AI is equal to or superior to radiologists in reading mammograms

[MDapplicant578124]

AI matches radiologist performance on mammo, DBT exams

A meta-analysis encompassing over 1.1 million screening breast imaging exams has concluded that artificial intelligence (AI) software performs equally to or better than radiologists, according to research published May 23 in Radiology.

www.auntminnie.com

Pretty significant meta analysis. I think the days of saying rads will never be replaced in any capacity by AI are over. Big changes will come in the next 3-4 decades when us med students are practicing

---

Members don't see this ad.

 

[Doctoscope]

Nice try, rads gunner.

 

[RangerBob]

I think rads is certainly higher risk than most other specialties. But more likely is we'll see physicians using AI to augment their work.

I'm not sure any company wants to assume the malpractice risk that comes with taking on making medical decisions. At least not yet. And that assumes they can even do it legally. As of right now you need to have a medical degree and license to practice medicine. I haven't seen any AI go to med school/get a license to do any of that. And the gov't has been very slow to even react to the detrimental effects of social media, which have been known for a while. So them getting together and saying an AI can make medical decisions (which the public likely overwhelmingly opposes, though I've seen no surveys on this), is a tall order. Realistically not happening for quite some time.

On the other hand, if I were an accountant, paralegal (or lawyer), and perhaps a good majority of US desk jobs, I'd be concerned about AI coming for my job.

They could replace politicians though. They would be more rational. Though the rational solution to climate change which threatens humanity is to limit our population and consumption. So that probably wouldn't go very well...

Realistically, either AI will be sorted out by the time it threatens physician jobs, or the apocalypse will have been here for some time already, since as I said, we're not the low-hanging fruit for AI.

 

[MateoGM416]

Realistically not happening for quite some time.

We said the same thing about computers, the internet, and mobile phones. As humans have we ever predicted accurately the rate of ascent of any form of technology, ever?

 

[Dantes]

We said the same thing about computers, the internet, and mobile phones. As humans have we ever predicted accurately the rate of ascent of any form of technology, ever?

I was promised jetpacks and flying cars by now so I'm still salty about that

 

[operaman]

Yeah the way things are done, until you can bill for an AI only read, I think jobs are safe. They will definitely use AI though to be more efficient and improve accuracy. It’ll be a hyper accurate version of the automatic EKG read that’s right a lot of the time, but still gets a human over read.

Biggest things I can think of offhand:

1) error correction. An extra set of eyes is always nice.

2) AI dictated notes that can be edited by the overreading radiologist.

3) EMR integrated AI that scans the whole chart and gives a one line summary of the patient and the key clinical questions for the scan itself. Pre dictated template above also addresses these directly as well.

ChatGPT can already do the second two even in its current iteration.

 

[ChordaEpiphany]

Every time someone posts something like this, people flock into the thread and say, "there's so much more to rads than reading an image" or "liability will keep doctors employed." We need to drop that attitude not just for rads/path but for every intellectual specialty. We have some of the biggest targets on our backs, and the profession needs to react soon.

People love to say, "if they automate doctors, they'll have automated everyone else first." The problem is that medicine is also evidence-based and extraordinarily well-documented. Other white collar jobs are less complex, but they are also more diverse and poorly defined (so it's hard to train an AI specific to the task). Further, these workers are more flexible, which is why we've maintained a ~94% employment rate through a century of earth-shattering automation/globalization. However, doctors have few transferable skills of equal value. Our demand is government-controlled and supply/exclusive rights to practice is our only leverage. Quality of care, patient satisfaction, and outcomes don't actually dictate physician pay. With AI, the capacity to supply care will increase. Since reimbursement is effectively government controlled, you need to convince the government to improve healthcare. We need to make sure it's known that AI + doctor is better and necessary. Equivalency cannot be an acceptable alternative to the highest quality of care.

There are also protective factors, which is why catastrophic predictions are wrong. You still need doctors to collect novel data (HPI), perform a physical exam, do procedures, and guide care towards patient needs/desires. Finally, changes in healthcare are extremely slow, so you might have 10-15 years to prep before you're made totally redundant.

