Artificial intelligence in contrast-enhanced mammography: Current evidence, clinical integration, and future perspectives – a narrative review

Abstract

Introduction: Contrast enhanced mammography (CEM) combines standard mammographic imaging with functional enhancement patterns and is increasingly used in diagnostic and staging pathways. Artificial intelligence (AI) has shown strong potential in breast imaging, yet its application to CEM remains limited. Objectives: This narrative expert review summarizes current evidence on AI applied to CEM, identifies methodological gaps, and outlines future directions for clinically integrated, precision-oriented models. Methods: A narrative review of peer‑reviewed studies was conducted, focusing on AI for lesion detection, classification, segmentation, radiomics, workflow support, and emerging multidimensional frameworks involving CEM. Given the heterogeneity of available studies and the early developmental stage of the field, a narrative expert review was considered more appropriate than a protocol‑driven systematic approach. Results: Existing studies demonstrate that AI can improve lesion conspicuity, support automated detection, and extract quantitative features from CEM. Reported performance metrics—such as area‑under‑the‑curve (AUC) values between 0.87 and 0.93 in small datasets—indicate promising diagnostic potential. However, progress is limited by small sample sizes, heterogeneous acquisition protocols, variability in background parenchymal enhancement, lack of external validation, and the absence of models incorporating clinical, hormonal, or multimodal information essential for real‑world interpretation. Emerging frameworks suggest that AI may also support multidimensional risk profiling by integrating imaging features with patient‑specific variables. Conclusions: AI has the potential to enhance the diagnostic and clinical value of CEM, but meaningful implementation will require standardized acquisition, larger multicenter datasets, explainable architectures, and integration of clinical context. Future models should reflect real-world senological workflows and support personalized, risk‑stratified decision‑making.