Self-Supervised Denoising Under Variations Between Adjacent Slices for Fluorescence Microscopy Image Stacks

Fluorescence microscopy is a critically important imaging technique for biological studies, enabling researchers to vi-sualize cellular structures and processes with high resolution and molecular specificity. Because the presence of noise in fluorescence microscopy is inevitable in practice, image de-noising is required in a wide variety of applications. Considering the challenge of acquiring paired images for supervised training of denoising models, this work proposes a self-supervised denoising algorithm. The algorithm is inspired by Noise2Noise, utilizing the similarity between adjacent slices in microscopy image stacks to construct training pairs. Concurrently, it incorporates additional regularization terms to mitigate the influence of variations in the underlying signals between adjacent slices. The algorithm has been tested on publicly available real-world datasets, and its effectiveness was validated through a series of comparative experiments and ablation studies. The code for this work is openly accessible at https://github.com/chill-luo/SDUV.