Gary Marcus

Seriously? Hinton said we should stop training radiologists now because with advances in AI by the time all current radiologists retired we would not need anymore. Not that radiologists would disappear in 3 years. You might not agree with him, but he isn’t a fool.

false. what he actually said is not at all how you are reinterpreting him, see part after semicolon: “People should stop training radiologists now; its just completely obvious that within five years, deep learning is going to do better than radiologists."

Radiology is a not a static image recognition problem. For most non-binary images, it's a metadata problem where clinical state of the patient need to be considered as input. The data standard to summarize patient's physiological state does not exist. Radiologists will remain.

yes! we make that point (in different terms)

I would not put the blame of @geoffreyhinton Even if successful, such a CV tool only replaces part of a radiologists job. Inference is not the same as a detection - the same way a diagnostic detection is not the same as a prognostic prediction. My 0.02.

not Hintons fault that the problem is hard, but not great that he overstated when he said “I think if you work as a radiologist, you’re like the coyote that’s already over the edge of the cliff… within five [or 10] years, deep learning is going to do better than radiologists”

To be fair to Geoff, he said this *might* take ten years (so, up to 2026); and really, I suspect he was being provocative in order to try to get the field to engage to see if DL could help. Still, there are challenges - tech, regulatory, cultural - that might mean 2026 too early.

my problem is really with the coyote part and premature suggestion that we should stop training radiologists “I think if you work as a radiologist, you’re like the coyote that’s already over the edge of the cliff…“

Good article but it seems spurious to bring in adversarial attacks in this context. Xrays are made in a pretty controlled environment, and if you drop that constraint I’m pretty sure it would be trivial to confuse human radiologists too.