Physics-informed self-supervised denoising for fluorescence microscopy via structure and intensity priors

Abstract

Summary Fluorescence microscopy images contain smooth, low-intensity backgrounds dominated by random noise and brighter biological signals with complex textures. We incorporate these structural and intensity characteristics into self-supervised denoising through two complementary modules. Structure-adaptive probabilistic masking (SAPM) uses local variance to reduce masking in textured regions while increasing masking in redundant backgrounds. Intensity-aware weighted loss (IA-Loss) reweights optimization according to local brightness to improve preservation of high-intensity signals under signal-dependent noise. Using Noise2Void (N2V) as the baseline, we developed structure- and intensity-prior-guided Noise2Void (SI-N2V) with a Mamba-UNet backbone for global contextual modeling. Experiments on an in-house pathological dataset and public fluorescence microscopy datasets show that SAPM and IA-Loss independently improve denoising and achieve the largest gains when combined. IA-Loss also consistently improves several self-supervised denoising frameworks, supporting its use as a plug-and-play optimization module.