AI threat to rad onc

Get help with your application

Use all the free resources available to you from SDN: articles, guides, expert advising, forums discussions, and school research.

DrProtonX

Full Member
2+ Year Member
FC356F0E-DCC3-43F8-B058-35F8D871C127.png

“62% subsitutable by AI” … what do you guys think? Is future of rad onc in danger?
 
So 62% of radiation oncology payments are classified as "nonphysical," and therefore 62% of our specialty is potentially replaceable by AI? That's quite the intellectual leap, even by SDN standards. AI will absolutely (and already has, in some cases) automate substantial portions of our workflow. Contouring, treatment planning, documentation, image review, plan QA, etc. Frankly, bring it on. I didn't spend half my adult life in training to spend my days drawing parotid glands and clicking through MOSAIQ.

But there's a rather significant difference between automating a task and replacing the physician responsible for the clinical decisions. Someone still has to decide whether the patient should receive radiation in the first place, what to treat, what not to treat, how to balance competing risks, and what to do when the patient has already received three courses of radiation and the spinal cord has approximately 0.2 Gy of remaining tolerance. The legitimate concern is economic. If AI allows one RO to do the work previously requiring two, that has obvious implications for physician demand and compensation. Pretending otherwise would be foolish.

That said, I'd still rank CMS reimbursement, hypofractionation, payer shenanigans, and our profession's remarkable enthusiasm for producing more graduates than available jobs considerably higher on my list of existential threats.
 
So 62% of radiation oncology payments are classified as "nonphysical," and therefore 62% of our specialty is potentially replaceable by AI? That's quite the intellectual leap, even by SDN standards. AI will absolutely (and already has, in some cases) automate substantial portions of our workflow. Contouring, treatment planning, documentation, image review, plan QA, etc. Frankly, bring it on. I didn't spend half my adult life in training to spend my days drawing parotid glands and clicking through MOSAIQ.

But there's a rather significant difference between automating a task and replacing the physician responsible for the clinical decisions. Someone still has to decide whether the patient should receive radiation in the first place, what to treat, what not to treat, how to balance competing risks, and what to do when the patient has already received three courses of radiation and the spinal cord has approximately 0.2 Gy of remaining tolerance. The legitimate concern is economic. If AI allows one RO to do the work previously requiring two, that has obvious implications for physician demand and compensation. Pretending otherwise would be foolish.

That said, I'd still rank CMS reimbursement, hypofractionation, payer shenanigans, and our profession's remarkable enthusiasm for producing more graduates than available jobs considerably higher on my list of existential threats.
Funny enough if you think about it if we had not gone to computerized planning and were still using china markers on films and spending an hour or more in the simulator for each sim and hand drawing blocks ... we'd be A.I. untouchable. In fact, now that I think about it, a pure benign RT practice could be A.I. untouchable if it wanted to set itself up to be so, and the plan and RT quality would be totally fine. A pure brachy practice would be pretty A.I. untouchable.
 
The table is obviously not complete.

I do wonder however, how the jobs of
- intensive care physicians
- oncologists
- infectiologists
will change, just to name a few.

Yes, ICU physicians still have to do all the manual stuff, like inserting tubes and so on. But alot of their medical decisions, esp. concerning circulation and ventilation are exactly the stuff that AI can perform in an excellent manner. The ICU is the optimal enviroment for that: an ICU is monitored more than any other patient 24/7 and all those values are there. Wouldn't AI do a better job than an ICU doctor in deciding stuff like: "should we lover O2 today", " what dose of catecholamines does the patient need today", etc.

Yes, oncologists still have to talk to patients and so on (pretty much like we have to do), and may do an actual procedure now and then, like a bone marrow biopsy. But wouldn't AI make a better judgement on what drug to propose as next line of treatment or what dose, having access to all the molecular, pathology, radiology data and the patient's lab values. No, if the nurse has actually put in the patients CTCAE toxicities in a chart, or (God forbid!) the patient has actually done that from home on his phone with the PROMs-app, who is actually making decisions?

Yes, infectiologists still have to go on rounds and discuss with the internal medicine physicians about the patients. But, can't AI actually judge better, what antibiotics a patient needs, based on all the lab values, antibiograms and vital parameters, which are all on the EMR?
 
The table is obviously not complete.

I do wonder however, how the jobs of
- intensive care physicians
- oncologists
- infectiologists
will change, just to name a few.

