Editors' ChoiceDERMATOLOGY

Turning skin “check” into checkmate

See allHide authors and affiliations

Science Translational Medicine  01 Mar 2017:
Vol. 9, Issue 379, eaam9856
DOI: 10.1126/scitranslmed.aam9856

Abstract

Artificial intelligence algorithm performs on par with dermatologists to distinguish malignant from benign skin tumors.

Twenty years ago, the world anxiously watched the chess match in which Deep Blue beat Gary Kasparov. This etched into popular culture that computers can outperform humans, at least in pattern recognition. Current artificial intelligence technology produces computers that are infinitely better, train in hours, and impressively, train themselves. In their report, Esteva et al. tested the ability of a computer-based deep learning algorithm to distinguish benign from malignant lesions using only nonstandardized images.

In this effort, investigators trained a convolutional neural network (CNN) using about 1.41 million images to help the CNN classify melanomas versus benign nevi and keratinocyte carcinomas versus benign seborrheic keratosis. Uniquely, they relied only on nonstandardized images, thus making the technology readily accessible. The investigators tested the CNN against 21 Board certified dermatologists. The CNN performed on par with dermatologists, demonstrating ability to opt for a biopsy when the lesion is malignant and avoid sampling benign neoplasms.

This study using deep learning allows a peek into future medical care, as this technology is likely to change how medicine and other fields evolve. Being able to reliably tell benign lesions from malignant based on a smartphone picture can help deliver care where dermatologists are unavailable and also, though yet untested, may help augment the dermatologist’s current decision process. This is not likely to put dermatologists out of business, as, importantly, dermatologists’ decision to biopsy a lesion is based on a physical exam of a live patient, not an image, and dermatologists’ differential diagnosis is often not binary as they consider more than two options, as well as consider the patient’s preferences. Yet, deep learning is evolving rapidly. Therefore, we may be only a few years away from CNN turning this skin “check” into “checkmate.”

Highlighted Article

View Abstract

Navigate This Article