Uncertainty-Aware Deep Learning for Skin Cancer Diagnosis

Abstract

Skin cancer is one of the most common cancers worldwide. Timely detection substantially improves prognosis, particularly for melanoma. Deep learning models based on convolutional neural networks, Transformers, ensembles, and multimodal learning have advanced automated skin lesion classification and segmentation; however, most remain deterministic and provide no estimate of predictive confidence, which limits their safe use in clinical settings. This systematic literature review examines uncertainty-aware deep learning for skin cancer classification and segmentation across 64 studies published between 2022 and 2026. The reviewed approaches are categorized and compared across Bayesian deep learning, Monte Carlo Dropout, entropy-guided learning, deep ensembles, Transformer-based architectures, explainable artificial intelligence, multimodal learning, federated learning, active learning, and semi-supervised learning, together with quantification techniques such as predictive entropy, mutual information, Expected Calibration Error, conformal prediction, and test-time augmentation. The synthesis shows that although recent models achieve high diagnostic accuracy, uncertainty reporting remains limited, uncertainty thresholds are rarely validated, calibration practices are inconsistent, and external clinical evaluation is scarce. The review maps these gaps to the formulated research questions and outlines future directions toward calibrated, explainable, privacy-preserving, and clinically validated uncertainty-aware dermatological AI systems.