Unsupervised Microscopy Video Denoising

TL;DR

This work tackles unsupervised denoising of microscopy videos captured at fixed locations where noise types are unknown and ground-truth frames are unavailable. It introduces DeepTemporal Interpolation, which embeds a temporal filter into a UNet-based denoiser that operates on a 7-frame stack $F_t=[f_{t-k},...,f_t,...,f_{t+k}]$ with $M=2k+1$ and $k=3$, to produce the denoised center frame $\hat{f_t}$ without explicit motion estimation. Extensive experiments on synthetic DAVIS/SET8 data with Gaussian, Poisson, and Impulse noise, as well as real calcium imaging and fluorescence microscopy, show the method outperforms state-of-the-art unsupervised denoisers (and is competitive with supervised baselines) by leveraging fixed-noise self-supervision and temporal information. The approach offers robust, adaptable microscopy video restoration suitable for real-world biomedical analysis where noise characteristics are diverse and ground-truth references are unavailable.

Abstract

In this paper, we introduce a novel unsupervised network to denoise microscopy videos featured by image sequences captured by a fixed location microscopy camera. Specifically, we propose a DeepTemporal Interpolation method, leveraging a temporal signal filter integrated into the bottom CNN layers, to restore microscopy videos corrupted by unknown noise types. Our unsupervised denoising architecture is distinguished by its ability to adapt to multiple noise conditions without the need for pre-existing noise distribution knowledge, addressing a significant challenge in real-world medical applications. Furthermore, we evaluate our denoising framework using both real microscopy recordings and simulated data, validating our outperforming video denoising performance across a broad spectrum of noise scenarios. Extensive experiments demonstrate that our unsupervised model consistently outperforms state-of-the-art supervised and unsupervised video denoising techniques, proving especially effective for microscopy videos.

Figures (3)

• Figure 1: Visual result for different noise types. (a) shows the denoising outcome for Gaussian noise with a standard deviation ($\sigma$) of 30, (b) showcases the denoising results for Poisson noise with a parameter ($\lambda$) of 30, and (c) presents the output achieved by denoising an image corrupted with Impulse noise with a probability of 0.2. These results demonstrate the effectiveness of our method in reducing different types of noise while preserving the intricate details of the frames.
• Figure 2: Model Architecture
• Figure 3: The illustration of the linear filter output