A neural network system that analyzes photographs can rank and distinguish suspicious, potentially precancerous skin lesions, which can turn into the deadly skin malignancy melanoma if not caught and removed early. The system accurately scoped out suspicious lesions from 68 patients in a manner that mostly matched tried-and-true evaluations from dermatologists. The results suggest the platform could help clinicians spot suspicious lesions during clinical visits faster and on a larger scale, potentially allowing for earlier diagnosis and treatment. Melanoma is the deadliest form of skin cancer, but outcomes can be very good for patients who have their melanomas removed during the disease's earliest stages, when the lesion is still thin and has not spread deep into the skin. To screen for melanomas, clinicians typically evaluate large skin surfaces with the ABCDE set of criteria, searching for "ugly duckling" lesions that show signs of being pre-cancerous. Authorities have also started to roll out large skin cancer screening programs to reduce the burden of melanoma, but clinics lack scalable tools that can assess lesions in large numbers of patients. Here, Luis Soenksen and colleagues designed a neural network platform that takes photographs of skin lesions - even those taken with a cell phone camera - and rapidly identifies suspicious markings that may need follow-up testing. The team trained their technology with 38,283 photographs, including skin photos from 133 patients, and observed the method distinguished suspicious lesions from nonsuspicious ones with a sensitivity and specificity of 90.3% and 89.9%, respectively. In a separate experiment, the strategy also ranked "ugly duckling" lesions on the skin of 68 patients, yielding rankings that mostly matched assessments from 3 dermatologists. The authors add that future improvements may help address some current limitations with the system, such as by making it work with a wider range of cameras, light settings, and photographers.
Science Translational Medicine