ESTRO meets Asia 2026 Presentation: Prostate AI plan in < 1 sec, quality comparable to human Dosimetrist

Started by Gfunk6
Apply Smarter with MyERAS
This forum made possible through the generous support of SDN members, donors, and sponsors. Thank you.
Get help with your application

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

Browse articles Compare schools My application plan Read Interview Feedback
Practice for interviews Ask an Advisor
Gfunk6

Gfunk6

And to think . . . I hesitated
Moderator Emeritus
Lifetime Donor
Verified Member
Physician
Faculty
PhD
Verified Expert
20+ Year Member
Gold Donor
Advertisement - Members don't see this ad

AI End-to-End Radiation Treatment Planning Under One Second

Artificial intelligence-based radiation therapy (RT) planning has the potential to reduce planning time and inter-planner variability, improving efficiency and consistency in clinical workflows. Most existing automated approaches rely on multiple dose evaluations and corrections, resulting in...
arxiv.org arxiv.org

AI End-to-End Radiation Treatment Planning Under One Second​

Simon Arberet, Riqiang Gao, Martin Kraus, Florin C. Ghesu, Wilko Verbakel, Mamadou Diallo, Anthony Magliari, Venkatesan Karuppusamy, Sushil Beriwal, REQUITE Consortium, Ali Kamen, Dorin Comaniciu

Artificial intelligence-based radiation therapy (RT) planning has the potential to reduce planning time and inter-planner variability, improving efficiency and consistency in clinical workflows. Most existing automated approaches rely on multiple dose evaluations and corrections, resulting in plan generation times of several minutes. We introduce AIRT (Artificial Intelligence-based Radiotherapy), an end-to-end deep-learning framework that directly infers deliverable treatment plans from CT images and structure contours. AIRT generates single-arc VMAT prostate plans, from imaging and anatomical inputs to leaf sequencing, in under one second on a single Nvidia A100 GPU. The framework includes a differentiable dose feedback, an adversarial fluence map shaping, and a plan generation augmentation to improve plan quality and robustness. The model was trained on more than 10,000 intact prostate cases. Non-inferiority to RapidPlan Eclipse was demonstrated across target coverage and OAR sparing metrics. Target homogeneity (HI = 0.10 ± 0.01) and OAR sparing were similar to reference plans when evaluated using AcurosXB. These results represent a significant step toward ultra-fast standardized RT planning and a streamlined clinical workflow.
Click to expand...

This is the kind of AI in radiation oncology that actually matters. >10,000 prostate cases, CT + contours in, deliverable single-arc VMAT plan out in under one second, including leaf sequencing, with dosimetry non-inferior to Eclipse RapidPlan on AcurosXB. Yes, prostate is the easiest possible sandbox, and yes, I want to see this on H&N, pancreas, reirradiation, etc. But it is getting harder to pretend the future of treatment planning is humans spending 45 minutes lovingly nudging objectives around until the DVH looks pretty. Routine planning is going to become a commodity. The useful work shifts to contouring, prescription strategy, QA, and catching the occasional spectacularly stupid machine-generated plan. If this generalizes, bragging about optimization speed is eventually going to sound like bragging that your calculator does division really fast.
 
Advertisement - Members don't see this ad
Gfunk6 said:
The useful work shifts to contouring
Click to expand...
But that will be "easily" A.I. solvable too

I remember early in my career, treating ~30-40/day, I had to devote a full half day on the schedule each week just for contouring

Are we still going to need more and more rad oncs if we are simultaneously needing less and less dosimetrists
 
TheWallnerus said:
But that will be "easily" A.I. solvable too

I remember early in my career, treating ~30-40/day, I had to devote a full half day on the schedule each week just for contouring

Are we still going to need more and more rad oncs if we are simultaneously needing less and less dosimetrists
Click to expand...
As the technology advances and the support staff needs are reduced, radiation oncology will find its day to day more and more limited to the clinical aspects of care. I can see our daily practice resembling Med Onc
 
fiji128 said:
Not hard to believe that things won't be heading this way in the not too distance future. Actually kinda surprised it isn't already here.
Click to expand...
I think there have been cultural forces at play keeping this from happening.

Planning is made for neural network/iterative learning type stuff IMO. There were initiatives regarding this almost 20 years ago that worked pretty well.

