Advances in Stimulated Raman Scattering Microscopy via Deep Learning

Abstract

Stimulated Raman scattering (SRS) microscopy is a powerful chemical imaging technique that acquires images based on the vibrational-spectral “fingerprints” of molecules within an imaged field of view often without the need for exogenous fluorophores or labels. SRS microscopy has found an established niche in biophotonics with many examples of translational clinical applications and demonstrations of imaging various biological systems on subcellular to tissue spatial orders. Concurrent to the development of SRS microscopy, computational advancements have seen a democratized adoption of deep learning platforms for a wide variety of computer vision tasks. In this work I document my contributions in integrating SRS microscopy and deep learning towards advancing the capability to study biological systems. Specifically, deep learning will be shown to address technical limitations of SRS microscopy such as imaging noise and ultimate imaging depth in tissue samples. Deep learning will also be shown to improve analysis of SRS images via the development of a novel convolutional neural network architecture designed to handle a variety of chemical imaging techniques and perform a variety of computer vision tasks. Finally, I will show how these advancements and novel architecture can be used to diagnose thyroid cancer in label-free human tissue samples and to classify and study T cells based on label-free images.