Advanced deep learning for digital holography: A review

Deep learning (DL) has become an important tool in digital holography (DH), enabling data-driven learning of complex inverse mappings directly from holographic measurements. Early DL-assisted DH demonstrated the feasibility of this approach but often relied on large labeled datasets and lacked explicit physical interpretability. In recent years, advances in DL methodologies have begun to address these limitations. This review examines three major developments that are reshaping the field: innovative training strategies for data-efficient learning, the integration of physical priors for improved interpretability and physical consistency, and advanced network architectures capable of extracting richer representations from holographic data. By organizing recent progress from this methodological perspective, we provide an updated overview of how modern DL techniques are expanding the capabilities of DH. We conclude with a forward-looking perspective, identifying promising but underexplored DL directions expected to drive significant future advancements in DH applications.