Interpretable cardiac anatomy modeling using variational mesh autoencoders.

Area of Science:

• Computational geometry
• Medical imaging
• Machine learning

Background:

• Cardiac anatomy exhibits significant population variability, impacting clinical diagnosis and treatment.
• Existing computational methods aim to model this variability for applications like disease prediction and virtual population generation.

Purpose of the Study:

• To introduce a novel geometric deep learning approach, the variational mesh autoencoder (mesh VAE), for modeling population-wide cardiac shape variations.
• To evaluate the mesh VAE's performance in reconstructing 3D cardiac meshes and its utility in clinical prediction and virtual population synthesis.

Main Methods:

• Developed a hierarchical variational autoencoder (VAE) framework incorporating multi-scale graph convolutions and mesh pooling.
• Applied the mesh VAE to a dataset of 3D cardiac meshes from over 1,000 patients with acute myocardial infarction.
• Compared mesh VAE performance against a voxelgrid-based deep learning benchmark.

Main Results:

• Achieved low reconstruction errors for 3D cardiac meshes, with mean surface distances below image resolution.
• Outperformed a voxelgrid-based benchmark in mean surface and Hausdorff distances, with significantly reduced memory requirements.
• Demonstrated improved prediction of major adverse cardiac events and generation of realistic virtual cardiac populations.

Conclusions:

• The mesh VAE is an efficient and effective deep learning method for modeling cardiac shape variability.
• This approach offers potential for enhanced clinical diagnosis, treatment planning, and the creation of synthetic cardiac datasets.