A comprehensive review of deep learning applications in the segmentation and classification of skin cancer

Skin cancer (SC) is one of the most prevalent forms of cancer worldwide. Both melanoma and non-melanoma types pose major challenges for early detection, accurate diagnosis, and proper treatment. Conventional diagnostic approaches, such as biopsy and visual examination, are often time-consuming, subjective, and prone to human error. Recent advances in artificial intelligence (AI) and deep learning (DL) have greatly improved the accuracy of SC diagnosis. This systematic review explores the applications of DL techniques in the segmentation and classification of skin lesions between 2014 and 2024. Following the Preferred Reporting Items for Systematic reviews and Meta-Analyses guidelines and applying predefined inclusion and exclusion criteria, a total of 77 experimental studies out of 540 were analyzed from major databases, including Scopus, IEEE, PubMed, and MDPI. Convolutional neural networks (CNNs) were identified as the most widely used for classification, while U-Net and its variants dominated segmentation tasks. Hybrid and ensemble frameworks demonstrated superior performance on benchmark datasets such as the International Skin Imaging Collaboration (ISIC) archive and HAM10000. Moreover, this work incorporates a formal risk-of-bias analysis, revealing critical concerns about class imbalance and data leakage. Almost all reviewed studies for the classification task achieved an average accuracy of 96% for the ISIC dataset, while the HAM10000 dataset attained an average accuracy of 93%. Despite these advances, challenges such as class imbalance, limited dataset diversity, and insufficient clinical validation persist. Addressing these issues through data augmentation, explainable AI, and federated learning could further enhance the generalizability and clinical applicability of AI-driven diagnosis systems. Additionally, this study identifies a clear paradigm shift from standalone CNNs to hybrid frameworks and multi-source feature fusion strategies, aiming to improve SC diagnosis.