Telling a melanoma from a mole on a photograph · Contested (since 2017)

In January 2017 Nature published a convolutional network trained on 129 450 clinical images that matched 21 board certified dermatologists on the classification of keratinocyte carcinomas and of melanomas.

Both sides were given the same thing, a photograph with a biopsy proven answer behind it, and nothing else. A dermatologist in clinic has the history, the feel of the lesion under the finger, a dermatoscope, the rest of the skin and the option of asking the patient to come back in three months, and none of that was in the test. The result depends heavily on which images are chosen, and it has not become a standing clinical practice in the years since. Sorting a photograph is genuinely contested. Deciding to cut, and carrying the word melanoma to the person it belongs to, are not the same act.