Multiclass convolutional neural network vs. 96 dermatologists in skin lesion diagnosis, an international study.

BACKGROUND AND OBJECTIVES: Artificial intelligence was shown to improve diagnostic accuracy for skin cancer detection. While most clinically approved models provide binary "benign/malignant" classifications, multiclass predictions may offer greater clinical utility. Yet, comparisons between multiclass convolutional neural networks (CNNs) and dermatologists are scarce. METHODS: In an international web-based reader study, dermatologists (n = 96) and a prototype multiclass CNN (FotoFinder Systems, Germany) assigned diagnoses to 100 skin lesions by using nine disease categories (melanoma; basal cell carcinoma; squamous cell carcinoma; intraepithelial carcinoma; melanocytic nevus; benign keratinocytic lesion; dermatofibroma; vascular lesion; "other"). The main outcome measure was overall mean sensitivity (micro-averaged percentage of correct diagnosis) of dermatologists versus multiclass CNN. RESULTS: Dermatologists achieved an overall mean sensitivity (95% CI) of 69.4% (68.4%-70.3%) with dermatoscopy-only data (level-I), which improved to 76.0% (75.2%-76.9%) when provided with full clinical information (level-II). The CNN's top-rank predictions showed a higher overall mean sensitivity of 82.0% (73.3%-88.3%). The CNN significantly outperformed all dermatologist subgroups, except for dermatoscopy experts at study level-II (overall mean sensitivity 81.1% [79.9%-82.3%], pequivalence = 0.0085). CONCLUSIONS: Multiclass CNN predictions outperformed most dermatologists in diagnostic accuracy, supporting their potential to enhance clinical decision-making, particularly in settings with limited dermatological expertise.