AI Breakthrough Can Detect Skin Cancer Better Than Doctors

A new artificially intelligent computer can diagnose skin cancer more accurately than doctors, according to researchers.

The specially programed device deep learning convolutional neural network (CNN) was able to identify skin cancer more accurately than 58 dermatologists from 17 countries. Some 30 of the dermatologists involved in the study were considered experts as they had more than five years' experience, while 11 had between two to five and 17 had less than two.

The team, from Germany, the U.S. and France, behind the study published in the journal Annals of Oncology calibrated the CNN by showing it over 100,000 images of malignant and benign skin cancers, and inputting a diagnosis for each.

Melanoma is the deadliest form of skin cancer, and is expected to kill an estimated 9,320 people in the U.S. by the end of the year, according to the American Cancer Society. Worldwide, an estimated 232,000 new cases will be diagnosed in 2018, causing around 55,500 deaths. The condition is treatable when caught soon enough, but is all too often diagnosed when the cancer is more advanced and harder to treat.

Professor Holger Haenssle, senior managing physician at the Department of Dermatology, University of Heidelberg, Germany, and lead author of the study, explained that the CNN works like the brain of a child. The computer is made up of an artificial network of nerves that mimic the processes of the brain as it computes information taken in from the eyes. In a technique known as machine learning, the device was able to quickly evaluate information researchers presented it, and improve its ability to pinpoint skin cancers.

"Only dermoscopic images were used, that is lesions that were imaged at a 10-fold magnification," Professor Haenssle said in a statement. "With each training image, the CNN improved its ability to differentiate between benign and malignant lesions."

To see who would emerge victorious between the doctors and the CNN, the team presented each with 100 skin lesion images and instructed them to make a diagnosis and recommended a follow-up action. In a second test, only the dermatologists were presented with further information such as age, sex, and the position of the lesion alongside the same images.

On average during stage one, the dermatologists identified 86.6% of melanomas and 71.3% of lesions that were benign, whereas the CNN diagnosed 95% of melanomas. In stage two, the dermatologists improved their score to 88.9% of malignant melanomas and 75.7% non-cancerous lesions.

Professor Haenssle commented the CNN could be used to diagnose skin cancer faster and also decide whether clinicians should perform a biopsy.

"The CNN missed fewer melanomas, meaning it had a higher sensitivity than the dermatologists, and it misdiagnosed fewer benign moles as malignant melanoma, which means it had a higher specificity; this would result in less unnecessary surgery," he said.

However, the authors acknowledged that the study was limited in several ways, including the fact the dermatologists may have been unconsciously skewed because they knew they were not diagnosing a real patient whose life could be at risk.

More research needs to be carried out ensure the CNN is accurate when diagnosing areas difficult to image such as the fingers, toes and scalp, and to train it to pick up unusual melanomas, as well as lesions that patients haven't yet found.