Early detection is key to surviving melanoma, a type of malignant tumor responsible for more than 70% of skin-cancer-related deaths worldwide, but “suspicious pigmented skin lesions” (SPLs) are so common it’s impractical for doctors to check them all out. Now MIT researchers have developed a tool that can analyze skin photos taken with a smartphone to determine which SPLs should be evaluated by a dermatologist.

The researchers, who include professors Martha Gray, SM ’81, PhD ’86, James Collins, and Regina Barzilay and postdoc Luis Soenksen, PhD ’20, made use of deep convolutional neural networks, machine-learning algorithms often used to classify images.

COURTESY OF THE RESEARCHERS

The team had dermatologists visually classify the lesions in 20,388 images from 133 patients at the Hospital Gregorio Marañón in Madrid and a number of publicly available images. The system was trained on 80% of those images and tested with the rest. It distinguished more than 90.3% of SPLs from nonsuspicious lesions, skin, and complex backgrounds. It also was able to classify the level of suspiciousness. 

“Our research suggests that systems leveraging computer vision and deep neural networks, quantifying such common signs, can achieve comparable accuracy to expert dermatologists,” Soenksen says. The screenings could be done during routine primary care visits, or even by patients themselves.