Image2TetMesh: A hybrid image/graph convolutional neural network for generating volume meshes of the total cavopulmonary connection from 3D MRI

Motivation: The Fontan procedure palliates single-ventricle patients with the total cavopulmonary connection (TCPC). Hemodynamic analysis and 3D modeling of the TCPC are possible using volume-meshes, however creating these meshes is labor-intensive and difficult due to the complex anatomy. Goal(s): Automate volume-mesh generation of the TCPC. Approach: We present Image2TetMesh, a hybrid image/graph deep learning model that directly generates patient-specific TCPC volume-meshes from 3D MRI. Results: Image2TetMesh produces anatomical volume-meshes with accurate vessel diameters, achieving a median Dice of 0.87, outperforming the conventional 3D UNet across all segmentation metrics. It works 30x faster than the 3D UNet (with post-processing) and 3000x faster than manual methods. Impact: Image2TetMesh is a deep learning model that rapidly generates accurate volume-meshes of the total cavopulmonary connection from 3D MRI in single ventricle patients. Our model has the potential to support hemodynamic analysis and personalized treatment planning for this patient cohort.