Document Type : Systematic Review and Meta-analysis
Authors
1
Zakout hudklinikk AS, Kongens gate 30, 4610 Kristiansand, Norway.
2
Department of Information Technology, Faculty of Engineering and Information Technology, Al-Azhar University, Gaza, Palestine.
3
School of Electronics and Computer Science, Faculty of Engineering and Physical Sciences, University of Southampton, SO17 1BJ, UK.
Abstract
Background: Skin cancer is the most common malignancy worldwide and can be lethal if not detected early, especially melanoma. Primary care is often the first point of contact, but limited dermoscopy expertise and increasing service pressure can delay diagnosis. Artificial intelligence (AI) offers a route to expand access to earlier detection by supporting assessment and triage decisions in primary care. Methods: We conducted a narrative review of peer-reviewed literature on AI systems for skin-lesion analysis, including traditional machine-learning pipelines, convolutional neural networks (CNNs), vision transformers, ensemble and hybrid models, and mobile/on-device tools. We prioritized studies that compared AI with clinicians, reported diagnostic performance and calibration, evaluated subgroup performance (e.g., skin tone and device), and examined implementation in primary care, teledermatology, or real-world outpatient workflows. Results: AI systems can achieve high sensitivity for melanoma and other malignant lesions on curated image sets, often approaching specialist-level discrimination. Reader studies and pragmatic evaluations suggest AI assistance can improve non-dermatologists’ sensitivity and diagnostic confidence, but specificity, calibration, and robustness across devices and under-represented populations remain variable. Recent deployment-focused evidence highlights risks of automation bias, alert fatigue, and inequitable performance, underscoring the need for careful threshold selection, external validation, and post-deployment monitoring. Conclusion: AI-enabled skin cancer detection can support earlier diagnosis and more efficient triage in primary care, but clinical benefit depends on evaluation beyond headline accuracy. Diverse training data, robust external and prospective validation, explicit calibration and equity reporting, and governance frameworks with human oversight and continuous monitoring are essential for safe implementation.
Keywords
Main Subjects
)