Stimulated Raman scattering (SRS) imaging of fresh tissue is an emerging approach to render label-free pathology and diagnosis based on the chemical contrasts of native biomolecules. However, the contrast of cell nuclei is often too weak to perform reliable segmentation and quantification. To advance the development of label-free digital pathology with SRS, we explored three deep learning-based techniques with multicolor and hyperspectral SRS imaging data, including U-Net, Mask R-CNN, and Autoencoder. Our results show that a combination of machine learning and SRS imaging is a promising pathway to transforming the methodology of SRS label-free pathology for real clinical use.
|