The software and scanning device was designed by computer scientists at Stanford. The software uses artificial intelligence for the diagnosis, based on a preparatory algorithm for skin cancer. The learning part of the program utilizes a database compiled from some 130,000 skin disease images. Over time, by providing ‘positive’ and ‘negative’ inputs, the researchers ‘trained’ the algorithm to visually diagnose potential signs of skin cancer. For the training, the algorithm was transmitted into the device as raw pixels, together with a disease label.
The algorithm was developed by Google. It was a software code that had been built to identify 1.28 million images from 1,000 object categories. What the researchers had to do was to use bespoke software designed to differentiate between images of animals on the Internet to detect malignant tissue growth. This was a slow process, given that no huge dataset of skin cancer images existed. The images came from high-quality, biopsy-confirmed pictograms provided by the University of Edinburgh’s medical department and from the International Skin Imaging Collaboration Project.
Initially the researchers were not sure if they could design a program that was capable of detecting signs of cancer successfully. According to Sebastian Thrun, who works at the Stanford Artificial Intelligence Laboratory: “We realized it was feasible… that’s when our thinking changed. That’s when we said, ‘Look, this is not just a class project for students, this is an opportunity to do something great for humanity.”
The ultimate test was pitching the software against 21 experienced dermatologists. Here various skin lesions were shown, some which were cancerous and some that were benign. In the study, where 370 test images were used, the algorithm matched the performance of dermatologists. Evaluation consisted of a sensitivity-specificity curve, which detailed the accuracy against the opinion of the dermatoligists.
Screening for skin cancer is important. If it is detected early then the survival rate is over 95 percent; however, if detected late the chances of survival fall to 15 percent or lower. It remains that visual examination is the primary way to detect cancer formation. This is aided by the use of a type of microscope called a dermatoscope. Where the visual check is unclear, a biopsy can be performed.
Given the success of algorithm the next step is to develop a functioning computer unit, suitable for commercialization.
The research is published in the journal Nature and it is titled “Dermatologist-level classification of skin cancer with deep neural networks.”
