Deep Learning for Automatic Skin Cancer Detection and Classification: A Narrative Review of Architectures and Performance

Abstract

Skin cancer, including melanoma and non-melanoma subtypes such as basal cell carcinoma and squamous cell carcinoma, is among the most common malignancies worldwide, and clinical outcomes depend strongly on early and accurate diagnosis. Visual dermoscopic examination by dermatologists remains the diagnostic standard but is time-consuming, subjective, and reliant on specialist availability, motivating the development of automated, image-based detection systems. This paper presents a structured literature review of deep learning-based approaches for automatic skin cancer detection and classification, anchored on Gururaj et al.'s (2023) IEEE Access study, DeepSkin, as the base paper, and extended through a synthesis of related peer-reviewed literature spanning convolutional neural networks (CNNs), transfer learning, ensemble architectures, spiking neural networks, and cloud-based deployment models, evaluated primarily on the HAM10000, ISIC, and DermIS dermoscopic image datasets. Across the reviewed studies, transfer-learning-based CNN architectures (e.g., DenseNet, ResNet, VGG, EfficientNet) consistently outperformed classical image-processing pipelines, with reported classification accuracies ranging from approximately 83% to above 94%, and one multi-classification ensemble network reporting an area-under-the-curve value of 99.43%. Ensemble and hybrid architectures that combine multiple pretrained backbones, or that integrate segmentation with classification, generally reported higher and more stable performance than single-backbone models. The reviewed evidence indicates that deep learning, and particularly transfer-learning-based CNN pipelines, is an effective and increasingly mature approach for automated skin cancer detection, though challenges of class imbalance, dataset diversity across skin tones, interpretability, and clinical validation remain open. Future work should prioritize prospective clinical evaluation and standardized, demographically diverse benchmarking.