Abstract
Tree-like structures, such as blood vessels, often express complexity at veryfine scales, requiring high-resolution grids to adequately describe theirshape. Such sparse morphology can alternately be represented by locations ofcentreline points, but learning from this type of data with deep learning ischallenging due to it being unordered, and permutation invariant. In this work,we propose a deep neural network that directly consumes unordered points alongthe centreline of a branching structure, to identify the topology of therepresented structure in a single-shot. Key to our approach is the use of anovel multi-task loss function, enabling instance segmentation of arbitrarilycomplex branching structures. We train the network solely using syntheticallygenerated data, utilizing domain randomization to facilitate the transfer toreal 2D and 3D data. Results show that our network can reliably extractmeaningful information about branch locations, bifurcations and endpoints, andsets a new benchmark for semantic instance segmentation in branchingstructures.