That said, AI is very real and very much coming for physicians. We need to redefine our value, and in a world where AI performs most intellectual labor, data is the valuable thing we bring to the table. Top AI companies will be clamoring for the most updated and thorough data sets to secure contracts with hospital systems. Currently, we give our data away for free. We need to consider:

1) Reimbursement to physicians for the generation of data used to train AI models. A standard part of the physician's contract should be that if a health system sells EMR data to a company, the physician gets a piece of that.

2) Licensing fees for published advancements incorporated into AI models.

3) Unified efforts to ensure physicians do not cannibalize each other by selling data for cheap.

This is exactly why we created copyright laws and intellectual property when it became extremely easy to reproduce novel ideas. Why would anyone create a new product if someone else can just copy it and sell it the same as you? Why would anyone create new data for AI models if some company is just going to observe your actions, have an AI program learn from and copy them, and cut you out of the equation? Data is the new labor.

 

[MateoGM416]

1) Reimbursement to physicians for the generation of data used to train AI models. A standard part of the physician's contract should be that if a health system sells EMR data to a company, the physician gets a piece of that.

2) Licensing fees for published advancements incorporated into AI models.

3) Unified efforts to ensure physicians do not cannibalize each other by selling data for cheap.

This is exactly why we created copyright laws and intellectual property when it became extremely easy to reproduce novel ideas. Why would anyone create a new product if someone else can just copy it and sell it the same as you? Why would anyone create new data for AI models if some company is just going to observe your actions, have an AI program learn from and copy them, and cut you out of the equation? Data is the new labor.

I agree with you, but then would patients not be entitled to the money being made by the health systems and physicians from their own data? Why would health systems ever willingly sign up to split those profits even to physicians

 

[voxveritatisetlucis]

The whole “augment their work and make them more efficient” argument is pretty dumb. If a radiologist + AI can do the work of 5 radiologists, then it would still destroy the market for all intents and purposes due to extreme competition for limited jobs

 

[ChordaEpiphany]

I agree with you, but then would patients not be entitled to the money being made by the health systems and physicians from their own data? Why would health systems ever willingly sign up to split those profits even to physicians

Who's entitled to what is a function of leverage. Some patients are paid for their data. Some studies just hand a consent form over to patients and ask if it's okay if their biopsy is included in some clinical research. Those patients have no leverage because if they say no, the researchers can just wait for the next person to agree. Other studies require physical presence 5-6 times during work hours. Those patients get paid for showing up because patients who can do that are somewhat scarce.

How much leverage physicians have will be a function of how much data and what data is needed to make the best models. It will also be a function of overall education on the topic and efforts to organize to ensure we don't cannibalize each other.

 

[MDMechEngr]

I would not want to board any aircraft that does not have a human pilot in the cockpit.

 

[voxveritatisetlucis]

I would not want to board any aircraft that does not have a human pilot in the cockpit.

Yes but patients don’t care if they are seen by somebody with 10 years of training versus somebody who did a 1 year online degree. So why would they care about AI vs human

 

[MDMechEngr]

I said human pilot and not a flight attendant with some flight training (NP).

 

[operaman]

I would not want to board any aircraft that does not have a human pilot in the cockpit.

Thankfully the pilot is by far the least expensive part of hourly costs for a commercial aircraft!

 

[Esmolol]

Wonder when AI will figure out that fudging diagnoses and withholding interpretation is a nice and good way to get rid of mankind...
#SKYNET4EVER !!!

 

[Cornfed101]

I would love it if AI removed the need to chart or greatly reduced it. I already had chatgpt help me write my ERAS PS in 30 min and it’s not half bad

 

[Rads125125]

Several of my coresidents in my DR program have already started talking about potentially switching specialties. It seems like people are starting to take this seriously. Kinda crazy how just 3 months ago, DR was talked about as this super competitive specialty, and now people are starting to drop out of it. I feel bad for the M4s who just matched DR a few months ago

 

[Dantes]

Several of my coresidents in my DR program have already started talking about potentially switching specialties. It seems like people are starting to take this seriously. Kinda crazy how just 3 months ago, DR was talked about as this super competitive specialty, and now people are starting to drop out of it. I feel bad for the M4s who just matched DR a few months ago

This reads like an M4 trying to make the upcoming rads cycle less competitive by any means necessary

 

[Esmolol]

This reads like an M4 trying to make the upcoming rads cycle less competitive by any means necessary

Guys! Just keep in mind - Anesthesia is already almost all machines. #Gunner4Life

 

[Anakinmemer]

This reads like an M4 trying to make the upcoming rads cycle less competitive by any means necessary

It's certainly working, I know of 2 classmates not doing radiology solely due to AI fears. One switched to IR early on and another to EM or anesthesia.

