Multimodal and Uncertainty-Aware Deep Learning for Skin Cancer Detection: A Review

Skin cancer is one of the most prevalent and increasing cancers globally, and melanoma is the most aggressive type of skin cancer. While deep learning has been able to approach dermatologist-level accuracy in diagnosis using only an image, current methods mostly focus on image-only predictions, overlooking other diagnostic features used to inform decisions. This review paper examines recent developments in multimodal and uncertainty quantification deep learning techniques for skin cancer diagnosis between 2018 and 2025. The authors discuss image-only, multimodal fusion (dermoscopic image and meta-data), uncertainty quantification and explainable artificial intelligence (XAI) methods. The results show that multimodal models achieve better performance compared with image-only models by incorporating additional meta-data, such as age, lesion site and family history of melanoma, which boost performance and sensitivity to melanoma. Sophisticated multimodal fusion approaches such as cross-attention and gating also help. Predictive uncertainty, such as Bayesian neural networks and Monte Carlo dropout, helps with confidence in predictions, and explainability helps with interpretability. But there are issues with class imbalance, inconsistencies in the evaluation protocol, patient-based validation and diversity of data. This article addresses these problems and suggests future research for clinically relevant AI models