Google’s latest foray into healthcare has been a web tool that uses artificial intelligence to help people determine their skin, hair or nail condition. The company previewed the tool at its I/O event and announced that it plans to launch a pilot later this year.
The use of the tool is as follows: People take three pictures of the problematic area on the skin or nails with their phone’s camera. They then answer a series of questions about their skin type and other symptoms. The tool then presents a list of possible conditions it has been trained to recognize.
Google Health chief medical officer Karen DeSalvo says that because of the prevalence of skin problems, Google aims to tackle them using artificial intelligence: “People come to Google to ask questions about skin problems. About 10 billion skins a year. We’re questioning the situation.” DeSalvo says there is a significant shortage of dermatologists in the world, and he hopes this tool will help people quickly get accurate information about potential conditions without having to spend hours doing their own online research.
The team trained the model with millions of skin problem images, thousands of healthy skin images, and 65,000 images from clinical settings. The model takes factors such as age, skin type, gender and race into account when suggesting possible conditions. When tested on nearly 1,000 skin problem images from various patients, Google was able to identify the top three recommendations with 84 percent accuracy.
In fact, the new system builds on Google’s past work using artificial intelligence tools to identify skin conditions. The company published the first update of its deep learning system in Nature Medicine last month. This article showed that the system could identify 26 common skin diseases as accurately as dermatologists and more accurately than primary care physicians. In April, the company also published another study showing that the system could help non-dermatologists diagnose skin conditions more accurately.
Google is also working with the Stanford University research team to test how well the tool works in a healthcare setting.
28590
