Advancing Stimulated Raman Scattering Microscopy through Deep Learning and Gel-Based Tissue Engineering

Stimulated Raman scattering (SRS) microscopy is a highly effective label-free imaging method for investigating the molecular composition of biological systems. Its broader use has been held back by spatial resolution, imaging speed, and large-scale tissue imaging compatibility. Breaking these limitations requires an integrated approach beyond the development of optical hardware. This thesis introduces a compilation of techniques that leverage gel-based tissue engineering and deep learning to enhance the capabilities of SRS microscopy. The first chapter, Gel-Enabled Super-Resolution Label-Free Volumetric Vibrational Imaging, introduces VISTA, a sample-expansion vibrational imaging technique that achieves label-free super-resolution imaging of protein-dense biological structures with resolution as fine as 78 nm. By enabling isotropic expansion and protein retention, VISTA allows for high-throughput, unbiased volumetric imaging without labeling, with further enhancement using deep learning-based component prediction. The second chapter, High-Resolution Imaging of In Vivo Protein Aggregates, applies VISTA to image amyloid-beta and polyQ aggregates in biological samples with high specificity. Combined with segmentation using convolutional neural networks, this technique is capable of mapping aggregate structure and microenvironments, enabling new insights into neurodegenerative disease pathology. The third chapter, High-Throughput Volumetric Mapping Facilitated by Active Tissue SHRINK, introduces SHRINK, a hydrogel-based sample shrinkage method that isotropically shrinks tissue while maintaining structural integrity. Active shrinkage enhances imaging throughput and signal sensitivity and enables rapid, large-scale, three-dimensional whole-organ mapping with SRS microscopy. The fourth chapter, Deep Learning-Augmented Metabolic Profiling in Live Neuronal Cultures, presents a tandem deep learning platform for live-cell metabolic imaging. By integrating a recurrent convolutional neural network and U-Net segmentation model with deuterium-labeled metabolic tracing, this platform enables non-invasive, high-speed profiling of lipid, protein, glucose, and water metabolism in neuronal subtypes under physiological and pathological conditions. These developments represent multidimensional strategies that expand the application of SRS microscopy to high-resolution, high-throughput, and dynamic imaging in a variety of biological systems. The integration of deep learning and gel-based tissue engineering techniques opens new avenues for SRS microscopy to explore complex biological questions.