Google is using AI to detect skin conditions and tuberculosis

Medical applications of artificial intelligence (AI) is not a new domain. During the ongoing pandemic, organizations and researchers have been utilizing the technology for diagnostics, managing COVID-19 patients, and ensuring social distancing. Despite a lot of work being done in this space already, there's still significant potential for other applications too. Google is currently building AI tools to detect skin issues as well as tuberculosis (TB).

Starting with dermatology, Google is building an AI tool which enables you to detect skin conditions using just your phone's camera. The company says that billions of people suffer from skin conditions but there is a deficit of specialists, which is where this tool comes into play. By utilizing a web-based experience, you can simply take three photos of the problematic area from different angles, after which you'll be asked some questions about it. Based on your photos and your answers, the AI model will narrow down possible conditions with reliable information for you to research further. Currently, the model contains information about 288 different conditions.

Google claims that its tool is trained on millions of data points from various demographics, but it is still meant for assistance only. It has not been evaluated by the U.S. Food and Drug Administration (FDA) and as such, should not be a replacement for any kind of review by a medical professional. Google plans to launch a pilot of the tool later this year and those interested in it can sign up here.

When it comes to TB, Google is using deep learning systems to detect potential patients and recommend them for follow-up testing. The model essentially uses chest X-ray images as a preliminary step for screening before people directly opt for the diagnostic test which is much more expensive.

Google claims that its model returned false positives and false negatives at rates similar to those of 14 radiologists. To build its AI system, Google used de-identified data from nine countries and then tested it on cases from five countries. The model returns an output ranging between 0 and 1, which means that it outputs a probability of whether a follow-up diagnostic test is necessary or not. It is up to clinics which utilize the technology to determine a threshold to recommend additional tests, should they choose not to use the default configuration offered by Google.

The research paper for these findings can be viewed here. The company will be expanding its research by partnering with hospitals in India and Zambia later this year.