Artificial Intelligence and Precision Dermatology across Inflammation, Cancer, and Infection: A Review

Abstract

Introduction: Artificial intelligence (AI) is rapidly transforming dermatology, with emerging roles in the triage, diagnosis, and monitoring of inflammatory, neoplastic, and infectious skin diseases. Translation to routine care is limited by dataset imbalance, domain shift, and inconsistent external validation, calibration, and fairness reporting. The aim of this study was to review current applications of AI in dermatology, evaluate the performance of AI systems across selected inflammatory, neoplastic, and infectious skin diseases, and identify key challenges and priorities for their safe, equitable, and effective clinical implementation.

Materials and Methods: This narrative review synthesizes recent evidence on AI workflows for eczema and psoriasis severity scoring (EASI/PASI), pigmented tumors (melanoma, basal cell carcinoma, naevi, and benign keratosis-like lesions), and selected infections (superficial mycoses and mpox). We summarize commonly used public, metadata-rich datasets (e.g., HAM10000, ISIC challenges, BCN20000, SIIM-ISIC 2020, PAD-UFES-20, Derm7pt) and methodological trends including convolutional and transformer backbones, detection and segmentation models, label-efficient pretraining, and multimodal/vision-language systems.

Results: Across disease groups, the most mature evidence supports pigmented-lesion classification, where discrimination is high on curated dermoscopy benchmarks and clinician-AI collaboration improves triage decisions. Segmentation-derived EASI and PASI estimates show strong correlation with clinician assessments for longitudinal monitoring. For superficial fungal infections and mpox, image-based models are approaching dermatologist-level performance, but generalization remains sensitive to capture device, setting, and population.

Conclusion: Key priorities for equitable and reliable precision dermatology include leakage-safe, skin-tone-balanced datasets, routine external and temporal validation with subgroup analyses, calibrated and uncertainty-aware outputs (including selective referral strategies), privacy-preserving deployment (e.g., on-device inference and federated learning), and adherence to standardized health-AI reporting frameworks.