CVP Prototypes Skin Cancer Detector for Smartphones
Mobile apps have become ubiquitous and Artificial Intelligence (AI) is disrupting entire industries. As a new partner of Google Cloud, CVP wanted to evaluate its cloud computing services and see if we could challenge ourselves to quickly build a mobile app that used AI on the Google Cloud Platform (GCP). The use case we chose was the preliminary diagnosis of whether or not a skin lesion was a benign or a malignant melanoma, based on how skilled dermatologists had classified similar images in the past. The application–which we developed in just two weeks–now works on any iOS or Android smartphone, is over 85% accurate on an independent test data set, and is hosted in Google Cloud at a cost of pennies per month.
The great success of this experiment now serves as a blueprint for how we can quickly build sophisticated, AI-driven mobile apps for our clients. In two weeks, we were able to go from concept to an app on a device that had performance levels similar to a dermatologist performing the same task.
To build the app, we used:
- Automated Machine Learning as a service — Rather than spend weeks hand-coding and designing a complex neural network, CVP used an automated machine learning service that uses reinforcement learning to autonomously design a machine learning model. This cut the development time required from weeks to hours.
- Cloud computing — Instead of setting up dedicated servers to host the back-end for the application, CVP used something called “serverless computing” where we upload only code to the cloud and it is the cloud provider’s (e.g., Google, AWS) responsibility to run the servers, keep them patched, and scale them up in response to more activity. In a matter of hours, the system’s back-end was running in Google Cloud and costs were under a dollar per month to run.
Due to the FDA’s regulation which classifies apps that diagnose a disease as medical devices, CVP is unable to publicly release the app in the short-term, but we are now looking into how this research can benefit the health industry in other ways. At future conferences, CVP plans to present on how healthcare organizations can build similar apps to make automated telemedicine a reality for a wide variety of conditions.