Self-supervised learning for breast cancer detection: A review

Breast cancer is one of the most prevalent and life-threatening malignancies worldwide, prompting continuous improvements in both early detection and precise diagnosis. Deep learning methods have shown promise in computer-aided detection (CAD) systems, yet they often rely on extensive labeled datasets, which are costly and time-consuming to annotate. Self-supervised learning (SSL) offers a compelling alternative by leveraging large volumes of unlabeled medical images to learn robust feature representations. In this review, we analyze the application of SSL methods across the major stages of the breast cancer detection pipeline: screening, diagnosis, grading and staging. We focus on the predominant imaging modalities, including mammography, digital breast tomosynthesis, ultrasound, MRI, and histopathology, outlining how SSL can reduce annotation demands, enhance generalization under domain shift, and improve lesion localization and segmentation. Our survey reveals a growing number of successful SSL applications in mammography and ultrasound for early detection, as well as in MRI and histopathology for more detailed characterization of suspicious lesions. Notably, however, we find a marked absence of SSL-based research in PET imaging, despite its recognized value in diagnostic and staging contexts. We discuss the potential of extending SSL to these underexplored modalities and highlight future research directions, including multi-modal data fusion, domain-adaptive pretext tasks, and explainability-driven models for clinical integration. Overall, this review underscores the transformative role SSL can play in breast cancer imaging, offering scalable solutions with fewer reliance on expert annotations and the promise of improved detection and prognostication.