Segmentation of Microvascular Networks Embedded in Gigavoxel 3D Images Using RSF Level Sets with OpenVDB

Microvascular networks are vital for tissue function and disease progression, but their complex three-dimensional structure makes them difficult to analyze. Recent milling-based microscopy methods can capture images of these networks in whole organs at high resolution, though the resulting gigavoxel-scale images are challenging to segment. Convo-lutional neural networks (CNNs) are commonly used for this task, but they cannot account for the network's shape and topology. This paper presents a solution using a fully automated milling microscope to create a gigavoxel-scale dataset of mouse liver microvasculature. A CNN is trained to create an initial segmentation of the vascular network. The vessels are then refined using a parallel RSF-based level set model. To make this model practical on such large volumes, it is implemented in parallel using a sparse OpenVDB data structure that reduces the grid size to approxately 4% of the original.