Deep Learning-Based Pulmonary Artery Surface Mesh Generation

Abstract

Properties of the pulmonary artery play an essential role in the diagnosis and treatment planning of diseases such as pulmonary hypertension. Patient-specific simulation of hemodynamics can support the planning of interventions. However, the variable complex branching structure of the pulmonary artery poses a challenge for image-based generation of suitable geometries. State-of-the-art segmentation-based approaches require an interactive 3D surface reconstruction to prepare the simulation geometry. We propose a deep learning approach to generate a 3D surface mesh of the pulmonary artery from CT images suitable for simulation. The proposed method is based on the Voxel2Mesh algorithm and includes a voxel encoder and decoder as well as a mesh decoder to deform a prototype mesh. An additional centerline coverage loss facilitates the reconstruction of the branching structure. Furthermore, vertex classification allows for the definition of in- and outlets. Our model was trained with 48 human cases and tested on 10 human cases annotated by two observers. The differences in the anatomical parameters inferred from the automatic surface generation correspond to the differences between the observers’ annotations. The suitability of the generated mesh geometries for numerical flow simulations is demonstrated.