I fully expect rads apps to slow down their growth rate or even take a dip this upcoming cycle from the combination of people now recognizing it as being competitive and scared to apply with lower strength apps, as well as fear of AI

 

[Esmolol]

another to EM or anesthesia.

You mean EM, right?

 

[VA Hopeful Dr]

We said the same thing about computers, the internet, and mobile phones. As humans have we ever predicted accurately the rate of ascent of any form of technology, ever?

My family got our first computer in 1996. The hospital where I did residency didn't get rid of paper charts until around 2015 or so.

Medicine is always slow to adopt new tech as it's a fairly conservative field.

It's also a heavily government regulated field which slows down big changes as well.

 

[ChordaEpiphany]

Guys! Just keep in mind - Anesthesia is already almost all machines. #Gunner4Life

Sometimes automation really does just make the job easier. Pilots and anesthesiologists are basically 90% there, but they've retained their employment and wages. The protective factors are liability and the inability to scale these jobs. You still need at least one person in the OR and at least one person in the cockpit. Even with some ORs staffing 1:5 MD:CRNA, anesthesiologists haven't lost any comp since 2011 ($325K vs. $448K, which is $447K vs. $448K inflation adjusted).

Where you'll run into the most trouble with AI is scalable specialties where patients don't really care if they see a doctor. You can easily imagine a world where basic outpatient visits are completed on an app with an AI that sends its recommendations over to a single MD (or degree mill NP) who rubber stamps hundreds of prescriptions per day.

However, it's important to point out that it's much easier to predict which current income streams will vanish than it is to imagine the new income streams that will emerge. The prospect of running your own private practice on main street and pulling in a top 1% income while doing the same thing for 40 years is obviously dead, but the value of highly intelligent, knowledgeable, sociable, and savvy individuals isn't changing any time soon.

 

[Vyzzle]

If I were a radiologist I would put a big chunk of money into publicly traded companies that are using AI for imaging and or other healthcare applications. Worse case, they lose their job but have a 50x investment

Or you hedge wrong, bet on the wrong companies, get replaced AND have your portfolio tank

 

[voxveritatisetlucis]

Or you hedge wrong, bet on the wrong companies, get replaced AND have your portfolio tank

 

[Anakinmemer]

Sometimes automation really does just make the job easier. Pilots and anesthesiologists are basically 90% there, but they've retained their employment and wages. The protective factors are liability and the inability to scale these jobs. You still need at least one person in the OR and at least one person in the cockpit. Even with some ORs staffing 1:5 MD:CRNA, anesthesiologists haven't lost any comp since 2011 ($325K vs. $448K, which is $447K vs. $448K inflation adjusted).

Where you'll run into the most trouble with AI is scalable specialties where patients don't really care if they see a doctor. You can easily imagine a world where basic outpatient visits are completed on an app with an AI that sends its recommendations over to a single MD (or degree mill NP) who rubber stamps hundreds of prescriptions per day.

However, it's important to point out that it's much easier to predict which current income streams will vanish than it is to imagine the new income streams that will emerge. The prospect of running your own private practice on main street and pulling in a top 1% income while doing the same thing for 40 years is obviously dead, but the value of highly intelligent, knowledgeable, sociable, and savvy individuals isn't changing any time soon.

I have the MGMA 2010 report and it has anesthesia at 419k and rads at 515k, inflation adjusted that's 602,000 and 739,000 respectively. Anesthesia and rads are nowhere near those inflation adjusted averages today and are actually among the worst performing since 2010. It might be hard to believe because everyone talks about how "hot" the job markets are, but the data is there (MGMA sample sizes are very good). Surgical subs and procedural IM specialties have kept their earnings up much better.

People don't realize that rads used to be among the highest earners in all of medicine (rads average was just 10-20k below ortho and interventional cards...). I think rads would be derm level competitive if it kept up with IC and ortho

 

[ChordaEpiphany]

I have the MGMA 2010 report and it has anesthesia at 419k and rads at 515k, inflation adjusted that's 602,000 and 739,000 respectively. Anesthesia and rads are nowhere near those inflation adjusted averages today and are actually among the worst performing since 2010. It might be hard to believe because everyone talks about how "hot" the job markets are, but the data is there (MGMA sample sizes are very good). Surgical subs and procedural IM specialties have kept their earnings up much better.

