Unsupervised Hyperspectral Stimulated Raman Microscopy Image Enhancement: De-Noising and Segmentation via One-Shot Deep Learning

Hyperspectral stimulated Raman scattering (SRS) microscopy is a powerful label-free, chemical-specific technique for biomedical and mineralogical imaging which can suffer from low signal-to-noise ratios due to requirements of low input laser power or fast imaging, or from optical scattering and low target concentration. Here, we demonstrate a deep learning neural net model and unsupervised machine-learning algorithm for rapid and automatic de-noising and segmentation of SRS images based on a ten layer convolutional autoencoder: UHRED (Unsupervised Hyperspectral Resolution Enhancement and De-noising). UHRED is trained in an unsupervised manner using only a single (one-shot) hyperspectral image, with no requirements for training on high quality (ground truth) labelled data sets or images. Importantly, although we illustrate this method using SRS, the hyperspectral index (signal as a function of a laser parameter) may be any imaging modality such as harmonic generation, linear and/or nonlinear fluorescence, CARS, Pump-Probe, Thermal Lensing and Cross-Phase microscopy. UHRED significantly enhances SRS image contrast by de-noising the extracted Raman spectra at every image pixel. Applying a k-means clustering algorithm provides automatic, unsupervised image segmentation based on Raman vibrational spectra of the sample constituents, yielding intuitive chemical species maps, as we demonstrate for the case of a complex lithium ore sample.