Modelling Multi-Phase Cardiac Anatomy Using Point Cloud Variational Autoencoders

The dynamic behaviour of the cardiac anatomy has a considerable impact on cardiac function and disease progression. Geometric deep learning methods have shown promising results for simulating the shape of biventricular anatomy, primarily emphasizing individual phases from the cycle. However, a more comprehensive understanding of cardiovascular disease could be achieved by analysing abnormalities in motion, which can be captured through modelling multiple phases together in the cardiac cycle. In this work, we propose a novel geometric deep learning architecture capable of encoding and predicting 3D shapes of the biventricular anatomy for end-diastolic and end-systolic phases of the cardiac cycle using only a single extreme phase as input. The reconstruction and prediction performance of the proposed architecture are accurate to sub-pixel resolution, with an average Chamfer distance of 1.58 (±0.16) mm. We also investigate the generative properties of the proposed model, finding the model can generate a range of diverse and realistic 3D cardiac shapes and has interpretability in its latent space.