The world's best dermatologist

Artificial intelligence has surpassed dermatologists in the diagnosis of melanoma

Natalia Pelezneva, Naked Science

Australian researchers reported that an algorithm based on convolutional neural networks coped with the diagnosis of melanoma better than most experienced dermatologists who participated in the experiment. The program was able to accurately detect melanoma in 95% of cases, and people – in less than 90%. The study is published in the journal Annals of Oncology.

The main way to diagnose melanoma is dermatoscopy, examination of dark formations on the skin with a magnifying glass.

Distinguish a malignant tumor from an ordinary mole (nevus) allow characteristic features: asymmetry of the spot, heterogeneity of color, uneven edges. The algorithm was trained to recognize tumors using a database of one hundred thousand photos of various formations on the skin – both malignant and not dangerous to health.

Then the program was tested on another, unfamiliar database of 300 photos. One hundred of them were offered to 58 practicing dermatologists from 17 countries to view. 29% of doctors were just starting practice, they worked in the specialty for no more than two years, 19% practiced for two to five years. 52% of the participants in the experiment had more than five years of experience.

At the first stage of the experiment, doctors and the program detected melanoma only on the basis of photographs. People were successfully diagnosed with cancer in 86.6% of cases, artificial intelligence – in 95%. Harmless moles that do not require treatment were detected by the program and people equally often – in about 71% of cases. At the second stage, four weeks later, dermatologists received detailed data about patients, including their age, gender and the area of the body where the suspicious neoplasm is located. After that, they were able to recognize 88.9% of melanoma cases. Experienced doctors coped with the diagnosis better than beginners, but on average worse than artificial intelligence.

Example of a test image (Haenssle et al.)

The authors emphasize that the programs will not replace experienced doctors, but they can become an effective and accurate tool for diagnosing malignant tumors. Melanoma is a widespread disease, in 2015 it affected more than three million people worldwide.

Earlier, the United States approved for the first time the use of a program capable of independently diagnosing eye diseases.

Portal "Eternal youth" http://vechnayamolodost.ru