Deep Multimodal Fusion for Breast Cancer Classification: Taxonomy, Datasets and Open Challenges

Breast cancer remains one of the leading causes of mortality among women worldwide, necessitating early and accurate detection for improved patient outcomes. Deep learning has significantly advanced breast cancer classification; however, unimodal approaches relying on a single source of information, such as imaging or genomics, often fail to capture the complex heterogeneity of the disease. Multimodal fusion offers a promising solution by integrating diverse data sources, including histopathology, radiology, genomics, and clinical records, to enhance predictive performance and reliability. This paper presents a comprehensive taxonomy of deep multimodal fusion approaches for breast cancer classification, categorizing fusion at feature, representation, and decision levels while highlighting state-of-the-art DL strategies such as convolutional neural networks (CNNs), transformers, and hybrid models. Additionally, we review key datasets spanning imaging, omics, and electronic health records, discussing their strengths, limitations, and integration challenges. Open issues such as data scarcity, standardization, interpretability, and privacy are critically examined, alongside emerging directions in explainable AI, federated learning, and edge computing. By consolidating methodologies, datasets, and research gaps, this study provides an organized framework to advance multimodal fusion techniques toward clinically viable breast cancer diagnostic systems.