Deep Learning in Multimodal Breast Cancer Imaging: From Image Reconstruction and Segmentation to Diagnosis and Treatment Response Prediction

Breast cancer imaging is central to screening, diagnosis, staging, treatment monitoring, and post-treatment surveillance, but image interpretation remains limited by variable image quality, interobserver variability, false-positive findings and heterogeneous tumor biology. This narrative review summarizes current applications of deep learning in multimodal breast cancer imaging, with emphasis on image reconstruction, image enhancement, lesion detection, segmentation, classification, biomarker prediction, treatment response assessment, prognosis and clinical implementation. A structured literature search was performed across major biomedical and technical databases, focusing on studies involving mammography, digital breast tomosynthesis, ultrasound, MRI, PET/CT, digital pathology and multimodal fusion approaches. Current evidence indicates that deep learning can support image denoising; super-resolution, low-dose, and accelerated reconstruction; lesion localization; tumor segmentation; benign–malignant classification; and molecular or biomarker-related prediction. Multimodal models integrating radiological imaging, histopathology, clinical variables, and molecular markers show particular promise for treatment response prediction, recurrence risk estimation, and personalized decision support. However, clinical translation remains limited by retrospective study designs, small and imbalanced datasets, domain shift, inconsistent annotations, limited explainability, bias, lack of prospective validation, and regulatory challenges. Deep learning should therefore be viewed as a decision-support framework that may improve breast cancer imaging workflows if validated in diverse, prospective, and clinically representative settings.