Ventricle Surface Reconstruction from Cardiac MR Slices Using Deep Learning

Reconstructing 3D ventricular surfaces from 2D cardiac MR data is challenging due to the sparsity of the input data and the presence of interslice misalignment. It is usually formulated as a 3D mesh fitting problem often incorporating shape priors and smoothness regularization, which might affect accuracy when handling pathological cases. We propose to formulate the 3D reconstruction as a volumetric mapping problem followed by isosurfacing from dense volumetric data. Taking advantage of deep learning algorithms, which learn to predict each voxel label without explicitly defining the shapes, our method is capable of generating anatomically meaningful surfaces with great flexibility. The sparse 3D volumetric input can process contours with any orientations and thus can utilize information from multiple short- and long-axis views. In addition, our method can provide correction of motion artifacts. We have validated our method using a statistical shape model on reconstructing 3D shapes from both spatially consistent and misaligned input data.