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

purposeInternal anatomical motion challenges precise radiation delivery during external beam radiotherapy. Estimating and compensating for anatomical motion are essential for improving planned dose delivery to target volumes while sparing organs-at-risk. This research achieves accurate motion prediction using only planar X-ray imaging from conventional linear accelerators, without surrogate signals or invasive fiducial markers.

methodsWe propose Deep-Motion-Net: a patient-specific end-to-end graph neural network (GNN) enabling 3D volumetric organ reconstruction from single in-treatment kV planar X-ray images at arbitrary projection angles. A 2D convolutional neural network (CNN) encoder extracts image features, which four feature pooling networks fuse to a 3D template organ mesh. A ResNet-based graph attention network then deforms the feature-encoded mesh. Training uses synthetically generated organ motion instances and corresponding kV images, created by deforming a reference CT volume aligned with the template mesh, generating digitally reconstructed radiographs (DRRs) at required angles, and applying DRR-to-kV style transfer via conditional CycleGAN.

resultsQuantitative testing on synthetic respiratory motion scenarios and qualitative assessment on in-treatment images from four liver cancer patients demonstrated overall mean prediction errors of 0.16 ± 0.13 mm, 0.18 ± 0.19 mm, 0.22 ± 0.34 mm, and 0.12 ± 0.11 mm across datasets. Mean peak prediction errors were 1.39 mm, 1.99 mm, 3.29 mm, and 1.16 mm.

conclusionThis approach leverages accessible in-treatment imaging, avoiding expensive MRI systems or invasive markers. To the best of our knowledge, this is the first deep learning framework reconstructing volumetric 3D organ models from single-view images at arbitrary angles throughout an entire in-treatment scan series. Our approach achieves sub-millimetre accuracy when validated on synthetic motion instances and demonstrates clinical feasibility on real-treatment kV images, for which volumetric ground truth is inherently unavailable. The code is available at https://github.com/isurusuranga/DeepMotionNet .