Thread by @woojinrad

"This free AI reads X-rays as well as doctors."
No, it doesn't! The first image I uploaded was an image of an apical pneumothorax from the web, and it completely missed it. This isn't something you want to miss.
snip.ly/quakd3 @FastCompany #radiology #AI

The creator's intention is a good one, but this type of headline is both inaccurate and dangerous. #DeepLearning has a lot of great potential in #medicine, but we have to be careful not to overhype #AI. Oh, btw, this patient does not have cardiomegaly. #radiology #ImagingAI

To be fair, the authors in the original article (arxiv.org/pdf/1901.11210…) do not claim to be as good as or better than radiologists but rather to "augment" or "assist" them. Good: they show potential for an #AI solution with minimal computational cost while preserving pt privacy.

The authors also looked at @DrLukeOR's evaluation of the NIH dataset... "There have been claims that this dataset contains many errors in labeling"... though I would say it's more than a "claim." In the end, #AI in #medicine needs AI expertise *and* clinical domain expertise.