Multi-Resolution Patch-Based Diffusion Priors for Short-Arc CBCT Reconstruction
Abstract
Purpose
To develop a kV-triggered short-arc intrafraction motion monitoring technique for prostate SBRT VMAT by enabling on-treatment reconstruction of a 3D prostate and nearby organs-at-risk (OARs) volume within seconds. We propose an iterative short-arc CBCT reconstruction framework regularized by a multi-resolution, patch-based diffusion prior to mitigate severe ill-posedness under limited-angle acquisition.
Methods
Two population diffusion priors were trained using patches from pelvic planning CTs of 66 patients: 64×64 (coarse) and 32×32 (refinement). Short-arc projection data spanning 90°, 60°, 45°, and 20° were simulated from planning CTs. Prostate-focused reconstruction was performed via diffusion posterior sampling with denoising diffusion implicit model (DDIM)-style updates and measurement-consistency steps within a prostate ROI defined by the CTV. Pixels outside the ROI were constrained using a strong prior derived from a setup (full-arc) CBCT to focus evaluation on prostate/OAR depiction. A coarse-to-fine strategy was employed, where the coarse model captured global anatomy and the refinement model recovered boundaries and fine detail, with data fidelity enforced through the measurement model. Reconstructions were compared using the Structural Similarity Index Measure (SSIM) with an iterative baseline using the same surrounding prior.
Results
Diffusion-regularized reconstructions improved qualitative prostate conspicuity relative to the baseline across all angular spans. Reconstructions using coarse model completed in ~10 s per volume on a modern GPU. SSIM for the 90° case increased from 0.876 to 0.901 relative to the baseline. Multi-resolution refinement improved edge sharpness and prostate/OAR boundary definition at moderate coverage (90°–45°). At severe sparsity (20°), reconstructions showed increased sensitivity to model-driven hallucination.
Conclusion
Diffusion-based posterior sampling is feasible for prostate-focused reconstruction from short-arc CBCT. A multi-resolution patch diffusion prior can enhance structural visibility under limited-angle conditions, while highlighting the need for explicit hallucination control as angular coverage decreases. These findings support further development toward markerless intrafraction monitoring and on-treatment assessment applications.