Among the things that AI can do is reference an essentially infinite catalogue of prior plans and build on existing plans based on anatomical/target homology. This is a self-improving process.

Among the things that some docs/dosimetrists do is to tweak plans within levels of subtlety that are undoubtedly clinically irrelevant. This has delayed automation IMO.
 
Dosimetry will undoubtedly be replaced by AI. It’s sad to think because I love my dosimetrists. I’ve even told some therapists not to go into dosimetry because AI makes it an uncertain future. My real fear is that just like neuro onc is can be a one year fellowship, I can see some surgeons or med oncs do a 1 year rad onc fellowship if contours and plans are automatically done for you (I’d say this is 20 years away?). Especially if it’s just say GU rad onc or lung rad onc only. I hope I’m wrong
 
RadOncMegatron said:
Dosimetry will undoubtedly be replaced by AI. It’s sad to think because I love my dosimetrists. I’ve even told some therapists not to go into dosimetry because AI makes it an uncertain future. My real fear is that just like neuro onc is can be a one year fellowship, I can see some surgeons or med oncs do a 1 year rad onc fellowship if contours and plans are automatically done for you (I’d say this is 20 years away?). Especially if it’s just say GU rad onc or lung rad onc only. I hope I’m wrong
Click to expand...

I doubt very seriously that another specialty would do more training to get into radonc when you can get them cheap nowadays
 
RadOncMegatron said:
Dosimetry will undoubtedly be replaced by AI. It’s sad to think because I love my dosimetrists. I’ve even told some therapists not to go into dosimetry because AI makes it an uncertain future. My real fear is that just like neuro onc is can be a one year fellowship, I can see some surgeons or med oncs do a 1 year rad onc fellowship if contours and plans are automatically done for you (I’d say this is 20 years away?). Especially if it’s just say GU rad onc or lung rad onc only. I hope I’m wrong
Click to expand...
If anything I see maybe a clinical oncology specialty emerging (a la UK) given the med onc shortage and the coming rad onc oversupply (2 year rad onc fellowship that would become the first useful/accredited fellowship in our specialty).
 
Do you guys really have to be cynical and negative about everything? We constantly hear that AI will replace humans, but most of this hype is pushed by people who have zero understanding of either AI or healthcare. This development will probably have little to no impact on radiation oncologists. It is much more relevant to dosimetrists. Honestly, this forum becomes hysterical about everything, and every discussion seems to have the same conclusion “Radiation oncology is doomed” or “radiation oncology is bad.”

How many times have we heard that AI will replace radiologists or pathologists, the most obvious targets? and now radiology is the hottest field in medicine with the best lifestyle and salary. Radiation oncology is probably among the least threatened specialties because we physically deliver potentially lethal radiation. The dose is irreversible and unsalvageable. A surgical robot can be stopped, and some surgical errors can be repaired, but radiation? one extra digit or the wrong target or any tiny mistake and its game over.... no matter what you do.

Meanwhile, Johns Hopkins already has a robot that performed a lengthy phase of gallbladder surgery on a lifelike patient without human help, adapted to unexpected situations, and is being developed toward complete autonomous surgery. Robot performs first realistic surgery without human help | Hub

Ophthalmic procedures are computer driven. Once the ophthalmologist selects the target, newer systems can perform the procedure autonomously. One system performed targeted retinal injections with 100% success and better accuracy than manual or surgeon-controlled robotic surgery. Autonomous robotic intraocular surgery for targeted retinal injections

I also find it hilarious that you think MedOnc will replace RadOnc and not the other way around. Feed a high-end AI model the history, pathology, imaging, molecular results, surgical findings, radiation options and guidelines, and it can produce a systemic therapy plan equal or superior to your MedOnc that you work with. A family physicain or medical student or pre-med or NP or PA can then order and administer it through Epic.

Even the MedOnc part of immunotherapy toxicity management can easily be automated by AI. Grade the toxicity, hold or permanently discontinue the drug, start steroids, plan the taper, reassess, escalate to second line immunosuppressant and determine if rechallenge is allowed. It is basically a guideline algorithm and checkpoint inhibitors are the same thing besically.

And when the toxicity becomes grade 3 or 4, especially grade 4, MedOnc don't manage the affected organ, the relevant specialist manage it the best thing they can do is write note of what happend in Epic and do what they are told by surgeons or relevant service. If AI selects the same treatment, the probability of toxicity is the same whether the person using it is a MedOnc, nurse or a janitor with AI access. That probability is not determined by planning or any special skills or knowledge from MedOnc.

