Exploiting Scale Invariance and Rotation Equivariance for Sparse and Dense Artery Orientation Estimation

Dieuwertje Alblas, Iris Vos, Julian Suk, Christoph Brune, Kak Khee Yeung, Birgitta Velthuis, Hugo Kuijf, Ynte Ruigrok, Jelmer M. Wolterink

Keywords: artery orientation, rotation equivariance, scale invariance, graph neural network

Abstract: We present SIRE, a modular estimator of local artery orientations that is Scale Invariant and Rotation Equivariant. These symmetry preservations are obtained by operating on *spherical* image patches at multiple scales in parallel, and allow generalisation to arteries of unseen sizes and orientations. We embed SIRE into two different artery centerline tracking algorithms: a sparse, iterative tracker starting at a single seed point and a dense image filter serving as a cost function for connecting two bifurcation points. We show that SIRE can be used to obtain centerlines of arteries of various sizes and tortuosities by including datasets containing abdominal aortic aneurysms, coronary arteries and intracranial arteries.

Submission Number: 80