Deep learning algorithm matches dermatologists at recognizing skin cancer

Worried about a strange mole on your back? Why not let an algorithm look at it!

That’s the broad idea behind a recent project created by computer scientists at Stanford University, which applied the extraordinary machine vision powers of cutting-edge deep learning neural networks to the world of dermatology.

Using a database of close to 130,000 images of skin diseases, the team was able to create an artificial intelligence algorithm able diagnose skin lesions with a performance level matching trained experts.

“[We trained it to] classify images of skin conditions as benign or malignant, and found that it matches the performance of over 21 board-certified dermatologists at three key diagnostic tasks: identifying keratinocyte carcinomas (the most common human cancer), identifying melanoma (the deadliest skin cancer), and identifying melanoma when viewed using dermoscopy,” co-first author Andre Esteva told Digital Trends.

The neural network the researchers used was one originally designed by Google and trained to recognize 1.28 million images, with the somewhat frivolous purpose of distinguishing cats from dogs.

“We saw that it was demonstrating superhuman performance at distinguishing between 200 different types of dog,” co-first author Brett Kuprel told us. “We thought we could apply this to something more useful, such as skin cancer diagnosis.”

Prior to the project, neither Esteva or Kuprel had any background in dermatology, which means the algorithm they created was able to achieve expert-level performance without benefitting from any specially encoded domain-specific knowledge.

However, if the algorithm was to be used by trained physicians, they could take advantage of a so-called “salience map,” revealing how important each pixel in an image was in the AI’s prediction process. In other words, rather than replacing dermatologists, this could prove to be a useful tool in their arsenal — the equivalent of a smart X-ray that offers its own interpretation about what it sees.

For now, though, that’s jumping way ahead. “There are definitely regulatory rules to get the FDA to approve it,” Kuprel said. “That would be important before any application could be deployed.” Beyond this, though, the investigators aren’t saying what is next.

“We are still deliberating on next steps and cannot yet comment,” Esteva said.