Reconstructing 3D Cardiac Anatomies from Misaligned Multi-View Magnetic Resonance Images with Mesh Deformation U-Nets

High-quality three-dimensional (3D) representations of cardiac anatomy and function are crucial for improving cardiac disease diagnosis beyond strictly volume-based biomarkers used in current clinical practice, as well as for the accurate simulation of cardiac electrophysiology and mechanics. However, current gold standard cardiac magnetic resonance imaging (MRI) protocols typically only acquire a set of 2D slices to approximate the true 3D morphology of the underlying heart. In this work, we propose a novel geometric deep learning method, the \emph{Mesh Deformation U-Net}, to reconstruct 3D cardiac surface meshes from 2D MRI slices as the key part of a fully automatic end-to-end pipeline. Its architecture combines spectral graph convolutions and mesh sampling operations in a hierarchical encoder-decoder structure to enable efficient multi-scale feature learning directly on mesh data. A targeted preprocessing step approximately fits a template mesh to the sparse MRI contours, before the Mesh Deformation U-Net corrects for motion-induced slice misalignment by simultaneously utilising information from multiple MRI views and the template-induced anatomical shape prior. We evaluate the Mesh Deformation U-Net on a large synthetic dataset of heart anatomies and outperform multiple benchmark approaches while achieving small reconstruction errors below the pixel size of the underlying image resolution for three different cardiac substructures. Furthermore, we apply the pre-trained Mesh Deformation U-Net as the key component of a 4-step reconstruction pipeline to cine magnetic resonance images of the UK Biobank and observe realistic heart reconstructions on both a local and global level. We calculate multiple widely used clinical metrics for the reconstructed meshes and obtain values in line with other large-scale population studies.