Sorry to ruin the self-loathing party better luck next time 🙁
 
Beerus said:
Do you guys really have to be cynical and negative about everything? We constantly hear that AI will replace humans, but most of this hype is pushed by people who have zero understanding of either AI or healthcare. This development will probably have little to no impact on radiation oncologists. It is much more relevant to dosimetrists. Honestly, this forum becomes hysterical about everything, and every discussion seems to have the same conclusion “Radiation oncology is doomed” or “radiation oncology is bad.”

How many times have we heard that AI will replace radiologists or pathologists, the most obvious targets? and now radiology is the hottest field in medicine with the best lifestyle and salary. Radiation oncology is probably among the least threatened specialties because we physically deliver potentially lethal radiation. The dose is irreversible and unsalvageable. A surgical robot can be stopped, and some surgical errors can be repaired, but radiation? one extra digit or the wrong target or any tiny mistake and its game over.... no matter what you do.

Meanwhile, Johns Hopkins already has a robot that performed a lengthy phase of gallbladder surgery on a lifelike patient without human help, adapted to unexpected situations, and is being developed toward complete autonomous surgery. Robot performs first realistic surgery without human help | Hub

Ophthalmic procedures are computer driven. Once the ophthalmologist selects the target, newer systems can perform the procedure autonomously. One system performed targeted retinal injections with 100% success and better accuracy than manual or surgeon-controlled robotic surgery. Autonomous robotic intraocular surgery for targeted retinal injections

I also find it hilarious that you think MedOnc will replace RadOnc and not the other way around. Feed a high-end AI model the history, pathology, imaging, molecular results, surgical findings, radiation options and guidelines, and it can produce a systemic therapy plan equal or superior to your MedOnc that you work with. A family physicain or medical student or pre-med or NP or PA can then order and administer it through Epic.

Even the MedOnc part of immunotherapy toxicity management can easily be automated by AI. Grade the toxicity, hold or permanently discontinue the drug, start steroids, plan the taper, reassess, escalate to second line immunosuppressant and determine if rechallenge is allowed. It is basically a guideline algorithm and checkpoint inhibitors are the same thing besically.

And when the toxicity becomes grade 3 or 4, especially grade 4, MedOnc don't manage the affected organ, the relevant specialist manage it the best thing they can do is write note of what happend in Epic and do what they are told by surgeons or relevant service. If AI selects the same treatment, the probability of toxicity is the same whether the person using it is a MedOnc, nurse or a janitor with AI access. That probability is not determined by planning or any special skills or knowledge from MedOnc.

Sorry to ruin the self-loathing party better luck next time 🙁
Click to expand...
So you’re saying dosimetrists are f****d
 
Beerus said:
Do you guys really have to be cynical and negative about everything? We constantly hear that AI will replace humans, but most of this hype is pushed by people who have zero understanding of either AI or healthcare. This development will probably have little to no impact on radiation oncologists. It is much more relevant to dosimetrists. Honestly, this forum becomes hysterical about everything, and every discussion seems to have the same conclusion “Radiation oncology is doomed” or “radiation oncology is bad.”

How many times have we heard that AI will replace radiologists or pathologists, the most obvious targets? and now radiology is the hottest field in medicine with the best lifestyle and salary. Radiation oncology is probably among the least threatened specialties because we physically deliver potentially lethal radiation. The dose is irreversible and unsalvageable. A surgical robot can be stopped, and some surgical errors can be repaired, but radiation? one extra digit or the wrong target or any tiny mistake and its game over.... no matter what you do.

Meanwhile, Johns Hopkins already has a robot that performed a lengthy phase of gallbladder surgery on a lifelike patient without human help, adapted to unexpected situations, and is being developed toward complete autonomous surgery. Robot performs first realistic surgery without human help | Hub

Ophthalmic procedures are computer driven. Once the ophthalmologist selects the target, newer systems can perform the procedure autonomously. One system performed targeted retinal injections with 100% success and better accuracy than manual or surgeon-controlled robotic surgery. Autonomous robotic intraocular surgery for targeted retinal injections

I also find it hilarious that you think MedOnc will replace RadOnc and not the other way around. Feed a high-end AI model the history, pathology, imaging, molecular results, surgical findings, radiation options and guidelines, and it can produce a systemic therapy plan equal or superior to your MedOnc that you work with. A family physicain or medical student or pre-med or NP or PA can then order and administer it through Epic.

