Amidst the swift progression of neural networks, numerous strategies have been put forward to address the issue of image denoising under the strong supervision of large-scale datasets. Leveraging self-supervised blind-spot networks to tackle spatially correlated noise in real-world scenarios presents a significant and daunting obstacle. In this paper, we introduce an innovative method for denoising based on multi-branch blind-spot network with PD-random replace refinement (MBPDR3) to overcome information loss. Extensive research has shown that the proposed method is substantially superior to other self-supervised methods, and the efficacy of the proposed approach is demonstrated through extensive experimentation, yielding superior performance on both synthetic and real datasets.