UCLA Advances Imaging Microscopy with AI
Biologists have long struggled with the necessity of imaging minute biological processes – such as flowing blood cells or neurons moving through the brain – across long time-scales, as even supercomputers often strain to produce 3D imaging of those processes at longer than a few milliseconds. At UCLA, researchers are aiming to bridge that gap, enabling more advanced dynamic imaging microscopy of tissue samples – with the help of AI.
Currently, optical technologies are not advanced enough to capture these processes at the desirable spatiotemporal resolutions – so the researchers took a computational imaging technique called light-field microscopy for 3D imaging and supercharged it with a neural network. “Different from conventional microscopy, the tool reconstructed the 3D biological sample based on one snapshot through post-processing instead of scanning in the captured stage,” explained Zhaoqiang Wang, a doctoral student in bioengineering at UCLA’s Samueli School of Engineering and lead author on the paper. “The resulting temporal resolution of the images was drastically improved.”
To train the neural network, the researchers used 3D image stacks paired with images from light-field microscopy, teaching the network to be able to reconstruct the 3D images based on the light-field imaging. The resulting tool was tested on roundworms and zebrafish, where it was (respectively) used to track fluorescent tags and record the movements of blood and cardiac cells. The tool achieved 200 cubic frames per second and identified processes that occurred at spatial resolutions smaller than a grain of salt.
“This new system allows us to see biological events live in what is essentially five dimensions — the three dimensions of space, plus time and the molecular level dynamics as highlighted by color spectra,” said Tzung Hsiai, a professor of cardiology at UCLA and co-author of the paper. “For doctors and scientists, this could reveal the fine details of what’s happening in microscopic spaces and over millisecond-length time scales in a way that has never been done before. This advance can go a long way in helping find new insights to understand and treat diseases.”
The research was published as “Real-time volumetric reconstruction of biological dynamics with light-field microscopy and deep learning” in the February 2021 issue of Nature Methods. The article was written by Zhaoqiang Wang, Lanxin Zhu, Hao Zhang, Guo Li, Chengqiang Yi, Yi Li, Yicong Yang, Yichen Ding, Mei Zhen, Shangbang Gao, Tzung Hsiai, and Peng Fei. It can be accessed at this link.