Even the MedOnc part of immunotherapy toxicity management can easily be automated by AI. Grade the toxicity, hold or permanently discontinue the drug, start steroids, plan the taper, reassess, escalate to second line immunosuppressant and determine if rechallenge is allowed. It is basically a guideline algorithm and checkpoint inhibitors are the same thing besically.

And when the toxicity becomes grade 3 or 4, especially grade 4, MedOnc don't manage the affected organ, the relevant specialist manage it the best thing they can do is write note of what happend in Epic and do what they are told by surgeons or relevant service. If AI selects the same treatment, the probability of toxicity is the same whether the person using it is a MedOnc, nurse or a janitor with AI access. That probability is not determined by planning or any special skills or knowledge from MedOnc.

Sorry to ruin the self-loathing party better luck next time 🙁
Click to expand...
I read through the whole thread and I'm struggling to find the self-loathing party you're referring to?

If our jobs become easier, we can see more patients per physician. This is already the case. We used to see 200-225 patients a year and generate a great income. Now, we treat shorter courses, more patients - about 275-300 new patients a year is an estimate often used - and we earn about the same as we used to, but definitely less when accounting for inflaysh.

If we soon see 400-500 patients a year (I know people that do this and do it well, including members here) and its realized that the workload overall is less - i.e. it takes as much time/cognitive energy - then the reimbursement will catch up. If the overall numbers of patients doesn't go up and oligometastatic treatment doesn't grow to make up for the decreased numbers, of course things will change.

But, other work will fill the time. Radiopharm will become a greater part of our practice, perhaps more of us will handle ADT and other prostate cancer meds (I think I can handle ordering and monitoring effects of abiretarone), survivorship/longevity medicine.

I wholeheartedly disagree that AI will have little or no impact on radiation oncologists. The auto contouring of lymph node regions is now incredible - I review, make a few adjustments and move on. What used to take 15-20 minutes takes a few minutes, leaving me more time for other tasks. I am spending a little more time with tx planning, because my numbers aren't going up. I'm a 300 patient / year doc. Average 7 consults a week including vacations + F/U + weekly treatment can easily be handled in 4 days. Maybe 3.5. In 2011, this amount of work would have taken significantly longer.

It is an exciting time to be alive, particularly in health care. So much will change. Some of it will be good. Some of it won't. But to see it won't impact us at all - man, that is quite a take!
 
Advertisement - Members don't see this ad
Beerus said:
Do you guys really have to be cynical and negative about everything? We constantly hear that AI will replace humans, but most of this hype is pushed by people who have zero understanding of either AI or healthcare. This development will probably have little to no impact on radiation oncologists. It is much more relevant to dosimetrists. Honestly, this forum becomes hysterical about everything, and every discussion seems to have the same conclusion “Radiation oncology is doomed” or “radiation oncology is bad.”

How many times have we heard that AI will replace radiologists or pathologists, the most obvious targets? and now radiology is the hottest field in medicine with the best lifestyle and salary. Radiation oncology is probably among the least threatened specialties because we physically deliver potentially lethal radiation. The dose is irreversible and unsalvageable. A surgical robot can be stopped, and some surgical errors can be repaired, but radiation? one extra digit or the wrong target or any tiny mistake and its game over.... no matter what you do.

Meanwhile, Johns Hopkins already has a robot that performed a lengthy phase of gallbladder surgery on a lifelike patient without human help, adapted to unexpected situations, and is being developed toward complete autonomous surgery. Robot performs first realistic surgery without human help | Hub

Ophthalmic procedures are computer driven. Once the ophthalmologist selects the target, newer systems can perform the procedure autonomously. One system performed targeted retinal injections with 100% success and better accuracy than manual or surgeon-controlled robotic surgery. Autonomous robotic intraocular surgery for targeted retinal injections

I also find it hilarious that you think MedOnc will replace RadOnc and not the other way around. Feed a high-end AI model the history, pathology, imaging, molecular results, surgical findings, radiation options and guidelines, and it can produce a systemic therapy plan equal or superior to your MedOnc that you work with. A family physicain or medical student or pre-med or NP or PA can then order and administer it through Epic.

