Deep Learning Techniques for Unmixing of Hyperspectral Stimulated Raman Scattering Images

Stimulated Raman Scattering (SRS) microscopy is a stain-free, laser-scanning imaging technology that utilizes two coherent laser beams (i.e., the pump and Stokes) to stimulate vibration of chemical bonds in molecules. Different bonds have different resonant vibrational frequencies, and thus SRS microscopy can achieve rapid chemical imaging at high-resolution, enabling live cell imaging and near-instant, stain-free pathological imaging. To increase the ability to resolve different chemical species, multiple Raman wavenumbers can be used with the hyperspectral SRS imaging data. In particular, this approach holds promise for quantifying DNA content, which is important to characterize cancer cell polyploidy. The SRS spectra is the mixture of the spectra of various pure substances present in each pixel so unmixing must be performed to find the relative abundances of these substances. We ran our SRS hyperspectral data of cancer cells through SciPy’s Least Square Error Linear Optimization algorithm (LSQ) [1] but found that it was not able to return the correct DNA content. Our proposed solution to this problem is to use an autoencoder neural network to unmix the spectra. We based the network on the findings in Palsson et al. (2018) [2]. Our initial results show that the network is effective at finding an accurate linear combination, but the noise in the collection of the SRS hyperspectral data significantly increases the number of low error solutions which makes it difficult for the network to find the true linear combination. Future work will be focused on using noise reduction techniques to help the network find the true abundance values.