[Workaccount2]

Just a note on the radiologist part, the current SOTA radiology AI is still tiny parameter CNN's from the mid-late 2010's running locally. NYT ran an article a few weeks about this, and the entire article uses the phrase "A.I.", which people assume means ChatGPT, but really can refer to anything in the last 60 years of A.I. research. Manually digging revealed it was an old architecture.

We don't know yet how a modern transformer trained on radiology would perform, but it almost certainly would be dramatically better.

[demosthanos]

> We don't know yet how a modern transformer trained on radiology would perform, but it almost certainly would be dramatically better.

Why? Is there something about radiology that makes the transformer architecture appropriate?

My understanding has been that transformers are great for sequences of tokens, but from what little I know of radiology sequence-of-tokens seems unlikely to be a useful representation of the data.

[fhd2]

On the surface, radiology seems like an image classification problem. That's indeed something a small NN can do, already 15 years ago. But it's probably not really all it is.

I can only imagine what people picture when they think about AI and radiology, but I can certainly imagine that it goes beyond that. What if you can feed a transformer with literature on how the body works and diseases, so that it can _analyse_ (not just _classify_) scans, applying some degree of, let's call it, creativity?

That second thing, if technically feasible, confabulations and all, has the potential to replace radiologists, maybe, if you're optimistic. Simple image classification, probably not, but it sounds like a great help to make sure they don't miss anything, can prioritise what to look at, and stuff like that.

[ptx]

> What if you can feed a transformer with literature on how the body works and diseases, so that it can _analyse_ (not just _classify_) scans, applying some degree of, let's call it, creativity?

Then it would make up plausible-sounding nonsense, just like it does in all other applications, but it would be particularly dangerous in this one.

[TechDebtDevin]

That wouldnt be that much different than current CNN/labeling methods used on medical imaging. Last time I got a ct scan the paperwork had the workstation specs and the models/nueral network techniques used.

[ptx]

Do you still have the links to what you dug up manually?

[Workaccount2]

https://www.mayo.edu/research/labs/radiology-informatics/ove...

https://pubmed.ncbi.nlm.nih.gov/31089974/

https://mayo-radiology-informatics-lab.github.io/MIDeL/index...