Source-linked AI summary
SynthRCT: Scalable Conditional Deformation Synthesis for Synthetic Repeat CT Generation
Tomas Guija-Valiente, Blanca Rodriguez-Gonzalez, Norberto Malpica
TL;DR
SynthRCT addresses the limited realism of predefined anatomical perturbations used to evaluate proton-therapy robustness. It learns conditional latent deformation modes and decodes them into locally generated SVFs that form coherent full-volume transformations. In respiratory 4DCT experiments, the method reduced peak allocated GPU memory while producing plausible, spatially regular deformations, within the scope of the evaluated data.
Problem
Predefined perturbations poorly capture complex, patient-specific anatomical variability, while scalable conditional deformation generation remains challenging for large-field-of-view 3D CT.
Method
SynthRCT learns an anatomy-conditioned latent deformation distribution and decodes sampled latent codes into local SVFs over axial slabs, which are assembled and integrated into full-volume transformations.
Results
41.6% reduction in peak allocated GPU memory compared with full-volume decoding, while generated transformations achieved competitive alignment quality and preserved spatial regularity.
Takeaways & Limitations
SynthRCT enables sampling and interpolation of plausible anatomical deformations for large field-of-view CT data.
Takeaways & Limitations
The respiratory 4DCT data have restricted variability dominated by breathing motion, limiting assessment of whether the latent space disentangles multiple deformation modes.
Abstract
from arXiv · showhide
In proton therapy, plans are typically optimized on a single planning CT, making robustness evaluation essential under anatomical changes. However, current scenarios often rely on simplified perturbations that poorly capture complex, patient-specific variability. We propose SynthRCT, a scalable conditional generative framework for 3D anatomical deformation synthesis. Based on a conditional variational autoencoder, SynthRCT learns a latent deformation space and decodes sampled latent codes into local stationary velocity fields conditioned on an input anatomy. Local fields are assembled into coherent full-volume transformations, enabling memory-scalable generation for large field-of-view CT data. We validate the approach on respiratory 4DCT data with multiple breathing-phase anatomies per subject. SynthRCT enables patient-specific sampling of plausible anatomical transformations beyond predefined robustness scenarios. Code available at: https://github.com/TomasGuija/SynthRCT.
1 Introduction
SynthRCT addresses the limits of predefined perturbations by generating patient-specific anatomical deformations with a scalable conditional framework. It combines latent deformation modeling with local SVF composition for coherent full-volume transformations and evaluates multiple aspects of synthesis quality.
- Motivation: Predefined planning-CT perturbations only approximate the complex, patient-specific anatomical variability relevant to proton-therapy robustness evaluation.Registration methods estimate transformations between observed image pairs, whereas SynthRCT targets multiple plausible transformations from a reference anatomy.
- Motivation: Existing deformation-generation approaches remain challenging to scale to large-field-of-view 3D CT because volumetric fields are memory-intensive and local predictions must remain coherent.
- SynthRCT: SynthRCT learns global latent deformation modes from intra-patient CT pairs and decodes sampled codes into local stationary velocity fields over axial slabs.The slabs cover the full in-plane field of view while limiting the axial extent of each prediction.
- SynthRCT: Local stationary velocity fields are assembled and integrated into coherent full-volume transformations, reducing memory requirements while preserving high-resolution local deformation modeling.This design targets large field-of-view CT data.
- Evaluation: The framework contributes patient-specific repeat CT synthesis and a scalable local SVF generation and composition strategy.Experiments assess registration accuracy, deformation regularity, latent-space consistency, and landmark-distribution agreement.
2 Methods
SynthRCT models patient-specific anatomical deformations with a conditional latent representation and generates them as locally decoded stationary velocity fields. Overlapping slab predictions are refined, assembled, integrated, and used to warp the input CT into a coherent synthetic repeat CT.
