When to refuse cosmetic treatment - AI-decision support for differentiating melanocytic and non-melanocytic facial lesions.

Abstract

Background: Flat pigmented skin lesions in chronically sun-exposed face and scalp areas present diagnostic challenges. Particularly, solar lentigo, pigmented actinic keratosis and lentigo maligna show overlapping features. Erroneous laser treatment of lentigo maligna for cosmetic reasons will inevitably delay diagnosis and compromise patient safety. This study assessed the performance of a market-approved convolutional neural network (CNN) versus dermatologists in differentiating melanocytic from non-melanocytic face and scalp lesions (FSL).

Patients and methods: Cross-sectional two-level reader study using 100 histologically validated FSL (level I: dermoscopy only; level II: expanded case information). The primary outcome was the diagnostic accuracy of the CNN in comparison to 64 dermatologists reviewing expanded case information.

Results: The CNN differentiated melanocytic and non-melanocytic FSL with an accuracy (95% CI) of 79.0% (69.7%-86.5%) and a ROC-AUC of 0.847 (0.772-0.923). Dermatologists reached a mean accuracy of 66.7% (64.7%-68.7%) at level I, which increased to 73.6% (72.1%-75.1%; p < 0.001) with expanded case information (level II). Compared with dermatologists at level II, the CNN demonstrated a significantly higher accuracy (79.0% vs. 73.6%; p < 0.001).

Conclusion: Our findings suggest that a CNN trained to differentiate melanocytic from non-melanocytic FSL may improve clinicians' diagnostic accuracy when used as an AI-based decision support tool for determining whether cosmetic laser procedures could safely be performed.