Skin cancer was found to be diagnosed more accurately by artificial intelligence than experienced dermatologists in a new international study. Researchers tested a form of machine learning known as a deep learning convolutional neural network (CNN) to reach this conclusion.

The study titled “Artificial intelligence for melanoma diagnosis: How can we deliver on the promise?” was published in the cancer journal Annals of Oncology on May 28.

Malignant melanoma accounts for 1 percent of all skin cancers but causes a majority of skin cancer-related deaths. The American Cancer Society estimates 9,320 people will die from melanoma in 2018 while 91,270 new cases will be diagnosed. As part of the new study, researchers from the United States, Germany, and France decided to test the performance of a CNN in diagnosing malignant melanomas by looking at images of moles.

“The CNN works like the brain of a child,” explained study author professor Holger Haenssle from the University of Heidelberg, Germany.

More than 100,000 images of malignant and benign skin cancers and moles were shown to the network along with the diagnosis for each image. 

“Only dermoscopic images were used, that is lesions that were imaged at a 10-fold magnification. With each training image, the CNN improved its ability to differentiate between benign and malignant lesions,” he added.

Once the network was trained, two sets of images were built using new pictures never seen by the CNN. The first set of 300 images was to test the abilities of the CNN alone, while the second set of 100 images was to test both the network and a group of doctors. 

From 17 countries across the world, 58 dermatologists agreed to take part in the test of man versus machine. 

The number of skin cancer cases missed by the CNN were fewer than those gone unnoticed by dermatologists, indicating a higher sensitivity. The misdiagnosis of benign moles as melanoma also saw a lower rate with the network, which could help avoid unnecessary surgery.

On average, the dermatologists correctly detected around 86.6 percent of melanomas while the CNN identified 95 percent of them. After the dermatologists were provided clinical information about the patients such as their age, their sex, and the location of the lesion, their success rate in diagnosing melanoma increased to 88.9 percent. 

“When dermatologists received more clinical information and images at level II, their diagnostic performance improved,” Haenssle said. “However, the CNN, which was still working solely from the dermoscopic images with no additional clinical information, continued to out-perform the physicians' diagnostic abilities.”

The researchers do not believe the CNN will replace human professionals, but rather, be used as a tool to reduce the risk of misdiagnosis. They added most dermatologists were already using digital dermoscopy systems and other such tools for cancer-related documentation, detection, and follow-up of patients.