Physics-assisted light-field microscopy enables high temporal-spatial resolution 3D observation

Abstract

Emerging light-field microscopy (LFM) overcomes limited 3D imaging speed by simultaneously encoding spatial and angular information through light-field capture for 3D reconstruction from a single 2D snapshot. However, conventional LFM is constrained by the inherent spatial-angular bandwidth product trade-off, resulting in insufficient spatial resolution. Here we developed a series of deep learning-based 3D reconstruction algorithms for LFM. These algorithms substantially enhance the spatial resolution of light-field microscopic imaging. Consequently, our approach enables high-speed (100 Hz volumetric imaging), super-resolution (120 nm), and long-term (60 hours) 3D observation of living cells. Furthermore, to circumvent reliance on ground truth data for processes where it is unobtainable, we propose an implicit neural representation network with a geometry-fluctuation dual-model for self-supervised light-field reconstruction. This network achieves high-resolution, low-artifact 3D reconstruction. Further integration with a digital adaptive optics module enables high-resolution 3D dynamic imaging of deep tissues in live mice.

Speaker

Dongyu Li
Huazhong University of Science and Technology
China

Discussion

Ask question