Abstract: Significant variations exist among positron emission tomography (PET) images acquired at different dose levels, leading to severe performance degradation in deep learning models. Current multidose PET image synthesis methods primarily focus on extracting differential information and spatialdomain features from varying-dose PET images, while neglecting their shared information and frequencydomain features. To address these limitations, we propose a dual-domain information adaptive learning approach for multi-dose PET image synthesis (SFMdose). The proposed method aims to develop a unified model capable of simultaneously extracting both shared knowledge and discriminative information across PET images with different doses. Moreover, we design a spatial-domain information guidance module to learn dose-related patterns and adaptively selects propagation routes to suppress dose-specific interference. Additionally,a frequency-domain guidance module is proposed to capture dose distribution characteristics in the frequency domain, enabling the model to better learn features from images with different doses and improve synthesis quality. Extensive experiments on both the UDPET dataset and real human brain datasets demonstrate that our SFMdose method achieves superior performance in synthesizing highquality PET images across multiple dose levels.