Abstract
Efficient algorithms are needed to segment vasculature in new 3D medical imaging datasets at scale for research and clinical applications. Manual segmentation of vessels in images is time-consuming and expensive whereas computational approaches have limited accuracy. We organize a global machine learning competition, engaging 1,401 participants, to promote development of deep learning methods for 3D blood vessel segmentation in Hierarchical Phase-Contrast Tomography (HiP-CT) datasets. This paper presents a meta-analysis of the top-performing solutions, focusing on segmentation accuracy and morphological analysis. The competition and subsequent analysis reveal convergent methodological innovations: pseudo-labeling approaches that exploit data distributions, metrics and loss functions that optimize for vessel surface and topology, and multi-scale approaches that handle data heterogeneity. Additionally, the paper presents techniques for building deep learning models for the defined task, metrics to assess and compare algorithm performance, and a dataset with manually annotated and curated gold standard segmentations for future studies in blood vessel segmentation within HiP-CT imaging.