A neural network that classifies skin cancer as well as dermatologists
- Article
- Health
Skin cancer is diagnosed mostly by looking. A clinician screens a lesion visually, and a biopsy and histopathological examination confirm what they suspect. Because the first step is visual, it is a natural fit for deep convolutional neural networks (CNNs), which classify images. A 2017 paper in Nature by Andre Esteva, Brett Kuprel, Roberto A. Novoa, Justin Ko, Susan M. Swetter, Helen M. Blau, and Sebastian Thrun tested how far that idea could go. If a CNN could match a trained dermatologist, the same model could later run inside a smartphone application and put a first layer of screening within reach of people who cannot easily see a specialist.
Approach
The team trained a CNN on 129,450 clinical images labeled across 2,032 different diseases. The images came from more than 18 open-access online repositories and from clinical data gathered at Stanford University Medical Center. The dataset was split into 127,463 training images and 1,942 validation images. Performance was measured with nine-fold cross-validation, using sensitivity and specificity as the comparison metrics, and the network’s calls were tested against the diagnoses of trained dermatologists.
Findings
The CNN matched the dermatologists on general skin lesion classification. In the first round of validation, a three-class disease partition, the network reached about 72 percent accuracy while two dermatologists scored close to 66 percent. In a harder nine-class partition, the network scored about 55 percent overall, close to the dermatologists’ 53 and 55 percent.
The most important test used images confirmed by biopsy and compared the network against 21 dermatologists on melanoma dermoscopy images. The CNN reached an area under the curve (AUC) of 0.91, and it did better on average than the dermatologists. On carcinoma images the AUC was 0.96, and on melanoma photographs it was 0.94.
What it means
The authors see a model with this much predictive power as a candidate for a smartphone application that screens for skin cancer. It cannot replace expert consultation, biopsies, or histopathological examination. What it can do is act as a preliminary layer of the diagnostic process, which matters most for the many people who never get to that first consultation at all.
Source: Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., and Thrun, S. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542, 115 to 118. doi:10.1038/nature21056