https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1665219/
That's been in use for some time. But like many parts of radiology it really only can be a second look tool that as you mentioned proposes alternatives or suggests things. The false positive rate for CAD is substantially higher than for humans because of the human ability to see symmetry and patterns in very diverse tissue sets like one sees in screening mammography.
And the nature of screening tests like mammography means that actually percentages like "89%" isn't really good enough. You have to be more specific and sensitive than that to have a successful program, and I'm not sure that ML will ever be able to get there...there's a lot of experience and human intuition involved at some point that would be hard to replicate...and I know that because people have been trying to do that for decades.