Self-supervised denoising method uses the inherent attributes of the data as a supervised signal to train the model without clean images. However, existing methods generally rely on pixel-level spatial domain, which makes the model unable to adaptively distinguish and excessively suppress effective features. Moreover, the locality of the convolution makes it difficult to decouple the global noise with long-range correlation. In this paper, we propose an approach that utilizes Fourier insights to interact with spatial domains. Firstly, through the Fourier transform we explored the physical property that amplitude mainly encodes noise and phase contains structural information. Based on this prior, Amplitude Spectrum Adaptive Weight (ASAW) module captures global noise distribution and explicitly suppresses the noise-dominated frequency band, while maintaining the integrity of the phase to avoid image distortion. Additionally, Multi-Scale Fusion (MSF) dynamically adjusts the contribution of different branches. We also propose Artifact Suppression Filtering (ASF), which suppresses artifacts via adaptive low-pass filtering, and preserves edge sharpness through intensity-aware weighting. Extensive experiments on SIDD and DND datasets verify that our method is significantly better than the existing self-supervised denoising methods and achieves high efficiency.