Structure-aware Learning

Mesh4D is a motion-aware deep generative model for reconstructing high-resolution, temporally smooth 3D+t cardiac meshes directly from multi-view cardiac MRI. The method integrates a multi-view cross-attention encoder, transformer-based variational latent dynamics, and a continuous deformation decoder for anatomically consistent and physiologically plausible 4D heart reconstruction.

Status: Accepted at MICCAI 2025