Towards a clinically integrated artificial intelligence tool for triage of skin cancer.

Skin cancer is one of the most prevalent malignancies worldwide, with early detection critical to improving outcomes. However, in low-resource settings such as rural regions, limited access to dermatologists and diagnostic tools delays diagnosis. This study presents a clinical validation of an artificial intelligence (AI)-based mobile application to assist generalist healthcare professionals in skin lesion triage. The tool classifies lesions into five priority levels according to malignancy risk, following a protocol developed by dermatologists. The AI model, based on a fine-tuned MobileNet-V3 architecture, was trained on the PAD-UFES-20+ dataset (13,569 images) and integrated into an offline-capable mobile app. Internal validation achieved a sensitivity of 0.79, specificity of 0.95, and AUC of 0.95. In clinical validation across two phases, 131 healthcare professionals from nine cities participated. In Phase 1, sensitivity and specificity improved from 0.64 and 0.90 (without AI) to 0.80 and 0.92 (with AI), matching the model's standalone performance. In Phase 2, with 57 community health workers in rural areas, AI assistance increased triage effectiveness by 17% and reduced unnecessary referrals by 30%. Participant feedback was positive, highlighting increased safety and confidence. These results indicate that AI-assisted mobile triage tools can enhance diagnostic performance and healthcare efficiency in resource-limited environments.