Editors' ChoiceDERMATOLOGY

Diagnosing skin diseases using an AI-based dermatology consult

See allHide authors and affiliations

Science Translational Medicine  17 Jun 2020:
Vol. 12, Issue 548, eabc8946
DOI: 10.1126/scitranslmed.abc8946

Abstract

Deep learning outperforms general practitioners in diagnosing 26 common skin conditions.

Skin conditions are extremely common, but with hundreds of different potential diagnoses and a shortage of dermatologists, primary care physicians and nurse practitioners may need help making the right decisions. Liu et al. set out to determine whether artificial intelligence could identify a shortlist of conditions for caregivers to consider. Using a deep-learning approach, they trained a neural network to diagnose 26 of the most common skin conditions by exposing it to 16,114 cases collected from a teledermatology practice. Rather than rely on a single photograph, each case was associated with multiple photographs, allowing nuances to be learned by the algorithm. They then compared the accuracy of their artificial intelligence (AI) diagnostic to board-certified dermatologists, primary care physicians, and nurse practitioners using 963 validation cases. The algorithm made the correct diagnosis 66% of the time, on par with dermatologists (63%) and superior to primary care physicians (44%) and nurse practitioners (40%). However, when asked to provide a differential diagnosis that included the three most likely diagnoses, the algorithm’s accuracy rose to 90%, outperforming even dermatologists (75%). The diversity of cases used in the training set, in terms of both skin conditions and patient demographics, made the tool similarly accurate regardless of age or race.

Although AI has previously been trained to diagnose specific conditions, the ability of a single tool to diagnose 26 distinct conditions could transform how skin conditions are triaged. This could be particularly impactful for rural communities, where dermatology referrals may be difficult, by allowing primary care providers decision support to determine an immediate course of action. In cases where the top three conditions reported by the tool share a common treatment, care can begin immediately. In cases where the top three conditions differ in severity or recommended treatments, targeted follow-up diagnostics can be performed to hasten the patient’s access to the right treatment. Although this tool is not yet approved for clinical use, deep-learning–based diagnostics and clinical decision support tools are gaining acceptance in many medical specialties and are poised to change how we experience medicine.

Highlighted Article

View Abstract

Stay Connected to Science Translational Medicine

Navigate This Article