Abstract
Fluorescence microscopy is an important imaging technique for biological research and applications. However, owing to various constraints in its engineering implementation and image acquisition, the acquired images often suffer from substantial noise. Although many denoising methods have been proposed, their applicability is limited by the distinct properties of fluorescence microscopy images. In particular, in real-world applications, it is often challenging and sometimes infeasible to acquire a large number of paired noisy-clean images for supervised training, and it is impractical to parameterize and estimate all types of real-world noise distributions. We propose ED-Diff, a zero-shot denoising algorithm based on diffusion priors. ED-Diff integrates the optimization solution of inverse problems with the diffusion sampling process and introduces a noisy image transformation module (NiTM) with an encoder-decoder structure to handle real-world noise scenarios. Extensive experiments on multiple real-world datasets validated the effectiveness of NiTM, demonstrating that ED-Diff exhibits competitive and robust performance.