mets are and provide one click in 3D per met, it can automatically be registered in the sections (axial vs sagittal) and with a little more creativity it can be registered across the different contrast phases too.

I agree with that. This example would be more consistent with an AI system that augments the radiologist - the AI highlights areas of high suspicion on the scan, which the radiologist can then confirm based on clinical acumen. But as you said, training a true binary classifier with high accuracy and positive predictive value (which in my opinion, would truly "replace" a radiologist) is quite some ways off. At this point, I would definitely welcome the ability of AI to make my job easier, and I also find it intellectually rewarding to contribute to the development of such products.

Of all the translational issues, I am least worried about cost. Companies would sell the products in a way that you would want to buy them.

In the intermediate term, I would predict that AI products would be designed to improve efficiency. Suppose as you read the study the AI includes a list of possible annotations, and supports each annotation with e.g. automatic measurements ("lesion is 22 mm by 15 mm"), highlights relevant information in the EMR, etc. AI would not be designed to replace the entire radiologist but would make parts of your job faster. If, on average, it saves you 10 minutes an hour, and AI is priced per scan, and the cost is a fraction of the time savings converted to dollars, we're looking at a potential win-win-win: findings are caught that might have been missed; hospital saves money; radiologist becomes more efficient. The number of radiologists that society needs would be reduced, but not dramatically so.

I definitely think this is the direction that radiology is headed... although I imagine that integrating AI into the EMR in such a way to accurately parse through the free text in the patient charts would be a separate challenge in natural language processing. The demand for imaging continues to grow, but it seems that ACR wants to suppress the number of radiologists being trained (I've heard that in New York state, there is a hard cap to the number of radiology residents, preventing program expansion). I imagine that they want to avoid what happened to rad onc and pathology - if you flood the market with too many graduates when the market is good, everyone loses when the market inevitably shrinks again. However, as a result, many residents and recent grads are overworked - continuous secular growth of imaging, but not enough radiologists to meet the demand. In this case, I think AI is a good solution for keeping the supply of radiologists low while maintaining a reasonable workload for each radiologist.

 

[shadowlightfox]

If AI hasn't even replaced physicians signing off on EKG readings I don't see how they can replace radiologists.

 

[DOFOSHO]

Its only a concern if your passion is finding lung nodules, analyzing thyroid nodules, or measuring lymph nodes.