Deep learning based virtual multimodal fluorescence staining of fresh thick tissue using 3D epi-mode quantitative phase imaging

Despite the disadvantages of labeling cells, fluorescence imaging remains a cornerstone of biological and biomedical imaging. However, quantitative phase imaging (QPI) is increasingly being recognized in the biomedical field as an equally indispensable tool given its label-free, non-destructive, highly sensitive and long-term quantitative imaging capabilities. Recently, deep neural network-based methods have been developed to generate multi-spectral virtual fluorescent images from QPI, providing the best of both worlds. Yet, such methods have been limited to thin transparent samples, given the inherent inability of traditional QPI methods to image thick scattering samples. In this study, we develop a modified quantitative back-illumination microscopy (qOBM) system to obtain 3D quantitative phase information from fresh, thick samples and fluorescence data in tandem. The system can provide perfectly co-registered multi-modal images which can also be used to train a deep neural network to translate phase information into fluorescent information. Here we present the theoretical foundation of this method, describe the simple multi-modal system, and showcase results from both imaging modalities (quantitative phase and fluorescence in 3D) as well as the virtually stained images. This innovative virtual staining technique has significant implications for the investigation of cells in complex scattering media (e.g., tissue), as it holds the potential to significantly reduce time, labor, and costs, while providing new capabilities for high throughput, label-free, high-contrast imaging without compromising cell health. Additionally, for clinical applications, this method holds great promise to enable high-fidelity label-free virtual histological imaging in-situ/in-vivo with known ground truth for nuclear contrast.