Melanomas are treatable with early detection—but what resources do those living where dermatologists are scarce or nonexistent have if they think they have a melanoma? Image recognition technologies may provide the answer and a team of Stanford researchers is one step closer to finding it.

You may be reminded of taped interrogations in which investigators interpret “microexpressions”—brief, involuntary facial expressions that reflect emotions—to determine whether a suspect is lying. Star Trek fans may remember the “tricorder”  that Dr. McCoy used to figure out what was ailing various crew members on the USS Enterprise. In this case, the image recognition technology is an algorithm-based for recognizing various skin cancers. The Stanford researchers were able to pull this off and published their findings in January 2017.

Skin cancers are the most common forms of the disease. They are highly-curable and less than five percent of skin cancers diagnosed in the United States are melanomas—malignant tumors that originate in the pigment-producing cells of the skin’s basal layer. Yet melanomas are responsible for almost 75 percent skin cancer-related deaths. Those diagnosed at a melanoma’s latest stage have only a 14 percent chance of surviving.

Using a convolutional neural network—a type of biologically-inspired machine-learning software—they enabled a computer to recognize over 2,000 skin conditions. They also gave the computer the ability to recognize over 1 million everyday objects.

“The algorithm performed as well as board-certified dermatologists at several key diagnostic tasks,” said Andre Esteva, a member of the team and a doctoral student in Electrical Engineering. The algorithm could eventually result in a mobile app for users in poor countries. They would be able to scan a skin aberration with their phone and cross-check it against an expansive database of cancerous growths. Researchers also see applications of this technology in other medical fields. However, many more studies will be required before then. One question that needs to be investigated is whether the algorithm can distinguish between skin diseases that look similar.