[home] [Personal Program] [Help]

---

Session: Vascular - II
Session starts: Friday 27 January, 11:30
Presentation starts: 12:00
Room: Room 559

---

Julian Suk (University of Twente)

Abstract:
Computational fluid dynamics (CFD) is used for the non-invasive estimation of hemodynamic biomarkers like wall shear stress (WSS) from 3D models of human arteries. However, long computation times limit the widespread applicability of CFD in clinical practice. As an alternative, we propose to use a graph convolutional network (GCN) operating on a surface mesh. Our GCN can produce accurate directional WSS estimates mapped to the mesh vertices in under five seconds [1]. For training, we randomly generate datasets of single and bifurcating coronary artery models and simulate steady and pulsatile blood flow, subject to varying boundary conditions. Surface meshes with WSS vectors mapped to their vertices are used to train a GCN consisting of gauge-equivariant mesh (GEM) convolution [2] and pooling layers. The network is provably SE(3)-equivariant: shifting the input mesh in space does not affect WSS vectors, while a rotation results in accordingly rotated WSS vectors. GEM-GCN can predict WSS fields jointly for a set of discrete time points in the cardiac cycle. We condition the network on simulation-specific coronary blood flow by appending it as scalar field to the input features. In previously unseen arteries, our neural network produces WSS estimations with an approximation error of 7.8 % and 11.9 % for the single and bifurcating arteries, respectively. Predicting directional WSS on a surface mesh takes less than 5 seconds, compared to more than ten minutes for the corresponding CFD simulation. Our results indicate that GEM-GCN has the potential to be a feasible surrogate for CFD in time-critical applications. Further research is needed to assess generalisation to patient- specific 3D artery models. References [1] Suk, J., de Haan, P., Lippe, P., Brune, C., and Wolterink, J.M. (2021). Mesh convolutional neural networks for wall shear stress estimation in 3d artery models. In Statistical Atlases and Computational Models of the Heart. [2] de Haan, P., Weiler, M., Cohen, T., and Welling, M. (2021). Gauge equivariant mesh CNNs: anisotropic convolutions on geometric graphs. In Proceedings of the 9th International Conference on Learning Representations.