Our understanding of tissue pathology would significantly benefit from comprehensive models of organ-scale microstructures. One of the most fundamental tissue microstructures are microvascular networks, which form meshworks of capillaries that transport nutrients to cells. However, microvascular models are extremely challenging to build on several fronts. First, these structures must be imaged across large three-dimensional tissue samples at resolutions near the optical diffraction limit. Second, the embedded models have to be reconstructed from gigavoxel-scale images. Finally, the resulting 3D models have to be integrated into simulations that integrate fluid and tissue dynamics with molecular interactions in the surrounding tissue. Our approaches focus on two challenges. First, we have developed a tissue scanning framework for low-cost three-dimensional wide-field imaging at resolutions sufficient to capture capillaries that form microvascular networks. Image acquisition relies on block-face fluorescence imaging using deep ultraviolet excitation. Tissue samples are labeled en bloc using traditional fluorescent dyes and embedded in a hard polymer or traditional paraffin wax. Three-dimensional images are acquired using MUSE microscopy, which leverages deep-ultraviolet light to limit acquisition to a thin layer at the sample surface. The imaged tissue is then ablated using a microtome or other automated milling system. By fully automating this procedure, the resulting three-dimensional image stack is aligned and processed to build an explicit model of the tissue. Next, segmentation algorithms are developed to take advantage of similarity and connectivity seen across large tissue volumes. Convolutional neural network (CNN) architectures such as U-Net are adapted to perform semantic segmentation as an initial pass \cite{saadatifard2020cellular}, providing a mean Dice coefficient of 0.89. We then apply highly parallel level sets to refine the microvascular structures embedded in the images improving performance to a mean Dice coefficient of 0.95. This includes taking advantage of GPU-based methods and sparse storage using B+ trees (OpenVDB). These methods are demonstrated on large volumes of mouse liver, which is highly vascularized. The resulting grids store 4% of the entire volume, which can be converted to an explicit geometric surface can be extracted for physically-based modeling.