Deep Learning for Breast Cancer Imaging in Public Health

Breast cancer imaging is central to early detection, yet current screening and diagnostic modalities remain limited by breast density, interpretive variability, and suboptimal specificity. This thesis investigates how deep learning can enhance breast cancer detection and diagnosis across mammography, digital breast tomosynthesis, and breast magnetic resonance imaging. Novel architectures are proposed to improve the detection of early radiographic indicators such as microcalcifications and to exploit long-range contextual information using transformer-based models. To address the challenges of volumetric imaging, the thesis introduces interpretable frameworks that condense three-dimensional information into synthetic two-dimensional representations suitable for clinical workflows. In breast MRI, a deep learning–based strategy is presented to suppress confounding vascular enhancement, improving lesion visibility and diagnostic clarity. Collectively, the proposed methods demonstrate how interpretable, domain-informed deep learning can support more accurate and transparent breast cancer care.