Generative AI Methods in Fluorescent Virtual Staining

Abstract

This thesis investigates the application of advanced deep learning architectures, specifically the U-Net and the conditional Wasserstein GAN with gradient penalty (cWGAN-GP), in predicting fluorescent staining from transmitted light microscopy images, using differential interference contrast (DIC) imaging. Fluorescence staining is a technique in microscopy used in biomedical and material sciences to gain detailed visualizations of cellular structures and functions through the use of a sizeable variety of available fluorophores, while also allowing simultaneous imaging of different cellular and subcellular components. However, while traditional fluorescence microscopy is informative, it is also hindered by high costs, phototoxicity, and labor-intensive preparations. This research addresses these limitations by utilizing AI-driven models on a diverse dataset from the Light My Cells Grand Challenge, which includes both transmitted light images and fluorescent targets. After a thorough review of the deep learning theoretical fundamentals, this study examines both single-layer and multi-layer (z-stack) input configurations to evaluate their performance in generating highquality predictions. Key challenges such as handling target image sparsity and the complexity introduced by multi-channel data are discussed, alongside the trade-offs in preprocessing techniques like resizing and tiling. The findings demonstrate the potential of AI in histological practices, providing cost-effective, and non-invasive alternatives. Future research directions include mixed modality datasets and improved normalization techniques. Overall, this thesis aims to contribute to the research of AI applications in the field of biomedical imaging with respect to in silico staining solutions.