So glad you asked! To start, you can’t just train a set on randomly submitted diagnostic information. You need good curated information about the patients’ history. You need appropriate numbers of both positive and negative examples of all target conditions. 1/4 #pathsky #medsky - ThreadSky

mopathguy.bsky.social • 110 days ago

So glad you asked! To start, you can’t just train a set on randomly submitted diagnostic information. You need good curated information about the patients’ history. You need appropriate numbers of both positive and negative examples of all target conditions. 1/4 #pathsky #medsky

Reposted from Doug Wolf

Doc, I detest (and distrust) Musk, but am generally in favor of ML/automation as tools to assist trained pathologists. What dangers are you anticipating here? Idiot users asking Grok to diagnose a dark spot on a scan?

Comments

gleason4plus5.bsky.social•110 days ago

My big fear about AI is the incorrect information. It's nearly impossible to tell in an AI chat whether it is hallucinating something or saying something factual. I had a chat with google awhile back, when I was curious what it knew about thymic SCC. https://g.co/gemini/share/9e0a94460973

mopathguy.bsky.social•110 days ago

To be fair, not all AI tools are large language models. We use a AI tools in NGS that don’t hallucinate.

mopathguy.bsky.social•110 days ago

You need to know the strengths and weaknesses of the diagnostic technology used to generate the data. You need to have the knowledge to know where to set your thresholds (what should count as positive). You should be aware of (and responsible for) the consequences of getting anything wrong. 2/4

mopathguy.bsky.social•110 days ago

All that’s just off the top of my head. None of the above can be accomplished by feeding random, uncurated, unannotated, findings into an AI. The real problem, though: it’s going to be too easy to make it look like the AI “is pretty good,” 3/4

mopathguy.bsky.social•110 days ago

and it doesn’t seem like the FDA under the next administration will be equipped or empowered to step in. All of that without even addressing the person driving it, and their motives/flaws. 4/4

ridertoast.bsky.social•110 days ago

The people pushing this stuff, have an outsiders understanding of what anatomic pathology is. there’s a reason the loudest proponents of this technology are often not AP trained

mopathguy.bsky.social•110 days ago

Yes it’s the “it can almost always tell the difference between a chihuahua and a chocolate chip cookie” school of medical diagnostics.

ridertoast.bsky.social•110 days ago

Thanks AI chatbot for telling me “this is stratified squamous epithelium with concerning features dysplasia. Clinical correlation is recommended”

wissensfan.bsky.social•110 days ago

Don’t sell Grok short!
Musks AI - in large parts fed by Putin-Bots, conspiracy theorists, MAGA-Fans and other natural not that bright geniuses - is at least intelligent enough to identify Musk as one of the biggest lying toad on Twitter - nowadays called X!

https://www.independent.co.uk/tech/elon-musk-grok-twitter-ai-misinformation-b2645906.html

ridertoast.bsky.social•110 days ago

For me, I think the main thing is just that for the hardest cases. The reason they are hard is because if you show 10 pathologists they’ll disagree 6 to 4 and then if you showed them again, they disagree 4 to 6. There is no standard to build a model on for most things.

darkmich.bsky.social•110 days ago

And few people understand how difficult it is to get case sets like this together without biasing the collection.

Comments

Posting Rules

Reply