What’s the worst thing about radiology?

[Steve_Zissou]

---

Members do not see this ad.

...yet

the fact is that current and future trainees need at least 20-30s to re-coup their med school investment. that's a long, long time for AI to progress. you can't deny that there are fields (clinical and procedural) much safer than rads.

my advice to med students - proceed with caution

Which clinical fields are much safer, in your opinion?

 

[SeisK]

...yet

the fact is that current and future trainees need at least 20-30s to re-coup their med school investment. that's a long, long time for AI to progress. you can't deny that there are fields (clinical and procedural) much safer than rads.

my advice to med students - proceed with caution

Alternatively, pursue what you’re passionate about as even in a much improved form from its current state CNNs are not an existential threat to radiology, for the reasons outlined in detail above.

 

[CallMeBigJeff]

...yet

the fact is that current and future trainees need at least 20-30s to re-coup their med school investment. that's a long, long time for AI to progress. you can't deny that there are fields (clinical and procedural) much safer than rads.

my advice to med students - proceed with caution

My problem with the "well what if!" advice is that if someone who was genuinely interested in Radiology (e.g. myself) were to heed your advice and instead choose the 2nd specialty they are most interested in, what happens in 20-30-40 years when AI hasn't replaced Radiologists and people are very happily practicing Rads? I cannot fathom the type of regret that would cause a person, at least speaking personally. Even only considering that factor, I think taking this sensible "risk" of pursuing a career in Radiology is a considerably better choice than picking a completely different career due to the risk of us not having a job that so many Non-Rads and Non-AI experts constantly propose.

 

[scoopdaboop]

Step 1. Fire 3/4 of radiologists
Step 2. Have the ones left oversee 4 mid levels with AI
Step 3. Profit???

This makes 0 sense. How does a radiologist whose worried about missing something get affected by AI which let’s say completes read quickly with 99 percent specificity. Over the 100s and 1000s of reads being done no radiologist or hospital would take a chance and not read something like they normally would. Hence, even with AI giving a preliminary read, the actual read will take the same amount of time for fear of mistake.

Secondly what benefit do midlevels have in this situation...you really believe they will be able to catch something AI might have missed...
Lastly, radiology is a field that’s in high demand right now (based upon current job market etc). I don’t see how nationwide radiology workforce can even be cut by 1/3 even with AI because you’re never not going to have a radiologist working.

 

[SeisK]

This makes 0 sense. How does a radiologist whose worried about missing something get affected by AI which let’s say completes read quickly with 99 percent specificity. Over the 100s and 1000s of reads being done no radiologist or hospital would take a chance and not read something like they normally would. Hence, even with AI giving a preliminary read, the actual read will take the same amount of time for fear of mistake.

Secondly what benefit do midlevels have in this situation...you really believe they will be able to catch something AI might have missed...
Lastly, radiology is a field that’s in high demand right now (based upon current job market etc). I don’t see how nationwide radiology workforce can even be cut by 1/3 even with AI because you’re never not going to have a radiologist working.

Love the enthusiasm. Pretty sure he was sarcastic though.

 

[deleted746658]

My problem with the "well what if!" advice is that if someone who was genuinely interested in Radiology (e.g. myself) were to heed your advice and instead choose the 2nd specialty they are most interested in, what happens in 20-30-40 years when AI hasn't replaced Radiologists and people are very happily practicing Rads? I cannot fathom the type of regret that would cause a person, at least speaking personally. Even only considering that factor, I think taking this sensible "risk" of pursuing a career in Radiology is a considerably better choice than picking a completely different career due to the risk of us not having a job that so many Non-Rads and Non-AI experts constantly propose.

just gonna leave this here...
https://pubs.rsna.org/doi/10.1148/radiol.2020190283

 

[Cognovi]

just gonna leave this here...
https://pubs.rsna.org/doi/10.1148/radiol.2020190283

Oh wow, test sets of 5 patients, training sets as low as 0 patients.

 

[Cognovi]

Oh wow, test sets of 5 patients, training sets as low as 0 patients.

I read it a bit more in depth. They used deep learning for lesion segmentation, and then processed the lesions with feature selection tools. The centerpiece is an inference engine, in which an expert hard codes the probabilities based on experience and the literature. The input boils down to like 23 parameters (eg, age, size, location, enhancement) and it ranks a differential diagnosis from the 18 options. That why you don't really need training data.

