THIS is the AI we need.
Agreed, this is exactly what reinforcement learning and neural networks are good at. Calling them AI is beyond dumb, but hey marketing will be marketing. It’s pattern recognition, which is cool, but nobody would call that intelligent otherwise. Another big issue with the marketing is they only report on the success rate and not the failure rate. Doctors praise the cases being caught, but dislike the models pointing out stuff that is clearly not a tumor. It wastes time for people already short on time. These models also risk doctors becoming over reliant on them, even though they can have serious blind spots and thus miss stuff a doctor would have caught. Or the other way around, have people receive treatment (often not without risk, discomfort and cost to the patient), where none was needed. The thing that bothers me the most is how it’s always framed as a win for AI. Like see AI is good at diagnosing cancer (which then gets extrapolated to curing cancer for some bizarre reason), so that useless chat bot is also good somehow. Because AI.
Robert Murphy’s lab at Carnegie Mellon has developing learning sets like this for 20 years.
This is not designed to replace medical opinion, it’s designed to cross check as pathologists and radiologist have about 1% misses which is not acceptable.
But that takes actual work. See how the LLM systems are constantly wrong? That is because after you get to about 80% accuracy the rest will murder you.
This would take time and actual investment. Not something big tech can handle.
canwants toFTFY
That’s particularly useful for pancreatic cancer, if it’s accurate, reliable, cost effective, and practical in the real world.
Hey lookie here, the statistical pattern matching algorithm has some uses that could help society maybe possibly. Sure beats replacing artists or building inefficient chat bots that give people the Eliza effect.
deleted by creator
False positives are at least as dangerous as false negatives and AI solutions like this have massive problems with over diagnosing.
Absolutely 100% wrong.
In pancreatic ductal adenocarcinoma, a false positive means a follow-up scan. A false negative means death, the 5-year survival is near zero once it’s caught late, but exceeds 80% when caught early.
In the study, the radiologists’ lower false positive rate is achieved by missing 78% of cancers. That’s not a safer trade-off, it’s just a different way to fail. “Overdiagnosis” also requires a disease that might not have harmed the patient, PDA doesn’t have a harmless form. Every missed case is a lost life while every false positive is an extra doctor’s appointment.
This system detects twice as many cancers and was flagging them, on average, 675 days (nearly 2 years!) before clinical detection.
deleted by creator
If I’m wrong, then feel free to support your position with evidence or an argument showing that my statement was specious.
I linked the, peer-reviewed, paper which contains the data that supports my statements on the topic.
You’ve made two conclusory statements and immediately resorted to insulting comments when challenged.
There is not a single aggressive pancreatic cancer where a false negative is more dangerous than a false positive.
Percutaneous biopsy has a mortality rate of approximately 0.2% even relatively non-malignant pancreatic cancers (say Solid pseudopapillary neoplasm) have 10-year survival rates in adults of around 88% and that number is from cases which received surgical intervention and chemotherapy something that would not happen with a false negative.
So even in the worst case, the false negative multiple times more deadly. A false positives’ most likely outcome is pancreatitis from the biopsy procedure.
Stop being a bad person, please.
You’re rude, arrogant, and wildly incorrect from a medical standpoint. Please delete your message and don’t make comments like this in the future.
This comment should be deleted soon
Saw your edit. I’m not a “home-bound tech worker” and I think you’re projecting. What’re your medical qualifications?
In other words: not useful at all. (Didn’t read the article because it already misuses the AI acronym in the title, indicating it was written by some idiot with nothing to say)
You used the AI acronym in the same way, so I’m confused by your arrogant sounding statement
Did I though? Are they using a model with any kind of abstraction layer that actually understands relationships between objects?
Yes and yes.
These are questions that you wouldn’t have to ask if you didn’t smugly decide that you didn’t need to read before contributing your opinion.
If you can’t be arsed to read the article, here’s the peer reviewed paper in the British Medical Journal: https://www.science.org/doi/10.1126/science.adz4433