Yes, ICU physicians still have to do all the manual stuff, like inserting tubes and so on. But alot of their medical decisions, esp. concerning circulation and ventilation are exactly the stuff that AI can perform in an excellent manner. The ICU is the optimal enviroment for that: an ICU is monitored more than any other patient 24/7 and all those values are there. Wouldn't AI do a better job than an ICU doctor in deciding stuff like: "should we lover O2 today", " what dose of catecholamines does the patient need today", etc.

Yes, oncologists still have to talk to patients and so on (pretty much like we have to do), and may do an actual procedure now and then, like a bone marrow biopsy. But wouldn't AI make a better judgement on what drug to propose as next line of treatment or what dose, having access to all the molecular, pathology, radiology data and the patient's lab values. No, if the nurse has actually put in the patients CTCAE toxicities in a chart, or (God forbid!) the patient has actually done that from home on his phone with the PROMs-app, who is actually making decisions?

Yes, infectiologists still have to go on rounds and discuss with the internal medicine physicians about the patients. But, can't AI actually judge better, what antibiotics a patient needs, based on all the lab values, antibiograms and vital parameters, which are all on the EMR?
The USMLE literally can’t write a question AI gets wrong last I heard
 
Funny enough if you think about it if we had not gone to computerized planning and were still using china markers on films and spending an hour or more in the simulator for each sim and hand drawing blocks ... we'd be A.I. untouchable. In fact, now that I think about it, a pure benign RT practice could be A.I. untouchable if it wanted to set itself up to be so, and the plan and RT quality would be totally fine. A pure brachy practice would be pretty A.I. untouchable.
Electrons, my friend. Electrons are still A.I.-untouchable.
Noone really knows where electrons land and what dose they deliver. It's one of life's mysteries.
 
In my opinion, distrust is what essentially all physicians have in their corner. I suspect it will be the better part of most of our careers before the general public will accept truly autonomous AI for health care, even if it were capable.

Agree with GFunk that AI to this point has made my life easier. Are AI generated contours perfect? Nope. I have yet to meet one that can reliable distinguish between bowel and pelvic vessels on a non contrasted scan. But are they close enough for a simple prostate plan? Absolutely. Even for an N+ case, I’m just tweaking the AI normal tissues instead of starting from scratch.
 
A study that puts the radonc market at $1.2B when 21C was doing $1B per year in annual revenue at its peak a decade ago is very not serious.
 
In my opinion, distrust is what essentially all physicians have in their corner. I suspect it will be the better part of most of our careers before the general public will accept truly autonomous AI for health care, even if it were capable.

Agree with GFunk that AI to this point has made my life easier. Are AI generated contours perfect? Nope. I have yet to meet one that can reliable distinguish between bowel and pelvic vessels on a non contrasted scan. But are they close enough for a simple prostate plan? Absolutely. Even for an N+ case, I’m just tweaking the AI normal tissues instead of starting from scratch.


bowel awful.

if you do a ton of abdominal sbrt it matters and it's such a time sink. hope it gets better
 
I've found that, despite AI-financed reports breathlessly extolling the virtues of their great machines...it's still faster for me to contour myself than go through slice-by-slice and correct what AI would have done, at least when it comes to GTV/CTV/PTV/etc.
 
Sure agree just the part by the high dose PTV but still it’s an annoying task IMO esp if there a lot of bowel motion near tumor and you want to account for it
 
bowel awful.

if you do a ton of abdominal sbrt it matters and it's such a time sink. hope it gets better
It depends on how complex it is. For isolated nodes which will be pretty conformal, just having to tweak around the PTV is still a win. For something like pancreatic SBRT where you really push the conformality against abutting structures and you know there will be high dose streaks…way less helpful. For those I delete all bowel AI contours and do them myself.
Sure agree just the part by the high dose PTV but still it’s an annoying task IMO esp if there a lot of bowel motion near tumor and you want to account for it
breath hold is a way better solution than AI.
 
On a related note, today my physicist wanted to show me the Brainlab Elements autocontour software for brain mets. Of the three mets in my patient, it got one exactly right, missed a slice on another, and missed the third met entirely.
 
On a related note, today my physicist wanted to show me the Brainlab Elements autocontour software for brain mets. Of the three mets in my patient, it got one exactly right, missed a slice on another, and missed the third met entirely.
Sounds like some of the rads reads I've seen over the years