Comment by dnautics

A transformer trained only on cancer images could probably do better but the old model is trained on a well curated data set to do one thing. A decision model can decide on a much broader universe of things, only a sliver of which is cancer (and probably there's a smaller cancer dataset too)