There really should be at this point. Annotated radiology datasets, patients numbering into the millions, are the easiest healthcare datasets to obtain. I suspect there are many startups, and know of several long since failed, who trained on these. I've met radiologists who assert most of their job comes down to contextualizing their findings to their colleagues, as well as within the scope of the case itself. That's relevant here - it doesn't matter how accurate or precise your model is, if it can't do that. Radiologists already use "AI" tools that are very good, and radiology is a very welcoming field for new technology. I think the promise of foundation models at the moment would be to ease burden and help prevent burnout. Unfortunately, those models aren't "sexy" - they reduce administrative burden, assemble contextual evidence for better retrieval (have interfaces that don't suck when integrated with the EMR).
