SynthRCT

SynthRCT is a conditional generative framework for producing plausible, patient-specific anatomical transformations from a single CT volume. It was developed to move beyond the simplified perturbations commonly used to test proton-therapy plans, which cannot represent the full complexity of respiratory motion and longitudinal anatomical change.

Method

The model combines conditional variational learning with deformable image registration. During training, a posterior encoder learns from pairs of longitudinal CT scans, while an anatomy-conditioned prior learns which deformations are plausible for the input patient. A latent-conditioned decoder then predicts stationary velocity fields over overlapping axial slabs. These local predictions are refined, stitched into a coherent full-volume field and integrated before warping the source CT.

This design provides:

• patient-specific sampling from one reference CT;
• memory-scalable synthesis for large three-dimensional volumes;
• smooth latent-space traversal and interpolation; and
• direct export of warped CTs, velocity fields and displacement fields.

Phase-conditioned synthesis

The phase-conditioned model can also generate a complete respiratory cycle from a single source CT. The axial and coronal views below show the deformation evolving through ten breathing phases while preserving the patient-specific anatomical context.

Preprint · Official Publication · Weights · Code

Citation

@inproceedings{GuijaValiente2026SynthRCT,
  author    = {Guija-Valiente, Tomas and Malpica, Norberto and Rodriguez-Gonzalez, Blanca},
  title     = {SynthRCT: Scalable Conditional Deformation Synthesis for Synthetic Repeat CT Generation},
  booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
  publisher = {Springer Nature Switzerland},
  volume    = {17274},
  year      = {2026}
}