People don't realize that rads used to be among the highest earners in all of medicine (rads average was just 10-20k below ortho and interventional cards...). I think rads would be derm level competitive if it kept up with IC and ortho

I used Medscape from 2011 vs. Medscape 2023. I know Medscape tends to underestimate, but at least it's the same source and should be internally comparable. There's also some serious variance in the data sets, so it would be nice to see this year's MGMA to compare. Also, rads hasn't been affected in any way by AI yet, so any loss of comp has more to do with fee schedules, PE, and the death of private practice. I'm honestly not sure what the last 10-15 years of rads compensation has to do with this discussion, but if you use Medscape data again it shows a different story ($360K in 2011 translates to $505K, and today Medscape reports $483K).

I'm going to take this opportunity to make a very important point. In 2011 a lot of people on these forums were claiming the sky was falling for any number of reasons, be it mid-level creep (CRNAs), the ACA, or whatever. 12 years later while practicing medicine looks almost nothing like it did in 2011, the sky hasn't fallen and most specialties have kept up with inflation (or at least kept up with employee wages in general). 12 years in the future medicine will look nothing like it does today. Memorizing facts will seem even stupider than it does today, and you'll be practicing in a completely different way, likely with tons of virtual interaction, AI-interfacing, collaboration with other non-physician healthcare professionals, and probably a whole host of non-clinical tasks. Keep a finger to the wind and adjust as necessary. Doctors will be fine, even if the profession of doctoring becomes unrecognizable.

Edit: This thread confirms that 2010 MGMA numbers looked very inflated to a bunch of SDNers back then

 

[Anakinmemer]

I used Medscape from 2011 vs. Medscape 2023. I know Medscape tends to underestimate, but at least it's the same source and should be internally comparable. There's also some serious variance in the data sets, so it would be nice to see this year's MGMA to compare. Also, rads hasn't been affected in any way by AI yet, so any loss of comp has more to do with fee schedules, PE, and the death of private practice. I'm honestly not sure what the last 10-15 years of rads compensation has to do with this discussion, but if you use Medscape data again it shows a different story ($360K in 2011 translates to $505K, and today Medscape reports $483K).

I'm going to take this opportunity to make a very important point. In 2011 a lot of people on these forums were claiming the sky was falling for any number of reasons, be it mid-level creep (CRNAs), the ACA, or whatever. 12 years later while practicing medicine looks almost nothing like it did in 2011, the sky hasn't fallen and most specialties have kept up with inflation (or at least kept up with employee wages in general). 12 years in the future medicine will look nothing like it does today. Memorizing facts will seem even stupider than it does today, and you'll be practicing in a completely different way, likely with tons of virtual interaction, AI-interfacing, collaboration with other non-physician healthcare professionals, and probably a whole host of non-clinical tasks. Keep a finger to the wind and adjust as necessary. Doctors will be fine, even if the profession of doctoring becomes unrecognizable.

Edit: This thread confirms that 2010 MGMA numbers looked very inflated to a bunch of SDNers back then

I wasn't commenting on AI I was just saying your statement that physicians haven't lost comp is not accurate. Medscape's sample size is about 10% that of MGMA's and puts far less effort into obtaining complete data. There's a reason MGMA costs 1k to get, it's extremely comprehensive and includes 170,000 providers from all practice types, not just a few who respond to a medscape email. I don't know any recruiter that says medscape is a good source or used in any way in contract negotiations. The most recent MGMA's I could compare to are 2019, 2020, and 2021, unfortunately do not have the latest. In all these radiology averages from 500-560k, so I can't imagine they'll jump to 740k in the newest report...

Your thread has anecdotal comments from a few people who might not even be doctors (with others saying it is in fact accurate) so I don't think that's a great source personally.

I wish it wasn't the case but the sad reality is that when adjusted for inflation, many physicians don't make as much as they did 15 years ago, sometimes by a wide margin. I guess my point is any innovation that speeds up reading imaging will end up being overcompensated for in cuts, which is exactly what's happened every year.

