3D Mitral Valve Surface Reconstruction from 3D TEE via Graph Neural Networks

Mitral valve insufficiency is a condition in which the valve does not close properly, and blood leaks back into the atrium from the ventricle. Valve assessment for surgery planning is typically performed with 3D transesophageal echocardiography (TEE). The simulation of the resulting valve dynamics can support selecting the most promising surgery strategy. These simulations require an accurate reconstruction of the open valve as a 3D surface model. 3D mitral valve reconstruction from image data is challenging due to the fast-moving and thin valve leaflets, which might appear blurred and covered by very few voxels depending and spatio-temporal resolution. State-of-the-art voxel-based CNN segmentation methods need an additional processing step to reconstruct a 3D surface from this voxel-based representation which can introduce unwanted artifacts. We propose an end-to-end deep-learning-based method to reconstruct a 3D surface model of the mitral valve directly from 3D TEE images. The suggested method consists of a CNN-based voxel encoder and decoder inter-weaved with a graph neural network-based (GNN) multi-resolution mesh decoder. This GNN samples feature vectors from the CNN-decoder at different resolutions to deform a prototype mesh. The model was trained on 80 sparsely annotated 3D TEE images (1 mm 3 voxel resolution) of the valve during end-diastole. Each time frame was annotated by two cardiovascular experts on nine planes rotated around the axis through the apex of the left ventricle and the center of the mitral valve. Our method’s average bidirectional point-to-point distance is 1.1 mm, outperforming the inter-observer point-to-point distance of 1.8 mm.