Deep Learning Approaches for Skin Cancer Detection: A Systematic Literature Review on Architectures, Multimodal Fusion Strategies, and Open Research Challenges

Skin cancer is among the most common and fastest-growing forms of cancer worldwide, and early, accurate detection substantially improves treatment outcomes. Over the past decade, deep learning has emerged as the dominant tool for automated skin lesion classification, evolving from single convolutional neural networks to multimodal ensembles that fuse dermoscopic images with patient metadata. This paper presents a systematic literature review of 26 peer-reviewed and DOI-indexed studies (2002-2026) on deep learning for skin cancer detection, selected and screened following the PRISMA 2020 methodology. The reviewed literature is organised into eight thematic categories: convolutional neural network (CNN)-based approaches, transformer-based architectures, multimodal fusion strategies, class-imbalance handling, explainability, model efficiency, benchmark datasets, and privacy-preserving learning. A comparative analysis of methods, datasets, and reported accuracy is presented, followed by a structured discussion of eight open research gaps, most notably data leakage through metadata features, improper patient-wise data splitting, and the lack of fair, class-balanced evaluation in high-performing pipelines. The review concludes that while reported accuracy in this field continues to rise, trustworthy evaluation methodology has not kept pace, and outlines concrete directions for future work.