Self-Supervised Synthetic Cerebral Vessel Tree Generation Using Semantic Signed Distance Fields

Abstract

Advances in in-silico clinical trails for the development of novel treatment and devices for acute ischemic stroke have driven the creation of synthetic virtual patient populations to address the lack of large real-world datasets. Recent work proposed a method for generating semantic vascular centerline tree of the major cerebral arteries using pointcloud diffusion. However, this approach relies on separate post-processing algorithms to reconstruct the vessel tree topology, which does not generalize well to more topologically complex trees. To overcome this limitation, we introduce semantic signed distance fields for modeling cerebral vessel trees in a fully self-supervised manner. Our approach bypasses the need for separate reconstruction of the tree topology, and can be trained directly on shape-surfaces. Our method combines a variational autoencoder for encoding shapes to robust latent shape representations with a latent-diffusion model for generating synthetic vessel trees. By generating surface geometry directly, our approach eliminates the need for post-processing steps, enabling the generation of high-quality and topologically complex cerebral vessel trees.