A New Algorithm Has Beaten 21 Doctors At Spotting Skin Cancer


Can a robot be your doctor? It’s an excellent question, and one that’s more urgent than you might think. Automating basic medical care like diagnosis and routine testing would save patients billions and doctors valuable time and guesswork. So news that an algorithm is as good at spotting potential skin cancer as a squad of doctors is excellent.

A Stanford University team realized that skin cancer, one of the most common cancers out there, is usually diagnosed visually. So they used Google’s “deep learning” image tools and assembled an enormous collection of freckles, moles, rashes, and other skin problems, some malignant, some benign, and started teaching the algorithm. When it was put head to head against a team of doctors, it had the same accuracy they did, at 91%.

Skin cancer is one of the most common cancers there is, with one in five Americans diagnosed with it at some point in their lives. If it’s an invasive melanoma, and nearly 80,000 cases of those are diagnosed a year, catching it early is crucial.

We’re still a while away from this becoming something you can use at home. The algorithm has been trained on photos, after all, not real people; and algorithms can often stumble if the humans feeding them data make mistakes. It’s an important step forward, however. If spotting a potential skin cancer is as simple as downloading an app and taking a shot of that weird mole on your arm, that might mean dangerous cancers are caught much, much earlier.

(Via The Verge)