A Deep Learning Approach for Direct Mesh Reconstruction of Intracranial Arteries

Abstract

The Circle of Willis (CoW) is a group of vessels connecting major circulations of the brain. Its vascular geometry is believed to influence the onset and outcome of serious neurovascular pathologies. These geometric features can be obtained from surface meshes to capture vessel topology and morphology. A recent deep learning technique to handle non-Euclidean data, such as meshes, is Geometric Deep Learning (GDL). To this end, this study aimed to explore a GDL-based approach to directly reconstruct surface meshes of the CoW from magnetic resonance angiography images, thereby eliminating the traditional postprocessing steps required to obtain such a mesh from volumetric representations. The network architecture includes both convolutional and graph convolutional layers, allowing it to operate with images and meshes at the same time. It takes as input an image volume and a template mesh and outputs a 3D surface mesh. Experiments were performed on five crops representing different vessels and bifurcations to capture both stability and variability within the CoW. The results showed that anatomy-specific template input meshes and enhancement of the image feature representation increase the accuracy of the reconstruction. Moreover, incorporating the curvature characteristics of the meshes showed promising capability of handling complex geometries and sharp edges. However, achieving a consistent performance across CoW regions remains a challenge.