Deep learning spectroscopic stimulated Raman scattering microscopy

Spectroscopic stimulated Raman scattering (SRS) is a label-free chemical imaging modality enabling visualization of molecules in living systems with high specificity. Among various spectroscopic SRS imaging methods, a convenient way is through linearly chirping two femtosecond lasers and tuning their temporal delay, which in turn corresponds to different Raman shifts. Currently, the acquisition speed using a resonant mirror is 3 seconds (80 microseconds per spectrum), which is insufficient for imaging samples with high motility. In this work, we aim to push the imaging speed using a 50-kHz polygon scanner as a delay line tuner, achieving a speed of 20 microseconds per spectrum. At such high speeds, to overcome the signal level decrease due to reduced signal integration time, we apply a U-Net deep learning framework, which first takes pairs of spectroscopic SRS images at different speeds as training samples, with high-speed, low-signal images as input and low speed, high-signal ones as output. After training, the network is capable of rapidly transforming a low-signal spectroscopic image to a high-signal version. Consequently, our design can generate ultrafast spectroscopic SRS image while maintaining the signal level comparable to the output with longer signal integration time.