Deep Implicit Neural Representations for End-to-End Anatomical Shape Estimation from Volumetric Images

Max-Heinrich Laves, Steffen Schuler, Ahmed Abbas, David Paik, Raphael Prevost, Oliver Zettinig

Keywords: graph convolutional network, 3D shape generation, segmentation, cardiac

TL;DR: Train mesh estimation networks without target meshes using deep signed distance fields.

Abstract: We present ImplicitMeshNet, an end-to-end approach for anatomical shape estimation from volumetric images using deep implicit neural representations. Our neural network directly reconstructs shapes as 3D meshes and is trained on voxel-based segmentation maps by utilizing a deep signed distance field transform, eliminating the need for explicit ground truth meshes. Evaluated on cardiac CT scans from the MMWHS challenge dataset, our method achieves a Dice score of 0.92 for the extraction of the left atrium and ventricle, while maintaining anatomical fidelity. This enables more accurate cardiac modeling for visualization and downstream analysis in clinical settings.

Submission Number: 19