Title: Virtual Population Synthesis via Conditional Latent Diffusion Models Leveraging Fourier-Based Motion Representation

Abstract: The execution of in-silico trials for medical devices necessitates the creation of virtual populations comprising anatomical structures. In the context of cardiovascular research, virtual anatomies are generally depicted as 3D+t meshes derived from generative models. Nevertheless, the majority of current mesh generation techniques are limited to static structures, whereas sequence-based models frequently fail to incorporate explicit periodicity. Addressing this gap, we introduce 4D F-MeshLDM, a conditional generative architecture that integrates a convolutional mesh Variational Autoencoder (VAE) for mesh encoding, a structural latent space that models motion through a truncated Fourier series, and a diffusion prior designed to learn the latent distribution of Fourier coefficient tokens. By modulating the diffusion process with clinical covariates through affine modulation, the framework facilitates controllable synthesis. The process involves sampling tokens and executing inverse Fourier synthesis to produce cycle-consistent latent trajectories, which are subsequently decoded into 3D+t cardiac mesh sequences. Evaluations conducted on a dataset of 5,000 UK Biobank subjects reveal that 4D F-MeshLDM surpasses leading baselines in terms of anatomical accuracy, achieving a near-zero cycle closure error. Additionally, the synthesized cohorts maintain clinical functional indices with high fidelity, underscoring the framework’s promise for dependable in-silico cardiac trials.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC