Stimulated Raman scattering (SRS) images often suffer from low signal to noise ratio (SNR) due to absorption and scattering of light as well as limited optical power. We use deep learning to significantly improve the SNR of SRS images. Our algorithm, based on a U-Net convolutional neural network, significantly outperforms existing denoising algorithms. The trained denoising algorithm is applicable to images acquired at different imaging powers, depths, and experimental geometries not explicitly included in the training. Our results identify potential towards in vivo applications, where ground-truth images are not always available to create a paired training set for supervised learning.
|