Hierarchical Mesh Variational Autoencoder for Shape Representation of Abdominal Organs

Abstract

Goal: The variability in soft organ shapes and positions across patients poses challenges for linear models in the reconstruction of significant local variations, while nonlinear models have difficulties with interpretability. This study aims to address these issues by proposing a mesh variational autoencoder with hierarchical latent variables (HMVAE) for 3D organ shape representation. Methods: Hierarchical latent variables capture both global and local organ features. Mesh templates ensure vertex correspondence across different resolutions. Liver and stomach meshes from 86 patients were used for training, with testing conducted in 19 patients. Results: The proposed method achieved mean vertex distances of 1.5 mm for the liver and 1.4 mm for the stomach, outperforming principal component analysis in interpolation tasks. Conclusions: The proposed HMVAE enables accurate and interpretable 3D organ reconstructions with hierarchical shape control.