Artificial intelligence vs. dermoscopy for malignancy risk stratification of pigmented skin lesions: a systematic review and meta-analysis for public health

Abstract

Background Accurate risk stratification of pigmented skin lesions is essential for early melanoma detection and for reducing unnecessary excisions. Although artificial intelligence (AI) is increasingly applied to dermoscopic image analysis, its diagnostic performance compared to dermoscopy remains unclear. Objective This study aimed to compare the diagnostic performance of AI, dermoscopy, and AI-assisted clinicians for malignancy risk stratification of pigmented skin lesions. Methods PubMed, Embase, Web of Science, and the Cochrane Library were systematically searched for studies evaluating AI, dermoscopy, or AI-assisted clinicians in diagnosing pigmented or melanoma-suspected skin lesions. Diagnostic performance metrics were calculated from extracted or reconstructed data, and study quality was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 tool (QUADAS-2) and QUADAS-comparative (QUADAS-C). Results A total of 2,571 records were identified, and 10 studies were included in the primary quantitative analysis, contributing 17 diagnostic arms. These included 10 dermoscopy arms, 6 AI-alone arms, and 1 AI-assisted clinician arm. In the dermoscopy group, sensitivity ranged from 0.418 to 0.966, and specificity ranged from 0.293 to 0.975. In the AI group, sensitivity ranged from 0.164 to 0.968, and specificity ranged from 0.374 to 0.983. AI-assisted clinicians showed a sensitivity of 1.000 and a specificity of 0.837 in the single available study. Overall, AI and dermoscopy showed overlapping diagnostic performance, although substantial variability was observed across AI algorithms and clinical settings. Deeks’ funnel plots showed no evidence of significant publication bias in either the AI group or the dermoscopy group. Conclusion Autonomous AI showed diagnostic performance broadly comparable to dermoscopy. Although AI achieved slightly higher pooled specificity, its sensitivity was lower, indicating no clear clinical advantage over conventional dermoscopic assessment. Evidence on AI-assisted clinicians was limited to a single diagnostic arm; therefore, this finding should be interpreted as hypothesis-generating rather than as evidence to support immediate clinical implementation. Further comparative studies using standardized AI-assisted workflows, comparable physician expertise, and patient-centered outcomes are required before practical recommendations can be made. Systematic review registration Registered with PROSPERO (CRD420261389834).