Artificial intelligence in dermatology: a systematic review of skin cancer detection and classification

Abstract

Abstract Skin cancer is one of the most common malignancies worldwide with melanoma being the deadliest form due to its high metastatic potential. Early and accurate detection is crucial for improving patient outcomes. Traditional diagnostic methods, including visual inspection and dermoscopic evaluation are limited by inter-observer variability and require substantial clinical expertise. In recent years, artificial intelligence (AI) has shown promise in enhancing dermatological diagnostics by automating skin lesion analysis with high accuracy. This systematic review aims to evaluate and synthesize current literature on the application of AI for the detection and classification of skin cancer focusing on current trends, AI technologies, data sources, clinical relevance, limitations and future direction. A comprehensive literature search was conducted across PubMed, Scopus, IEEE Xplore, and Web of Science for articles published between January 2010 and May 2025. The PRISMA guidelines were followed, and quality assessment was performed using the PROBAST tool. Out of 6296 identified records, 163 studies met the inclusion criteria. Convolutional neural networks (CNNs) were the most commonly used AI models, frequently trained on publicly available datasets such as ISIC and HAM10000. The reported classification accuracy ranged from 82% to 95%, with several models achieving dermatologist-level performance. However, the generalizability of these models was often limited due to dataset bias, lack of external validation, and insufficient reporting of demographic diversity. Few studies addressed model explainability or clinical integration. AI-based approaches demonstrate strong potential for enhancing the detection and classification of skin cancer with several models showing performance comparable to expert clinicians. Performance metrics including accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were discussed alongside the skin image datasets and the relevant AI tools used in the studies. However, significant challenges remain, including the need for standardized evaluation frameworks, diverse and well-annotated datasets, and rigorous clinical validation. Future research should prioritize model transparency, fairness, and integration into real-world clinical workflows to ensure safe and equitable deployment.

Ähnliche Arbeiten