I think the key insight is that AI is (undoubtedly going to be) better at analysis and diagnosis than radiologists, but isn't yet widely deployed because:

1. The medical world doesn't accept new technologies easily. Humans get a much higher pass on bad performance than technology and especially than new technology. Things need to be extensively tested and certified, so adoption is slow.

2. AI is legally very different than a radiologist. The liability structure is completely different, which matters a lot in an environment that deals with life or death decisions.

3. Image analysis is not language analysis and generation. This specific machine learning part is not the bit of machine learning that has advanced enormously in the past two years. General knowledge of the world doesn't help that much when the task is to look at pixels and determine whether it's cancer or not. Now this can be improved by integrating the image analysis with all the other possibly relevant information (case history etc.) and diagnosing the case via that route.