> This is a great example of where we need to get humans out of the equation now that a machine performs consistently better. No, it's not. The system is an i

> This is a great example of where we need to get humans out of the equation now that a machine performs consistently better. No, it's not. The system is an image classifier with a HUGE false positive rate (and false negative rate). When false positive rates exceed actual incident rates in the population (or far exceed by orders of magnitude, in this case), then it's practically worthless. This is something that the medical community just does not get about statistics (edit: I'm not making a wild generalization here, there are articles out there about this issue). What this study actually shows is that an image-only diagnosis of melanoma sucks and should never be used. It doesn't matter if it out-performs doctors, because in either case the diagnosis is garbage.