Abstract:Advances in microscopy imaging enable researchers to visualize structures at the nanoscale level thereby unraveling intricate details of biological organization. However, challenges such as image noise, photobleaching of fluorophores, and low tolerability of biological samples to high light doses remain, restricting temporal resolutions and experiment durations. Reduced laser doses enable longer measurements at the cost of lower resolution and increased noise, which hinders accurate downstream analyses. Here we train a denoising diffusion probabilistic model (DDPM) to predict high-resolution images by conditioning the model on low-resolution information. Additionally, the probabilistic aspect of the DDPM allows for repeated generation of images that tend to further increase the signal-to-noise ratio. We show that our model achieves a performance that is better or similar to the previously best-performing methods, across four highly diverse datasets. Importantly, while any of the previous methods show competitive performance for some, but not all datasets, our method consistently achieves high performance across all four data sets, suggesting high generalizability.

What problem does this paper attempt to address?

### Problems Addressed by the Paper This paper aims to address the issue of image denoising in high-resolution microscopy imaging. Specifically, the authors propose a method based on the Denoising Diffusion Probabilistic Model (DDPM) to restore high-resolution images from low-resolution and high-noise microscopy images. The paper mainly focuses on the following aspects: 1. **Image Noise Issue**: A common problem in high-resolution microscopy imaging is image noise, which can severely affect the accuracy of subsequent analysis. Reducing the laser dose can extend the measurement time but will lead to decreased image resolution and increased noise. 2. **Photobleaching and Phototoxicity of Biological Samples**: High light intensity can cause photobleaching of fluorescent molecules and phototoxicity to biological samples, damaging the integrity of the samples and limiting the observation time. Reducing light intensity can alleviate these issues but will introduce more noise. 3. **Limitations of Existing Denoising Methods**: Traditional denoising methods (such as Gaussian filtering and wavelet transform) often fail to preserve detailed information while removing noise. In recent years, deep learning-based methods have made significant progress in image denoising tasks, but there are still certain limitations when dealing with different types of microscopy data. ### Solution The authors propose a Conditional Denoising Diffusion Probabilistic Model (Conditional DDPM) to address the above issues in the following ways: 1. **Model Training**: By training on low-resolution and high-noise images, the model can predict high-resolution images. The Conditional DDPM uses low-resolution information as a condition to gradually remove noise and generate high-quality images. 2. **Multiple Sampling**: Utilizing the randomness of DDPM, multiple images are generated and averaged to further improve the signal-to-noise ratio. This method performs well on multiple datasets, enhancing image quality. 3. **Dataset Diversity**: The authors tested the model on four highly diverse datasets, including STED microscopy images, confocal, and Airyscan super-resolution images. These datasets vary greatly in imaging processes, sample types, noise levels, and structures, validating the model's generalization capability. ### Experimental Results Experimental results show that the proposed DDPM model outperforms or is at least comparable to the current best benchmark models on multiple datasets. Particularly, in handling STED images of live cell mitochondria and fixed cell microtubules, the DDPM model excels, restoring high-quality high-resolution images while maintaining sample integrity. In conclusion, this paper effectively addresses the image denoising problem in high-resolution microscopy imaging by introducing the DDPM model, providing new ideas and methods for future microscopy imaging technology.

Denoising diffusion models for high-resolution microscopy image restoration

Microscopy image reconstruction with physics-informed denoising diffusion probabilistic model

Resfusion: Denoising Diffusion Probabilistic Models for Image Restoration Based on Prior Residual Noise

Baikal: Unpaired Denoising of Fluorescence Microscopy Images using Diffusion Models

Diffusion Models to Enhance the Resolution of Microscopy Images: A Tutorial

Denoising Diffusion Probabilistic Models for Generation of Realistic Fully-Annotated Microscopy Image Data Sets

Complex-valued Retrievals From Noisy Images Using Diffusion Models

Performance of deep learning restoration methods for the extraction of particle dynamics in noisy microscopy image sequences

Unsupervised Denoising of Retinal OCT with Diffusion Probabilistic Model

Stimulating Diffusion Model for Image Denoising via Adaptive Embedding and Ensembling

PET image denoising based on denoising diffusion probabilistic models

High-resolution MRI synthesis using a data-driven framework with denoising diffusion probabilistic modeling

Denoising for Optical-Resolution Photoacoustic Microscopy Via a Score-Based Diffusion Generative Model

Residual Denoising Diffusion Models

Denoising Diffusion Probabilistic Models for Image Inpainting of Cell Distributions in the Human Brain

Image Denoising Via Multiscale Nonlinear Diffusion Models.

Reconstruct-and-Generate Diffusion Model for Detail-Preserving Image Denoising

Stimulating the Diffusion Model for Image Denoising Via Adaptive Embedding and Ensembling

Image Denoising via Multi-scale Nonlinear Diffusion Models

Denoising diffusion probabilistic models for generation of realistic fully-annotated microscopy image datasets