Ai-driven flexible and generalizable 3d vascular model building
Patryk Rygiel (EEMCS-MIA), Dieuwertje Alblas (EEMCS-MIA), Christoph Brune (EEMCS-MIA), Kak
Khee Yeung (Departament of Surgery, Amsterdam UMC) and Jelmer M. Wolterink (EEMCS-MIA)
Abstract
Personalized 3D vascular models can aid in a range of diagnostic, prognostic, and treatment-planning
tasks relevant to cardiovascular disease management. For optimal utility, models should be highly
accurate and topologically correct. Moreover, a user should have full control over the blood vessel
segments to be included in a vascular model. We propose an AI-driven model that achieves both goals.
We combine a global controller that leverages voxel mask segmentations to provide boundary
conditions for vessels of interest with a local, iterative vessel segmentation model. We introduce the
preservation of scale- and rotational symmetries in the local segmentation model, leading to
generalization to vessels of unseen sizes and orientations. Combined with the global controller, this
enables flexible 3D vascular model building, without the need for vessel-specific training data
annotation.
We demonstrate the potential of our method on a dataset containing abdominal aortic aneurysms
(AAAs). Our method provides a providing a watertight, smooth surface segmentation, with overlap
and distance metrics on par with a state-of-the-art segmentation model in the segmentation of AAAs,
iliac arteries and renal arteries. Moreover, we demonstrate that by adapting the global controller, we
can easily extend vessel sections in the 3D model. Additionally, to incentivize study of AAA shapes,
we release a public dataset “AAA-100”, semi-automatically segmented with a proposed method, that
includes 100 AAA lumen shapes containing abdominal aorta, both iliac arteries and both renal
arteries.