Medical imaging foundation models, as an important direction in the intelligent analysis of medical images, are driving a paradigm shift in this field—from task-specific modeling for individual tasks toward a new paradigm grounded in large-scale pre-training and general-purpose representation learning. These models have demonstrated considerable potential in multimodal information integration, cross-task transfer, multi-scenario adaptation, and generative interaction. However, most current studies remain based on benchmark datasets, retrospective cohorts, or controlled experimental conditions, and still face challenges such as insufficient data representativeness, task settings that deviate from real-world clinical workflows, inadequate validation of generalizability and robustness, and ambiguous boundaries of ethical oversight and accountability. This article reviews the research background, key technologies, and advances in the applications of medical imaging foundation models, and further discusses practical challenges including data governance, workflow integration, and ethical regulation, with the aim of providing guidance for related research and clinical translation.