ARTIFICIAL INTELLIGENCE IN DERMATOLOGICAL IMAGE CLASSIFICATION: DIAGNOSTIC ACCURACY, SKIN TONE DISPARITIES, AND CLINICAL GOVERNANCE

Abstract

Background: Artificial intelligence (AI) is increasingly used to classify skin lesions from dermoscopic and clinical images, yet high performance in curated datasets may not translate into safe or equitable clinical use. Objective: This structured narrative review evaluates diagnostic accuracy, skin-tone-related performance disparities, dataset representation, and governance requirements for AI-based skin-lesion classification. Methods: PubMed/MEDLINE was used as the primary database, supplemented by PubMed Central, Google Scholar, Semantic Scholar, journal websites, and reference-list screening. English-language, peer-reviewed literature published mainly between 2021 and 2026 was synthesised thematically across diagnostic performance, external validity, skin-tone representation, clinical implementation, and governance. Results: Meta-analytic evidence indicates high sensitivity for AI-based skin-lesion classification, including pooled sensitivity of 0.91 and an HSROC-AUC of 0.88, but lower specificity and substantial heterogeneity limit direct clinical translation. Performance declined in darker skin-tone groups in several external evaluations; in selected models, sensitivity fell between Fitzpatrick types I–II and V–VI. Limited reporting of skin tone, reliance on internally tested or curated datasets, and sparse prospective evidence further constrain generalisability. Conclusion: AI should be implemented as clinician-supervised decision support rather than as an autonomous diagnostic system. Equitable deployment requires representative datasets, external and subgroup-stratified validation, transparent reporting, defined clinical accountability, regulatory oversight, and continuous post-deployment monitoring.