Melanoma rates have increased for the past 30 years, according to the American Cancer Society. Skin cancer is now the most common of all cancers. While only about 1% of skin cancers are melanomas, the deadly tumors cause the large majority of deaths from skin cancer. For 2018 the American Cancer Society predicts about 91,270 new melanomas will be diagnosed and approximately 9,232 people will die of melanoma. In early 2017 we wrote about a study at Stanford University where an interdepartmental team trained a deep convolutional neural network (CNN) to classify skin lesions by appearance. In tests comparing the CNN’s diagnostic capability, the team found the CNN performed “on par” with 21 board-certified dermatologists.
In a larger international study just published in Annals of Oncology, a neural network outperformed the Standford study’s CNN. Based on Google’s Inception V4 CNN architecture, a convolutional neural network more accurately diagnosed skin cancer than a group of 58 dermatologists from 17 countries. German, French, and American researchers trained the AI engine with more than 100,000 correctly diagnosed and labeled images of skin cancers. In the study, the physicians were shown 100 images of skin cancers and asked to classify them as melanomas or benign lesions. Regardless of experience level, including 19 beginners, 11 skilled practitioners, and 30 experts, the doctors averaged 86.6% diagnostic accuracy. Four weeks later the same doctors diagnosed the same images, this time with additional clinical information about each patient and accompanying close up images. On the second test, the doctors averaged 88.9% accuracy. The CNN scored 95% accuracy on its single test with the initial images only.
More study is needed, and it’s interesting to note that 18 of the physicians in the international test scored higher than the CNN. The study’s authors noted that the results suggest not that AI replaces physicians, but that dermatologists, regardless of experience, could benefit from using AI dermoscopy screening as a diagnostic aid.
