Generating Virtual Populations of 3D Cardiac Anatomies with Snowflake-Net

High-quality virtual populations of human hearts are of significant importance for a variety of applications, such as in silico simulations of cardiac physiology, data augmentation, and medical device development. However, their creation is a challenging endeavor since the synthesized hearts not only need to exhibit plausible shapes on an individual level but also accurately capture the considerable variability across the true underlying population. In this work, we present Snowflake-Net as a novel approach to automatically generate arbitrarily-sized and realistic populations of 3D heart anatomies in the form of high-resolution point clouds. Our proposed method combines transformer components with point cloud-based deep learning to effectively and directly process 3D heart anatomies reconstructed from cine magnetic resonance images. We develop our approach on a large UK Biobank dataset of about 1000 subjects. We find that the Snowflake-Net achieves average reconstruction errors of 0.90 mm in terms of mean Chamfer Distances, which is considerably below the pixel resolution of the underlying MRI acquisition, and outperforms a prior state-of-the-art approach by ∼ 20%. Furthermore, we show the Snowflake-Net’s ability to create new 3D cardiac anatomies with a high degree of realism on both an individual and population level and observe the generated virtual and the true underlying populations to be highly similar in terms of multiple generation quality metrics. Finally, we investigate how the captured 3D shape variability is encoded in the low-dimensional latent space and its effect on model interpretability.