 

[deleted1139416]

I used Medscape from 2011 vs. Medscape 2023. I know Medscape tends to underestimate, but at least it's the same source and should be internally comparable. There's also some serious variance in the data sets, so it would be nice to see this year's MGMA to compare. Also, rads hasn't been affected in any way by AI yet, so any loss of comp has more to do with fee schedules, PE, and the death of private practice. I'm honestly not sure what the last 10-15 years of rads compensation has to do with this discussion, but if you use Medscape data again it shows a different story ($360K in 2011 translates to $505K, and today Medscape reports $483K).

I'm going to take this opportunity to make a very important point. In 2011 a lot of people on these forums were claiming the sky was falling for any number of reasons, be it mid-level creep (CRNAs), the ACA, or whatever. 12 years later while practicing medicine looks almost nothing like it did in 2011, the sky hasn't fallen and most specialties have kept up with inflation (or at least kept up with employee wages in general). 12 years in the future medicine will look nothing like it does today. Memorizing facts will seem even stupider than it does today, and you'll be practicing in a completely different way, likely with tons of virtual interaction, AI-interfacing, collaboration with other non-physician healthcare professionals, and probably a whole host of non-clinical tasks. Keep a finger to the wind and adjust as necessary. Doctors will be fine, even if the profession of doctoring becomes unrecognizable.

Edit: This thread confirms that 2010 MGMA numbers looked very inflated to a bunch of SDNers back then

What about accounting for number of hours worked?

 

[Kierkegaard's Bud]

Or you hedge wrong, bet on the wrong companies, get replaced AND have your portfolio tank

There will be in-fighting, massive lawsuits about who owns the data that these models were trained on as soon as some real money is to be made.

 

[Cognovi]

AI matches radiologist performance on mammo, DBT exams

A meta-analysis encompassing over 1.1 million screening breast imaging exams has concluded that artificial intelligence (AI) software performs equally to or better than radiologists, according to research published May 23 in Radiology.

www.auntminnie.com

Pretty significant meta analysis. I think the days of saying rads will never be replaced in any capacity by AI are over. Big changes will come in the next 3-4 decades when us med students are practicing

Radiology is a technology-heavy field that changes a lot over the course of a career. The oldest attendings working now trained in the days before CT, MRI, and PET, before digital film, before speech recognition software, before gigabit internet. Can you imagine the change in the work of a radiologist, and the increased accuracy and thus importance of radiology in general, that has taken place in that career? They could give up the angiography catheter and lead suits for a more comfortable living as a desk jockey working from home, while continually learning new anatomy, new physics, new diseases. The changes that will come to radiology in the next 30-40 years will be mind bottling. As long as you are willing to adapt, you won't go extinct. There are some dinosaurs out there who only read plain films, but their days are numbered. You have to plan to be a radiologist who can master the AI and use it as a tool, in combination with your experience and clinical acumen.

 

[GadRads]

We utilize AI in our radiology work flow, and it is very good at identifying head bleeds, cervical spine fractures, and PEs. I would say the miss rate is roughly 2-4% and the false positive rate ca. 2%. I still double check.

What AI will do to radiology in 10-20 years is anyone’s guess, but I am not concerned. We do more than detect abnormalities, and someone has to assume liability.

 

[voxveritatisetlucis]

We utilize AI in our radiology work flow, and it is very good at identifying head bleeds, cervical spine fractures, and PEs. I would say the miss rate is roughly 2-4% and the false positive rate ca. 2%. I still double check.

What AI will do to radiology in 10-20 years is anyone’s guess, but I am not concerned. We do more than detect abnormalities, and someone has to assume liability.

I think that it’s more a problem of one person or corporation being able to assume liability for a much higher volume of cases therefore creating a tight labour market and in turn lower wages

 

[ChordaEpiphany]

What about accounting for number of hours worked?

I can't say exactly, but the trend is towards fewer working hours (this study shows a 7.6% decrease from 2001 to 2021). However, I think most practicing physicians will say that they are seeing more patients on a daily basis and that reimbursements per patient are lower. Has technology sped up the rate of seeing patients, or are we just squeezing 60 hours of work into 49 hours these days?

We utilize AI in our radiology work flow, and it is very good at identifying head bleeds, cervical spine fractures, and PEs. I would say the miss rate is roughly 2-4% and the false positive rate ca. 2%. I still double check.

