High-Speed Chemical Imaging by Dense-Net Learning of Femtosecond Stimulated Raman Scattering

Abstract

Hyperspectral stimulated Raman scattering (SRS) by spectral focusing can generate label-free chemical images through temporal scanning of chirped femtosecond pulses. Yet, pulse chirping decreases the pulse peak power and temporal scanning increases the acquisition time, resulting in a much slower imaging speed compared to single-frame SRS using femtosecond pulses. In this paper, we present a deep learning algorithm to solve the inverse problem of getting a chemically labeled image from a single-frame femtosecond SRS image. Our DenseNet-based learning method, termed as DeepChem, achieves high-speed chemical imaging with a large signal level. Speed is improved by 2 orders of magnitude with four subcellular components (lipid droplet, endoplasmic reticulum, nuclei, cytoplasm) classified in MIA PaCa-2 cells and other cell types which were not used for training. Lipid droplet dynamics and cellular response to dithiothreitol in live MIA PaCa-2 cells are demonstrated using this computationally multiplex method.