News Release

Deep learning-powered denoising technique for high-speed dynamic fluorescence imaging

Peer-Reviewed Publication

Chinese Society for Optical Engineering

Denoising results for in vivo venule images in the cremaster muscles of mice.

image: 

In PhotoniX, researchers report a self-supervised deep learning method that denoises dynamic fluorescence images in vivo without requiring clean training data. The figure shows in vivo venule images in the cremaster muscles of mice (red: myeloid cells; green: leukocytes and platelets), comparing raw noisy images (left) and denoised results using the TeD model (right). Insets show magnified time-lapse frames highlighting moving blood cells. Scale bars, 50 μm.

view more 

Credit: W. Lee et al. 2025 Springer Nature Publishing Group (DOI 10.1186/s43074-025-00173-8)

A new deep learning-based approach has been developed to overcome one of the critical limitations in fluorescence microscopy: severe image degradation caused by noise in dynamic in vivo imaging environments. The technique, recently published in PhotoniX (May 23, 2025), introduces a self-supervised denoising network—TeD (Temporal-gradient empowered Denoising)—that improves image quality without requiring clean reference images, representing a breakthrough for applications involving rapid biological dynamics.

Fluorescence microscopy plays a central role in studying live biological processes at cellular and subcellular levels. However, low signal-to-noise ratios under photon-limited conditions often obscure critical biological signals, particularly during high-speed image acquisition. TeD addresses this challenge by incorporating a temporal gradient-based attention mechanism that detects spatial motion and adaptively adjusts the use of temporal redundancy for denoising.

By processing time-lapse image sequences with this strategy, the model selectively utilizes relevant spatiotemporal features, enabling robust denoising even in data with fast-moving structures such as circulating blood cells, beating vascular walls, and calcium transients in neurons. Unlike traditional supervised models, it does not require clean ground truth images for training, making it suitable for real-world in vivo imaging scenarios.

Validation across multiple imaging modalities, including confocal and two-photon fluorescence microscopy, demonstrated TeD’s ability to recover fine structural details under both static and dynamic conditions. Quantitative assessments also confirmed significant improvements in signal-to-noise ratio and structural fidelity when compared to existing methods.

This development is expected to aid the interpretation of a wide range of fluorescence images involving spatiotemporal dynamics, potentially enabling new pathological discoveries and advancing the understanding of complex biological processes.


Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.