Positron Emission Tomography (PET) is an advanced imaging technique that vividly reflects human physiological activity and plays an indispensable role in diagnosing Alzheimer’s disease (AD) and cancer. However, PET imaging involves injecting radionuclides into the body, inevitably leading to radiation exposure. Reducing the dose of radionuclide used during imaging is crucial for safer and more cost-effective PET imaging. However, reducing the dose in PET acquisition can degrade image quality, potentially failing to meet clinical requirements. To maintain high-quality PET imaging while reducing the radionuclide dose, besides developing imaging systems to improve sensitivity, another effective approach is to generate Standard-dose PET (SPET) from Low-dose PET (LPET) by generative technologies. In this work, we propose a novel and effective approach to estimate high-quality SPET images from LPET images. Specifically, We employ a semi-supervised training framework to fully utilize both the rare paired and the abundant unpaired LPET and SPET images. Additionally, using this framework as a foundation, we introduce a Region-adaptive Normalization (RN) and implement a structural consistency constraint to address task-specific challenges. RN customizes normalization procedures for distinct regions within each PET image, mitigating adverse effects stemming from significant intensity variations across different areas. Simultaneously, the structural consistency constraint ensures the preservation of structural details throughout the process of generating SPET images from LPET images. With extensive experimental validation, our approach can achieve superior performance over state-of-the-art methods, and shows stronger generalizability to the dose changes of PET imaging