Preview

PDF, English Download (30MB) | Terms of use

Abstract

This thesis presents the development and evaluation of a multi-wavelength, multimodal Raman light-sheet microscopy (RLSM) platform for fast, label-free 3D imaging of tumour spheroids and 3D cell cultures. The system combines selective plane illumination, dualwavelength excitation combination (near-infrared and visible), hyperspectral detection and deep-learning-based restoration to address key limitations of conventional Raman imaging, including weak scattering efficiency, slow volumetric acquisition, fluorescence background and low signal-to-noise ratio in thick, scattering samples. By simultaneously capturing Rayleigh scattering, Raman scattering and fluorescence, the microscope generates structurally and chemically rich 4D data cubes that support both qualitative visualization and quantitative analysis of treatment effects in realistic 3D cancer models. The main focus is also the use of advanced AI-algorithms such as zero-shot and self-supervised deep learning under realistic conditions where clean, high-SNR ground truth is unavailable. The thesis investigates ZS-DeconvNet, Noise2Noise, Noise2Void, Self2Self and, in particular, Deep Image Prior (DIP) as denoising and deblurring strategies that can be optimized directly on noisy RLSM volumes without external training data. Across multiple modalities, wavelengths and cell lines, most algorithms substantially suppress noise while preserving structural detail. Quantitative evaluation demonstrates that DIP consistently achieves the best overall performance. The enhanced volumes are shown to be biologically meaningful rather than merely visually improved: in cisplatin-treated spheroids, denoised Rayleigh and Raman channels reveal treatment-dependent changes in spheroid morphology, internal heterogeneity and Raman peak patterns that are consistent with altered viability and biochemical composition. The thesis argues that RLSM should be regarded as an integrated pipeline (optical setup and advanced AI algorithms) in which excitation wavelengths, lightsheet geometry, spectral selection and denoising algorithms are co-designed. Remaining challenges include the lack of true ground-truth volumes, computational cost, crossinstrument generalization and data management. Overall, the results demonstrate that multimodal RLSM combined with advanced AI-based restoration can provide reliable, quantitative 3D biochemical imaging for drug screening, treatment monitoring and, potentially, label-free tissue assessment.