An artificial intelligence outperformed a team of dermatologists when asked to diagnose skin cancer.
The recent study was published only a few days ago in the medical journal Annals of Oncology. Led by researchers from Germany, France and from the USA, it showed that a trained CNN (artificial neural network), called Inception V4, was able to detect possible melanomas, and to be better at it than regular doctors.
To test its abilities, scientists first showed a series of 100 pictures of different moles to the machine, then to a group of 58 international dermatologists. The computer successfully detected 95% of the dangerous moles. Doctors only found 86.6% of them, but found 71.3% of the benign cases, versus 63.8% for Inception V4.
When given additional informations about their patients, such as zoom-in pictures or other medical elements, the medical team’s success rates went up to 88.9% for the melanomas, and to 75.7% for the benign ones.
Last year, Brett Kuprel explained the operation of such a system and such a test during the TensorFlow Dev Summit.
Yet, if this AI seems to be more efficient than regular human doctors, the scientists behind this study still explained that this kind of tool are only thought to assist doctors, not to replace them, that it was made to give them a complementary opinion on their cases.