A team of German, American and French scientists has developed a neural network that is more capable than doctors at recognizing cases of skin cancer. Most dermatologists were beaten by artificial intelligence.
The deep learning convolutional neural network was trained by more than a hundred thousand images of both malignant and benign melanomas. The network’s performance was then compared with the diagnoses of 58 dermatologists from 17 countries, more than half of whom have more than five years of experience.
Images of malignant and benign melanomas were shown, supplemented with photos of moles. In the end, the doctors were found to recognize skin cancer in 86.6 percent of the cases, while the neural network came out at 95 percent. At a later stage, the doctors received further medical information about the patients and close-up images. On that basis, the accurate recognition of melanomas rose to 88.9 percent. The neural network only had to make do with the images and received no further information.
The neural network appears to have a higher sensitivity because it has missed fewer melanomas. In addition, artificial intelligence has made a misdiagnosis less often, with benign moles, for example, still being classified as malignant melanoma. According to the German professor Holger Haenssle, this has shown that neural networks can perform better than experienced dermatologists. He argues that the better performance of AI will probably lead to fewer unnecessary surgical interventions in the long run.
The research still has some limitations. For example, the dermatologists knew that they were in an artificial situation and thus did not have to make impactful decisions for real patients. In addition, the images used in the test mainly showed melanomas and moles in white-skinned people.
Malignant melanomas are becoming more common. Worldwide, there are an estimated 232,000 new cases each year, with 55,500 patients dying from the consequences. The research is published in the scientific journal Annals of Oncology, under the title Artificial intelligence for melanoma diagnosis: How can we deliver on the promise?