Baikal: Unpaired Denoising of Fluorescence Microscopy Images Using Diffusion Models

Fluorescence microscopy is an indispensable tool for biological discovery but image quality is constrained by desired spatial and temporal resolution, sample sensitivity, and other factors. Computational denoising methods can bypass imaging constraints and improve signal-to-noise ratio in images. However, current state of the art methods are commonly trained in a supervised manner, requiring paired noisy and clean images, limiting their application across diverse datasets. An alternative class of denoising models can be trained in a self-supervised manner, assuming independent noise across samples but are unable to generalize from available unpaired clean images. A method that can be trained without paired data and can use information from available unpaired high-quality images would address both weaknesses. Here, we present Baikal, a first attempt to formulate such a framework using Denoising Diffusion Probabilistic Models (DDPM) for fluorescence microscopy images. We first train a DDPM backbone in an unconditional manner to learn generative priors over complex morphologies in microscopy images. We then apply various conditioning strategies to sample from the trained model and propose an optimal strategy to denoise the desired image. Extensive quantitative comparisons demonstrate better performance of Baikal over state of the art self-supervised methods across multiple datasets. We highlight the advantage of generative priors learnt by DDPMs in denoising complex Flywing morphologies where other methods fail. Overall, our DDPM based denoising framework presents a new class of denoising methods for fluorescence microscopy datasets that achieve good performance without collection of paired high-quality images.