Iterative self-supervised denoising for low-SNR fluorescence microscopy

Low-signal fluorescence microscopy is fundamentally limited by the trade-off among photon budget, imaging speed, photobleaching, and phototoxicity. Although supervised deep denoising can effectively improve image quality, it typically requires paired noisy-clean data that are difficult to obtain in routine biological imaging. In this work, we present a selfsupervised denoising framework for fluorescence microscopy based on progressive target updating. Starting from synthetically degraded noisy observations, the method iteratively replaces the training targets with restored outputs from previous rounds, thereby improving supervision quality without requiring clean references. A U-Net backbone with residual prediction is adopted as the default implementation, and mixed Poisson-Gaussian corruption is used during training to improve robustness under different acquisition conditions. Experiments on widefield F-actin images show clear stage-wise improvement during training, and the proposed method achieves substantially better restoration than Noise2Noise and block-matching and 3D filtering while approaching the performance of fully supervised Noise2Clean. Additional results on real noisy widefield images demonstrate improved detail preservation, and transfer experiments on STED data further indicate effective denoising and improved decorrelation-based resolution estimates. The proposed framework provides a practical strategy for low-SNR fluorescence microscopy with reduced dependence on clean training data.