- Conditional latent deformation model: SynthRCT learns a conditional distribution of plausible target anatomies from an input anatomy using latent deformation codes sampled from an anatomy-conditioned prior.The prior is regularized toward N(0, I), while a posterior inferred from moving–fixed CT pairs approximates training deformation codes.
- SVF parameterization and integration: Stationary velocity fields parameterize deformation and are integrated with scaling and squaring into smooth transformations.The method uses SVFs because sufficiently smooth fields integrate to smooth invertible transformations, reducing the risk of non-physical foldings relative to unconstrained displacement prediction.
- Slab-wise deformation generator: To scale synthesis to large CT volumes, full-volume encoders capture global deformation modes while a generator decodes each axial slab and latent code into a local SVF.FiLM injects the latent code into a U-Net decoder, combining global latent conditioning with local anatomical decoding.
- Overlap refinement and full-volume assembly: Neighboring slabs share one latent code, and their overlapping SVFs are refined before assembly into a single full-volume field.A convolutional refiner uses local SVF predictions and moving-image context to predict residual corrections for overlap regions.
- Training and synthesis: The assembled full-volume SVF is integrated into a transformation that warps the input image to produce the synthetic repeat CT.Training uses sampled neighboring slab pairs, with local normalized cross-correlation and intermediate-resolution deep supervision for image-based learning.
- Training and synthesis: The latent space is trained to align the pair-based posterior with the anatomy-conditioned prior while regularizing that prior toward a standard Gaussian.The KL objective includes terms for posterior–prior matching and prior regularization, with weighting and KL warm-up factors specified in the training objective.
3 Experiments
SynthRCT is evaluated on respiratory 4DCT for alignment, deformation regularity, landmark-distribution agreement, and latent-space behavior. The experiments show topology-preserving transformations, millimetre-scale agreement with observed respiratory configurations, and smooth, interpretable latent deformations.
- Evaluation setup: Experiments use DIR4DCT with 10 thoracic patients, 10 respiratory phases, and 75 annotated landmark trajectories per subject.Training and evaluation use intra-patient phase pairs separated by at least two respiratory positions.
- Registration accuracy: SynthRCT improves alignment over the unregistered moving image but remains below deterministic registration baselines, consistent with its generative objective.Alignment is assessed using LNCC, RMSE, and lung Dice overlap.
- Deformation regularity: Approximately zero folding is observed across held-out patients, indicating topology-preserving full-volume transformations.Folding is defined as the fraction of voxels with det ∇ϕ ≤0.
- Landmark distribution evaluation: Generated landmark distributions closely match observed respiratory configurations, with millimetre-scale configuration discrepancies.Evaluation uses distribution-level distances, coverage, precision, and spread-ratio metrics.
- Latent-space interpolation: Latent interpolation produces intermediate respiratory anatomies with gradually increasing SVF magnitude between identity and the largest observed deformation.The interpolation is evaluated on a held-out patient using generated warped images, 3D overlays, and SVF magnitude.
- Latent-space organization: PCA analysis associates displacement along the main latent direction with increasing deformation magnitude and suggests distinct respiratory expansion and contraction behaviors.The analysis uses more than 300 posterior latent codes from moving–fixed training pairs.
4 Conclusion
SynthRCT generates plausible 3D anatomical deformations from a single input CT by learning global latent modes and assembling local stationary velocity fields into coherent full-volume transformations. It supports memory-scalable sampling for large-field-of-view CT, but its learned variability is constrained by respiratory 4DCT data.
- SynthRCT synthesizes plausible 3D anatomical deformations from a single input CT using global latent deformation modes.
- Local stationary velocity fields are assembled before integration into coherent full-volume transformations.
- 41.6% lower peak allocated GPU memory was achieved versus full-volume decoding for large field-of-view CT data.
- Generated transformations achieved competitive alignment quality while preserving spatial regularity, although pairwise registration was not the primary objective.
- The respiratory 4DCT data provide restricted variability, with breathing motion dominating the learned deformation space.