Simple Summary
Cutaneous melanoma is a potentially lethal skin cancer that can be difficult to diagnose accurately, especially in early or ambiguous cases. Traditional histopathology relies on expert evaluation of tissue slides, but this process is subjective and prone to variability. With the rise of digital pathology and artificial intelligence (AI), there is growing interest in using computational tools to assist melanoma diagnosis. This review explores how AI—particularly deep learning and interpretable models—can analyze digital slides, extract diagnostic features, and even predict genetic mutations from routine images. By summarizing recent advances across classification, spatial modeling, and explainable AI, this work highlights how these tools can improve diagnostic accuracy, reduce workload, and support decision-making. Our goal is to inform researchers, clinicians, and pathologists of the current state of AI-assisted melanoma diagnostics and guide future studies toward more robust, clinically integrated solutions.
Abstract
Background: Cutaneous melanoma (CM) poses significant diagnostic challenges due to its biological heterogeneity and the subjective interpretation of histopathologic criteria. While early and accurate diagnosis remains critical for patient outcomes, conventional pathology is limited by interobserver variability and diagnostic ambiguity, especially in borderline lesions. Objective: This narrative review explores the integration of digital pathology (DP) and artificial intelligence (AI)—including deep learning (DL), machine learning (ML), and interpretable models—into the histopathologic workflow for CM diagnosis. Methods: We systematically searched PubMed, Scopus, and Web of Science (2013–2025) for studies using whole slide imaging (WSI) and AI to assist melanoma diagnosis. We categorized findings across five domains: WSI-based classification models, feature extraction (e.g., mitoses, ulceration), spatial modeling and TIL analysis, molecular prediction (e.g., BRAF mutation), and interpretable pipelines based on nuclei morphology. Results: We included 87 studies with diverse AI methodologies. Convolutional neural networks (CNNs) achieved diagnostic accuracy comparable to expert dermatopathologists. U-Net and Mask R-CNN models enabled robust detection of critical histologic features, while nuclei-level analyses offered explainable classification strategies. Spatial and morphometric modeling allowed quantification of tumor–immune interactions, and select models inferred molecular alterations directly from H&E slides. However, generalizability remains limited due to small, homogeneous datasets and lack of external validation. Conclusions: AI-enhanced digital pathology holds transformative potential in CM diagnosis, offering accuracy, reproducibility, and interpretability. Yet, clinical integration requires multicentric validation, standardized protocols, and attention to workflow, ethical, and medico-legal challenges. Future developments, including multimodal AI and integration into molecular tumor boards, may redefine diagnostic precision in melanoma.