Is it a freckle, a benign cancer, or a malignant time bomb about to go off? Dermatologists train in part by looking at thousands of images of skin abnormalities as one part of their diagnostic training. Doctors use much more than a single image to make their determinations, but the first look is an important step. According to the American Academy of Dermatology, skin cancer is the most common cancer in the U.S and best estimates are that one in five Americans will develop skin cancer in their lifetime. The Skin Cancer Foundation estimated that 76,380 new cases of invasive melanoma would be diagnosed in the U.S. in 2016, and that 10,130 people would die of the malignancy. Melanoma is present in less than 1 percent of all skin cancers but accounts for the vast majority of deaths.
Since appearance is the first determinant in classifying skin cancer, researchers at Stanford University‘s Department of Electrical Engineering, Department of Dermatology and Department of Pathology first trained a deep convolutional neural network (CNN) to classify skin lesions by appearance. The network was trained with 129,450 clinical images associated with 2,032 different diseases. After training, the CNN was tested against 21 board-certified dermatologists in recognizing the most common cancers and the deadliest skin cancer. The neural net performed on par with the clinicians on each task. The conclusion is that artificial intelligence is already capable of classifying skin cancers by appearance on a level comparable to dermatologists.
The purpose of developing and testing the neural net’s diagnostic ability is not to replace doctors, but to act as a first line of screening and extend dermatologists’ reach outside clinical settings. Neural networks improve their accuracy with more data. The Stanford researchers estimate that by 2021, 6.3 billion camera-equipped smartphone subscription could provide “low-cost universal access to vital diagnostic care.”