Deep Learning Empowered Parallelized Metasurface Computed Tomography Snapshot Spectral Imaging

Abstract

Abstract Snapshot spectral imaging is an emerging technology for fast data acquisition in dynamic environments, capturing high‐volume spatial‐spectral information in a single snapshot. However, it suffers from bulky cascading optics and cannot be directly used in space‐restricted scenarios such as endoscope‐assisted brain microsurgery and real‐time cellular tissue imaging. In this work, an ultracompact strategy of parallelized metasurface computed tomography empowered by generative deep learning is proposed, which can effectively reduce the optics volume in snapshot spectral imaging from cm 3 scale to sub‐mm 3 scale while retaining high resolution and speed of imaging so that the above‐mentioned pain point problem is well addressed. The system comprises seven multifunctional sub‐metasurfaces simultaneously acquiring multi‐angle spectral projection and integration information of the target, uses the system‐calibrated point spread functions as wavelength and spatial position distributions, and incorporates a generative adversarial deep neural network for fast reconstruction of spatial‐spectral multiplexed images. Experimental results show that single snapshot imaging can be achieved in 38 ms with a spectral resolution of 10 nm in the spectral range of 450–650 nm. This technique paves the way for snapshot spectral imaging integration into various highly miniaturized microscopy and endoscopic imaging systems in applications such as advanced medical diagnosis.