This is an expert system.

It's neat to see how much the residents sucked. It'd be interesting to see if a resident equipped with the conditional probability spreadsheet of the inference engine could improve their performance.

 

[CallMeBigJeff]

just gonna leave this here...
https://pubs.rsna.org/doi/10.1148/radiol.2020190283

Did you read the paper or just the title?

"Radiologists were more accurate at diagnosing common versus rare diagnoses (78% vs 47% across all radiologists; P < .001) " This includes residents.
It was NOT better than Neurorads attendings (P value of 0.20, wow!). And, like the poster above me mentioned, check out that paper's very unrealistic generalizability for patient populations. Stop fear mongering because it genuinely does turn off other people that would've potentially been interested in Rads.

 

[Kuratz]

I would rather do Rads and be out of a physician job in 10 years than deal with the migraine-inducing nonsense of IM or the sacrificial lifestyle of surgery. Medicine as a whole just isn’t all that it’s cracked up to be and to choose something you don’t like, such as IM (as you are @cosine89) out of fear of AI is a total recipe for burnout. If AI takes my job in 2030 so be it. We’re ambitious individuals, and sometimes an unexpected push can yield success in unsuspecting places.

 

[Steve_Zissou]

just gonna leave this here...
https://pubs.rsna.org/doi/10.1148/radiol.2020190283

I’m still curious, what specialties are better protected from outside forces than Radiology?

 

[MouseChair]

AI is still a threat in the distant future. Midlevel creep is already happening NOW.

 

[Tiger100]

i will offer a contrasting opinion to the above: only pursue radiology if you cannot see yourself happy in any other field

the world is changing fast. and you are a second year medical student. you are nearly a decade removed from becoming a practicing radiologist.

i think radiologists are inherently defensive when it comes to the threat of AI to their field. the reality is that it is a very real threat to disrupt the field. other fields are not immune to AI, but those same fields are definitely more resistant

for example, i think COVID-19 is running a nice simulation of what is to come in regards to how decreased volume will disrupt the rads workforce. let me elaborate: a good portion - maybe most? - of radiology studies ordered are NORMAL. when (not if) the big players in AI (google, microsoft, facebook, etc backed by unlimited $$) automate NORMAL reads, the amount of imaging that needs to be reviewed by human eyes drops significantly and the efficiency of a single radiologist increases exponentially. this leaves less and less work to be done by human radiologists, and thus will lead to a massive oversupply of rads

what happens to the rest of rads workforce? do they re-train? re-enter a new field altogether? idk. but i do know that most of the younger rads will get screwed as they will have graduated with immense loans but no big paychecks to pay them off. imo, i'm not sure its worth the risk.

One of the most difficult tasks in radiology is differentiating normal from abnormal. It will be one of the most difficult things for AI to do it.

Also most scans have "some findings:, may not be significant. But TRUE NORMAL scans are not common, unlike what you said.

 

[CyrusHabs7695]

I would rather do Rads and be out of a physician job in 10 years than deal with the migraine-inducing nonsense of IM or the sacrificial lifestyle of surgery. Medicine as a whole just isn’t all that it’s cracked up to be and to choose something you don’t like, such as IM (as you are @cosine89) out of fear of AI is a total recipe for burnout. If AI takes my job in 2030 so be it. We’re ambitious individuals, and sometimes an unexpected push can yield success in unsuspecting places.

Bro this is like a yearbook quality quote, about to get this tattoo'd on me. absolute chills

 

[SeisK]

I’m still curious, what specialties are better protected from outside forces than Radiology?

Is hospital administration a specialty?

 

[Steve_Zissou]

Is hospital administration a specialty?

Only if you have BSN, MSN, RN, APRN, NP after your name. Gotta have all the right credentials.

 

[Tiger100]

just gonna leave this here...
https://pubs.rsna.org/doi/10.1148/radiol.2020190283

This is the interesting part:

Although the AI system produced errors, these were different from and perhaps complementary to those produced by humans. The types of errors addressed by this type of AI algorithm include distinctly human biases such as satisfaction of search (21) and heavily skewed probability calculations (22). With imaging demand increasing steadily and outpacing the capacity of radiologists (23), methodologies that simultaneously improve efficiency (24,25) and reduce errors should be a welcome addition to the radiologist’s repertoire.


