A Generative Shape Compositional Framework to Synthesise Populations of Virtual Chimaeras

TL;DR

The paper tackles the challenge of building realistic virtual populations of multipart anatomy when training data exhibit partial overlap across structures. It introduces a generative shape compositional framework with two part-aware generators (independent gcVAEs and a dependent graph-convolutional mcVAE) and a self-supervised spatial composition network that performs affine and nonrigid registration to assemble complete whole-heart shapes from independently synthesized parts. Empirical results on UK Biobank cardiac meshes show that the proposed models achieve superior generalisability and clinical relevance compared with PCA, with nonrigid composition improving boundary coherence. The approach enables synthesising plausible, diverse virtual chimaeras suitable for in-silico trials and has potential to generalise to other multipart organ ensembles by leveraging partially overlapping datasets and weak labels.

Abstract

Generating virtual populations of anatomy that capture sufficient variability while remaining plausible is essential for conducting in-silico trials of medical devices. However, not all anatomical shapes of interest are always available for each individual in a population. Hence, missing/partially-overlapping anatomical information is often available across individuals in a population. We introduce a generative shape model for complex anatomical structures, learnable from datasets of unpaired datasets. The proposed generative model can synthesise complete whole complex shape assemblies coined virtual chimaeras, as opposed to natural human chimaeras. We applied this framework to build virtual chimaeras from databases of whole-heart shape assemblies that each contribute samples for heart substructures. Specifically, we propose a generative shape compositional framework which comprises two components - a part-aware generative shape model which captures the variability in shape observed for each structure of interest in the training population; and a spatial composition network which assembles/composes the structures synthesised by the former into multi-part shape assemblies (viz. virtual chimaeras). We also propose a novel self supervised learning scheme that enables the spatial composition network to be trained with partially overlapping data and weak labels. We trained and validated our approach using shapes of cardiac structures derived from cardiac magnetic resonance images available in the UK Biobank. Our approach significantly outperforms a PCA-based shape model (trained with complete data) in terms of generalisability and specificity. This demonstrates the superiority of the proposed approach as the synthesised cardiac virtual populations are more plausible and capture a greater degree of variability in shape than those generated by the PCA-based shape model.

Figures (4)

• Figure 1: Schematic illustration of our proposed shape compositional framework, which consists of a part-aware generator to learn the shape representations of each part and a composition network to perform the affine and nonrigid transformation to spatially compose the synthesised parts into a whole heart shape assembly.
• Figure 2: Schematic illustration of the graph convolution blocks. (a) is the architecture of the residual graph convolution block in the encoder, and (b) is that in the decoder.
• Figure 3: Examples of virtual chimaeras generated by the independent generator trained with complete data (IG-CO). Each component of the latent vector was randomly sampled from the posterior Gaussian distribution learnt from the training dataset.
• Figure 4: Box plots compare the median and interquartile range for all four cardiac indices of interest across the synthesised virtual cohorts and the UK Biobank population. Boxplots for each synthesised virtual cohort are labelled according to the corresponding method used: PCA, IG-CO, IG-PO, DG-CO and DG-PO.