Fig. 1: Overview of fingerprint SRS microscopy by ultrafast tuning and spatial-spectral residual learning.

a The intrinsic cross-section of the coherent Raman scattering process and instrumentation define the conventional design space for SRS imaging, resulting in trade-offs between bandwidth (i.e., spectral resolution), speed, and signal-to-noise ratio (SNR). Deep learning can expand the design space through computational methods, enabling high-speed, high-SNR fingerprint SRS imaging of living cells and large-area tissues. b Setup of ultrafast delay-line tuning. A 55-kHz polygon scanner is used to scan the Stokes beam onto a Littrow-configured blazed grating to generate an SRS spectrum within 20 µs. By changing the angle between the grating blazed line and the laser-scanned line (θ), the effective delay range can be fine-tuned as Lsin(θ)tan(α), where α represents the grating blazed angle and L is the length of laser scan line. PBS polarizing beam splitter, QWP quarter-wave plate, HWP half-wave plate, PS polygon scanner, DM dichroic mirror. c Training a spatial-spectral residual net (SS-ResNet) deep neural network for SNR improvement, ground truth (GT) images are generated by averaging multiple acquisitions of the same field-of-view, equivalent to increasing the pixel dwell time. A trained network is then applied to recover the SNR of high-speed yet noisy images. d Spectral unmixing using least absolute shrinkage and selection operator (LASSO) to generate chemical maps. Int., intensity. a.u., arbitrary unit.