Deep-Motion-Net: GNN-based volumetric liver shape reconstruction from single-view 2D projections

1

Deep-Motion-Net utilizes a GNN to reconstruct 3D liver shapes from single-view kV X-ray images at arbitrary gantry angles.
2

The model integrates CNN-extracted features with mesh node displacements, enabling end-to-end training for patient-specific organ meshes.
3

Projection angle information is encoded as an additional input channel, allowing the model to learn angle-dependent anatomical features.
4

The architecture includes a 2D CNN encoder and a GNN-based mesh deformation network to enhance the accuracy of volumetric reconstructions.
5

Deep-Motion-Net addresses limitations of existing methods by reconstructing 3D anatomy from limited-FOV projections without fixed angles.

Int. Journal Of Computer Assisted Radiology And Surgery

by Isuru Wijesinghe, Michael Nix, Arezoo Zakeri, Alireza Hokmabadi, Bashar Al-Qaisieh, Ali Gooya, Zeike Taylor
May 13, 2026

1