That's totally meaningless without a control that shows how it did on a set of images without nipples and how it did on a mixed set of images both with and without (after all, these could simply be selected to show 'good' cases).
Detecting nipples when you know they are present is relatively easy.