Even the MedOnc part of immunotherapy toxicity management can easily be automated by AI. Grade the toxicity, hold or permanently discontinue the drug, start steroids, plan the taper, reassess, escalate to second line immunosuppressant and determine if rechallenge is allowed. It is basically a guideline algorithm and checkpoint inhibitors are the same thing besically.

And when the toxicity becomes grade 3 or 4, especially grade 4, MedOnc don't manage the affected organ, the relevant specialist manage it the best thing they can do is write note of what happend in Epic and do what they are told by surgeons or relevant service. If AI selects the same treatment, the probability of toxicity is the same whether the person using it is a MedOnc, nurse or a janitor with AI access. That probability is not determined by planning or any special skills or knowledge from MedOnc.

Sorry to ruin the self-loathing party better luck next time 🙁
Click to expand...
No one is saying we will be replaced but given the under supply of med onc and structural oversupply of rad onc in the future (not my words, check the ASTRO workforce study)

Docs in general worry too much about job replacement from AI. We take liability. Therefore we will not be replaced
 
Beerus said:
Do you guys really have to be cynical and negative about everything?
Click to expand...
Click to expand...

Calm Down Al Pacino GIF

This is the essence of this forum.

TheWallnerus said:
So you’re saying dosimetrists are f****d
Click to expand...
I presume we may see one dosimetrist sitting in front of 3 workstations, tweaking AI-generated plans, instead of working one plan at a time (even it's done per template, which already offers alot of automation).
But this is nothing new. Think of how brokers operated in the 90s in stock market, and how they do it today.
 
Last edited:
Seems like an auspicious moment to drop that old saw “it’s tough to make predictions, especially about the future”

M

Thread 'CMS changes supervision rule. Rad Oncs no longer needed for daily operation of clinics. Med Students. Please read. You deserve to know implications.'

This is kind of buried in multiple threads, but this is a big deal.

CMS just proposed a rule change for supervision level of hospital outpatient therapeutic services from direct to general. This doesn't sound like a big deal, but it is actually a HUGE deal.

What this means functionally, is that a radiation oncologist will not need to be present to treat patients in hospital settings (i.e. the majority of jobs currently). This means a single Rad Onc can provide supervision for 2-3 hospitals simultaneously. It has been, that the vast majority of clinics have required a radiation oncologist...
  • Like
  • Angry
  • Sad
 
TheWallnerus said:
Seems like an auspicious moment to drop that old saw “it’s tough to make predictions, especially about the future”

M

Thread 'CMS changes supervision rule. Rad Oncs no longer needed for daily operation of clinics. Med Students. Please read. You deserve to know implications.'

This is kind of buried in multiple threads, but this is a big deal.

CMS just proposed a rule change for supervision level of hospital outpatient therapeutic services from direct to general. This doesn't sound like a big deal, but it is actually a HUGE deal.

What this means functionally, is that a radiation oncologist will not need to be present to treat patients in hospital settings (i.e. the majority of jobs currently). This means a single Rad Onc can provide supervision for 2-3 hospitals simultaneously. It has been, that the vast majority of clinics have required a radiation oncologist...
  • Like
  • Angry
  • Sad
Click to expand...
You're showing this as evidence that the doomerism predictions haven't manifested in the 7 years since?
 
CaesarRO said:
You're showing this as evidence that the doomerism predictions haven't manifested in the 7 years since?
Click to expand...
Yes. I can't remember any substantive NBF predictions this board (or me!) has made which have manifested. But as long as any of us shall continue to live I am sure we will not give up the prediction business.

Funny thing I saw on twitter today:

"No AI can replace a radiologist as fast as Ozempic & Co did Bariatric Surgeons."
 
You must log in or register to reply here.

Similar threads

W
ESTRO Annual Meeting vs ASTRO Annual Meeting
Replies
23
Views
2K
TheWallnerus
TheWallnerus
Palex80
ESMO 2026
Replies
1
Views
750
thesauce
thesauce
Y
  • Poll Poll
AI is Evil
Replies
28
Views
3K
temujim
T
Palex80
ASCO 2026
2
Replies
97
Views
8K
TheWallnerus
TheWallnerus
allseasons
Match rate 2026
Replies
25
Views
2K
WildRivers
W