SynthRCT: Scalable Conditional Deformation Synthesis for Synthetic Repeat CT Generation

Summary

The paper presents SynthRCT, a conditional variational autoencoder framework that synthesizes plausible 3D anatomical deformations from a single planning CT, targeting robustness evaluation in proton therapy. Unlike pairwise registration, the model learns a distribution of deformations: a posterior encoder infers latent codes from moving–fixed pairs during training, while an anatomy-conditioned prior supports sampling at inference. To handle large field-of-view volumes, SVFs are decoded per axial slab (48 slices, 50% overlap) with FiLM-conditioned latent codes, merged by a learned convolutional refiner that corrects overlap discontinuities, and integrated via scaling-and-squaring into full-volume diffeomorphic transformations. Validation uses the DIR4DCT dataset (10 thoracic 4DCT patients, 75 landmark trajectories each). Posterior-encoded alignment (LNCC 0.751, lung Dice 0.950) improves over the unregistered baseline (0.667/0.928) but trails SyN (0.789/0.965) and VoxelMorph (0.783/0.962), which the authors attribute to the generative bottleneck. Sampled deformations yield approximately zero foldings and cut peak GPU memory by 41.6% versus full-volume decoding. Landmark-distribution agreement on one held-out patient is tight (Energy distance 1.44 mm, W1 2.61 mm, spread ratio 0.98), and PCA/interpolation show an interpretable latent magnitude axis. Main limitations: respiratory-only variability, no generative or dosimetric baselines despite the proton-therapy motivation, single-patient landmark evaluation, and implementation details deferred to the code repository.