Generalizable Unsupervised Microscopy Video Denoising via Weighted SpatioTemporal Sampling

Medical video denoising is essential for improving image quality and enhancing the reliability of clinical data. However, the limited availability of annotated datasets, the variability of noise patterns, and the absence of ground-truth reference frames present significant challenges for traditional supervised denoising approaches. On the other hand, current unsupervised video denoising methods often struggle to balance noise removal and motion preservation, leading to either excessive smoothing that degrades fine details or insufficient denoising that leaves residual noise. To address these challenges, we propose STS-UVD, a novel unsupervised video denoising (UVD) method that removes noise while preserving motion integrity. By refining the optical flow, our method ensures temporal consistency without compromising important motion details. Additionally, STS-UVD demonstrates strong generalization across different noise conditions and datasets, making it a robust solution for medical video analysis. Extensive experiments validate its effectiveness in enhancing video quality while maintaining structural and temporal coherence.