Learning a Conditional Generative Model for Anatomical Shape Analysis

Abstract

We introduce a novel conditional generative model for unsupervised learning of anatomical shapes based on a conditional variational autoencoder (CVAE). Our model is specifically designed to learn latent, low-dimensional shape embeddings from point clouds of large datasets. By using a conditional framework, we are able to introduce side information to the model, leading to accurate reconstructions and providing a mechanism to control the generative process. Our network design provides invariance to similarity transformations and avoids the need to identify point correspondences between shapes. Contrary to previous discriminative approaches based on deep learning, our generative method does not only allow to produce shape descriptors from a point cloud, but also to reconstruct shapes from the embedding. We demonstrate the advantages of this approach by: (i) learning low-dimensional representations of the hippocampus and showing low reconstruction errors when projecting them back to the shape space, and (ii) demonstrating that synthetic point clouds generated by our model capture morphological differences associated to Alzheimer’s disease, to the point that they can be used to train a discriminative model for disease classification.