Most people cite liability as the main reason but I also haven't seen any algorithm that reads chest radiographs better than radiologists. To be fair, rads look at images, correlate with the patient chart and symptoms, and that info is not provided to the algorithms. But also, humans need only one or two examples of a weird diagnosis to find it again and that still can't be tackled by one-shot/few shot learning algorithms (yet!).

https://towardsdatascience.com/why-ai-will-not-replace-radio...

This covers some points, but "never?" I'm not so sure about that... That's why if I can't beat 'em I'm trying to join em