image: Cross-domain Cross-domain knowledge transfer from large-scale natural-scene video datasets into holographic microscopy enables reconstruction of fast biological dynamics at millisecond time scale. This is made possible because the underlying low-level spatiotemporal features are shared by videos from different imaging modalities.knowledge transfer from large-scale natural-scene video datasets into holographic microscopy enables reconstruction of fast biological dynamics at millisecond time scale. This is made possible because the underlying low-level spatiotemporal features are shared by videos from different imaging modalities.
Credit: Liangcai Cao, Tsinghua University
Computational optical imaging has long faced a trade-off between information throughput and temporal resolution. Traditional methods often require multiple sequential measurements to reconstruct high-dimensional data, which inherently slows down the imaging speed and limits the ability to observe fast-moving biological phenomena.
In a study published recently, the team led by Professor Liangcai Cao introduced STRIVER (SpatioTemporally Regularized InVERsion), a framework that integrates physical imaging models with deep learning priors.
Bridging the Gap with Spatiotemporal Priors
The core innovation lies in the use of a Plug-and-Play (PnP) optimization theory. Instead of relying on traditional analytical regularizers, the researchers utilized a deep video denoising network (ViDNet) trained on massive datasets of natural videos.
"While natural scenes and microscopic images differ in semantics, they share highly similar low-level spatiotemporal features, such as texture details and topological structures," explains Yunhui Gao, the first author of the study. "By migrating these features into the microscopy domain, we can reconstruct high-quality videos from sparse, sequential data without sacrificing temporal resolution".
Real-World Applications in Life Sciences
The researchers applied this technology to a lensless computational microscopy system based on coded ptychography. The setup allows for a large field of view and integrated, label-free imaging.
The team successfully captured the dynamic activities of live microorganisms, such as Paramecium and Rotifers. The high-speed imaging allowed for the clear visualization of subcellular structures in rapidly moving cells (up to 2.7 mm/s), including the formation and release of food vacuoles and complex tumbling motions.
Future Outlook
"The spatiotemporal computational imaging philosophy and the core algorithms developed in this work are expected to be applied to various multi-frame computational imaging modalities," says Professor Liangcai Cao. "This will pave the way for future advancements in motion analysis and image quality enhancement across different fields of microscopy".
Journal
PhotoniX
Method of Research
Experimental study
Subject of Research
Not applicable
Article Title
Model-based deep learning enables time-resolved computational microscopy
Article Publication Date
7-Jan-2026
COI Statement
The authors declare no conflicts of interest.