Predicting breast cancer recurrence using multimodal deep learning and MRI: Systematic review

Abstract

Breast cancer is one of the most common cancers in women worldwide affecting approximately 2.3 million women annually and causing 685,000 deaths each year. Thanks to screening and treatment, the 5-year survival rate exceeds 90% when detected early. However, recurrence remains a major challenge, occurring in 20-30% of cases often 2–5 years after initial diagnosis. Current methods (physical examinations, mammograms, clinical scores) often miss subtle signals, highlighting the urgent need for reliable, non-invasive risk stratification tools. This paper presents a systematic review of deep learning approaches for predicting breast cancer recurrence from multimodal Magnetic Resonance Imaging (MRI) and clinicopathological data, according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines. Eight studies published between 2021 and 2025 were included. Emerging trends include three-dimensional (3D) volumetric fusion networks, transformer-based temporal modeling, and radiogenomic correlation with Oncotype DX. Key challenges include heterogeneity in MRI acquisition protocols, scarcity of annotated longitudinal datasets, and insufficient model interpretability. This review maps current methodological trends, identifies research gaps, and provides a structured roadmap for clinical deployment.