It is important to detect melanoma early to avoid dying from it. Melanoma is a malignant type of skin cancer that accounts for over 70% of all skin-cancer deaths. However, suspicious pigmented lesions (SPLs), are common enough that doctors can't check them all. Researchers at MIT have created a tool that analyzes skin photos taken using a smartphone. This allows them to identify which SPLs need to be examined by a dermatologist.Professors Martha Gray, PhD 86, James Collins and Regina Barzilay, as well as postdoc Luis Soenksen (PhD 20), used deep convolutional neural network, which is machine-learning algorithm often used to classify images.COURTESY OF RESEARCHERSThe dermatologists classified the lesions visually in 20 388 images taken from 133 patients at Hospital Gregorio Maran, Madrid, and a variety of public images. The system was tested on the remaining images. It was trained using 80% of these images. It was able to distinguish between suspicious lesions, skin and backgrounds with more than 90% of SPLs. It was also able to determine the level of suspicion.Soenksen states that computer vision and deep neural network systems can quantify common signs with comparable accuracy as expert dermatologists. These screenings can be performed during routine primary care visits or by patients.
