The concept of using smartphones to diagnose possible skin cancers has been around for more than a decade, but the technique hasn't progressed far enough to threaten a professional dermatologist's precision. However, researchers at MIT and Harvard recently created deep learning algorithms that can accurately detect skin cancer.
Until now, skin cancer detection algorithms have been equipped to scan individual skin lesions for unusual features that may indicate melanoma, but what the researchers from MIT and Harvard have created is a little different. Instead, the researchers used the "ugly duckling" criteria, which is based on the idea that most moles on a person would appear alike, and those that don't, dubbed "ugly ducklings," are considered melanoma warning signs.
The researchers claim that their method is the first of its kind to reproduce this procedure, and they began by compiling a collection of more than 33,000 wide-field photographs that included not only the skin of the patient but also other artifacts and backgrounds. The team used the ugly duckling approach by creating 3D maps of all the lesions in a given image – for example, scattered around a patient's back – and calculating how strange the characteristics on each lesion is. The machine assigns values and determines the lesions were harmful by measuring how rare some of these features were relative to those on other lesions in the picture.
Using 135 wide-field images, the technology was previously used to recognize suspected lesions in 68 different patients. Individual lesions were given an oddity score depending on how troubling their characteristics were, and the results were contrasted to those of three dermatologists. The algorithm approved 88 percent of the time with the dermatologists' consensus, and 86 percent of the time with actual dermatologists.
“This high degree of agreement between artificial intelligence and human clinicians is a significant advance in this field,” says co-author Jim Collins, “because dermatologists' agreement with each other is generally very high, about 90%.” “In essence, we've been able to gain dermatologist-level precision in diagnosing possible skin cancer lesions from photographs that anybody with a smartphone can take, which opens up tremendous potential for detecting and treating melanoma earlier.”