As a result, you can predict that AI + Radiologist will do better than AI or Radiologist itself. In other words, it seems AI makes some mistakes that are obvious to a human interpreter.

 

[Cognovi]

This is the interesting part:

Although the AI system produced errors, these were different from and perhaps complementary to those produced by humans. The types of errors addressed by this type of AI algorithm include distinctly human biases such as satisfaction of search (21) and heavily skewed probability calculations (22). With imaging demand increasing steadily and outpacing the capacity of radiologists (23), methodologies that simultaneously improve efficiency (24,25) and reduce errors should be a welcome addition to the radiologist’s repertoire.


As a result, you can predict that AI + Radiologist will do better than AI or Radiologist itself. In other words, it seems AI makes some mistakes that are obvious to a human interpreter.

I don't see how this system addresses satisfaction of search if all the cases have only one diagnosis.

 

[Clean7795]

AI is still a threat in the distant future. Midlevel creep is already happening NOW.

Agreed. Primary care has the greatest threat of disruption, with cases of physicians literally being fired and replaced by midlevels. The fact that midlevel proliferation is continuing to explode shows that there are clear areas where NPs can divide up, take over, and practice safely (enough) without direct physician supervision. The threat to primary care is twofold with both AI and midlevels being threats: What happens if someone develops AI that shows NP+AI is noninferior to a physician in diabetes/hypertension management?

Diagnostic radiology by nature cannot be easily divided up into "easy" cases and "hard" cases, nor can cases be partitioned up such that a radiologist only does half the work for a case while a midlevel/AI does the rest. Every case has a potential fatal abnormality. Some of these fatal abnormalities resemble anatomic aberrations. Every case must be fully ran through in the radiologist's mind. Many of these variants are rare enough that we cannot build good training sets for them.

This means we cannot rely on AI to call normals or even save time. Radiology in comparison to primary care is very safe from AI and midlevels. Where AI has great potential is in flagging important cases that are truly STAT (e.g. high sensitivity detection of 1+ acute processes), improving speed of advanced imaging reconstructions, improving hanging protocols, research, and other applications that will help future radiologists.

 

[neriticzone]

high sensitivity detection

just to be pedantic, sensitivity alone can be worse than no program at all.

An anecdote a neuro attending gave us was when he had to cancel a clinical study on stroke screening early when the average number of code strokes per 24 hours went from 10ish to 40-50.

I agree with you tho, I think there is potentially/probably a place for these algos in rads, but we still don’t know enough to say much with certainty beyond that.

 

[DOFOSHO]

i will offer a contrasting opinion to the above: only pursue radiology if you cannot see yourself happy in any other field

nice simulation of what is to come in regards to how decreased volume will disrupt the rads workforce. let me elaborate: a good portion - maybe most? - of radiology studies ordered are NORMAL. when (not if) the big players in AI (google, microsoft, facebook, etc backed by unlimited $$) automate NORMAL reads, the amount of imaging that needs to be reviewed by human eyes drops siggraduated with immense loans but no big paychecks to pay them off. imo, i'm not sure its worth the risk.

It takes more skill to be able to identify normal than it does to identify abnormal. In my experience i've been called in the reading room and asked "can you make sure this is normal" way more times than "can you make sure this is abnormal". For as long as a radiologist has higher specificity WITHOUT even looking at the study AI will continue to remain a joke.

 

[Aliboo]

Yeah you normally have to pay several hundred dollars for the MGMA access

PM me your email and I can send you that

Hey doll you also be so kind as to send me that data. I'm not rads, but currently considering offers right after residency. Have no way of knowing if what is being offered is fair. If you can, thanks and if not thanks anyway!

 

[90sportsfan]

I'm strongly considering specializing in Radiology (hopefully). I admit that I have been a little apprehensive about AI, but like many have mentioned, I think it would be an adjunct and not a complete replacement for radiologists. Someone still has to write the report and to do that, you'd need to review images before you attach your name to the report (and the liability that comes with that). So I don't see how it would replace radiologists (no other specialists will just attach their names to report without reviewing the image- and radiologists are the specialists qualified to review images in detail).

Even if AI were to become that advance, it would take decades and decades to perfect, and science/the body (as we are seeing by this pandemic) can always throw a wrench in things, which could cause the need for a reassess the technology. Basically, at least not during our careers, could I see AI getting to the point where it replaces radiologists. Maybe becomes a useful adjunct to help with efficiency, but not a replacement.