Cardiac Mesh Flow: One-Step Generation of 3D+t Cardiac Four-Chamber Meshes via Flow Matching

A team led by Qiang Ma has introduced Cardiac Mesh Flow, a novel generative framework for creating dynamic 3D+t cardiac meshes of the four heart chambers. Unlike previous approaches that either used volumetric representations lacking anatomical correspondence or relied on VAE-based models that traded off reconstruction fidelity for generative diversity, Cardiac Mesh Flow leverages the flow matching technique. This allows one-step generation of multi-scale free-form deformation fields that warp a single template mesh into realistic, time-varying heart geometries across the entire cardiac cycle. The model inherently preserves point-wise anatomical correspondence between different time points and subjects, a critical requirement for population-level shape analysis.

Cardiac Mesh Flow also supports conditional generation controlled by cardiac chamber volumes, giving researchers precise control over synthetic heart characteristics. Experimental results demonstrate state-of-the-art performance on both unconditional and conditional generation tasks, producing meshes with high fidelity and diversity. This work has significant implications for large-scale computational cardiology, enabling realistic synthetic datasets for training deep learning models, simulating pathologies, and studying heart shape and motion variability without the need for costly manual annotations or extensive patient scans.