I'm a doctor but not a radiologist. I imagine it would go like this:

Computer makes diagnosis on image of lesion that has already been biopsied and has a known correct diagnosis.

I don't know how it would work for human false negatives that were never biopsied. Possibly the AI could be trusted enough to make a biopsy on its recommendation.