What AI will do to radiology in 10-20 years is anyone’s guess, but I am not concerned. We do more than detect abnormalities, and someone has to assume liability.

I think the question is less, "what will AI be able to do?" and more, "what is the comfort level of regulators, payers, and employers wrt AI autonomy?"

We've already seen regulators trusting APPs with autonomy despite markedly poorer training and no convincing evidence of equivalency or safety compared to physicians. Assuming liability will represent far less value as the technology becomes better. When that 2-4% miss rate and 2% false positive rate becomes 0.1% and 0.1%, and AI can read far more pathologies than a specialized radiologist, I doubt the MBA writing checks to radiologists every month is going cling to liability as a reason to keep everyone employed. More likely radiologists will have to be more-or-less okay with rubber-stamping AI reads at multiple times their current pace and absorb the extra liability, because the productivity demands coming from admins will be too high, and someone will be willing to do it. The job will become as much about knowing when to trust AI blindly as it will be about making accurate reads with your own eyes, and the result will be a significant increase in radiology productivity (and decreased demand for radiologist's labor hours). This is all without considering the possibility of lawmakers greenlighting autonomous AI reads.

Predicting the next 30 years is impossible. I really don't see a future where radiologists are obsolete in our lifetimes. I do foresee a future 10-15 years from now where the market is tough on both new grads and old-timers who don't adapt with the technology.

 

[GadRads]

I can't say exactly, but the trend is towards fewer working hours (this study shows a 7.6% decrease from 2001 to 2021). However, I think most practicing physicians will say that they are seeing more patients on a daily basis and that reimbursements per patient are lower. Has technology sped up the rate of seeing patients, or are we just squeezing 60 hours of work into 49 hours these days?

I think the question is less, "what will AI be able to do?" and more, "what is the comfort level of regulators, payers, and employers wrt AI autonomy?"

We've already seen regulators trusting APPs with autonomy despite markedly poorer training and no convincing evidence of equivalency or safety compared to physicians. Assuming liability will represent far less value as the technology becomes better. When that 2-4% miss rate and 2% false positive rate becomes 0.1% and 0.1%, and AI can read far more pathologies than a specialized radiologist, I doubt the MBA writing checks to radiologists every month is going cling to liability as a reason to keep everyone employed. More likely radiologists will have to be more-or-less okay with rubber-stamping AI reads at multiple times their current pace and absorb the extra liability, because the productivity demands coming from admins will be too high, and someone will be willing to do it. The job will become as much about knowing when to trust AI blindly as it will be about making accurate reads with your own eyes, and the result will be a significant increase in radiology productivity (and decreased demand for radiologist's labor hours). This is all without considering the possibility of lawmakers greenlighting autonomous AI reads.

Predicting the next 30 years is impossible. I really don't see a future where radiologists are obsolete in our lifetimes. I do foresee a future 10-15 years from now where the market is tough on both new grads and old-timers who don't adapt with the technology.

Those numbers I gave you are for appropriately scanned patients. When patients are not scanned correctly and there are artifacts, the AI software fails massively. Many times the system doesn’t even bother to spit out a finding. If I include these studies the system is at best 80% accurate for the very very few pathologies the system can assess. A radiologist with more than a 20% significant miss rate will lose his/her license in due time.

Practicing radiologists are not concerned about AI. I can assure you that. It might make us more productive but we aren’t even close to that reality.

 

[deleted1139416]

It might make us more productive but we aren’t even close to that reality.

What about the acceleration of AI development? The speed at which it has been developed for the last 10 years is likely a lot lower than for the next 10 years, so we might think a certain level of AI success might be achieved by 2050 but continued acceleration might make it possible by 2035.

 

[GadRads]

What about the acceleration of AI development? The speed at which it has been developed for the last 10 years is likely a lot lower than for the next 10 years, so we might think a certain level of AI success might be achieved by 2050 but continued acceleration might make it possible by 2035.

Geoffrey Hinton thought radiologists would be obsolete by now, but we are still here. If AI is able to replace radiology, then most jobs aren't safe. The best possibility is synergy between radiology and AI. Cardiologists are still over-reading EKGs. Imaging is more complex.

 

[Cognovi]

Nobody can predict the future. In 20 years, maybe radiology will consist of me sitting in my vacation cabin wearing my Apple Vision Pro to experience the images and sign off on AI-drafted reports.