Limited-Angle CBCT Reconstruction Via Conditional Diffusion-Based Projection Extrapolation
Abstract
Purpose
Limited-angle cone-beam CT (CBCT) reconstruction suffers from missing projection data, leading to severe streak artifacts, structural distortions, and degraded image quality. This study proposes a conditional diffusion-based projection extrapolation framework to recover missing sinogram regions from limited-angle CBCT data, enabling improved reconstruction quality while preserving physical consistency.
Methods
A conditional denoising diffusion probabilistic model (DDPM) was developed to extrapolate incomplete CBCT sinograms in the projection domain. Full-angle and limited-angle projection data were generated using the TIGRE toolbox under full-fan geometry. Paired sinogram slices were extracted from 10 simulated CBCT cases and resampled along the angular dimension to form 7,680 training pairs. The diffusion model was trained to denoise synthetically corrupted projection data, where full-angle projections were partially masked along the angular dimension to simulate limited-angle acquisition. The masked projections served as conditional inputs, enabling the model to learn a probabilistic mapping from incomplete to complete projection data. During inference, the reverse diffusion process was initialised from the limited-angle projections and applied slice-by-slice in the projection domain, with measured angular regions explicitly constrained to enforce data fidelity. The extrapolated projections were then reconstructed using the Feldkamp–Davis–Kress (FDK) algorithm. Reconstruction quality was evaluated using PSNR, SSIM, and RMSE for limited-angle scans of 120°, 150°, and 180°.
Results
The proposed diffusion-based approach effectively extrapolated missing projections, reducing wedge artifacts and improving image uniformity. During training, loss decreased from approximately 9×10⁻³ to below 6×10⁻⁴, with validation loss closely matching. PSNR remained stable at 10.4–10.6 dB, indicating good convergence and generalization.
Conclusion
This study demonstrates that conditional diffusion models provide an effective and flexible framework for limited-angle CBCT projection extrapolation. By modeling projection completion as a probabilistic denoising process, the proposed method improves reconstruction quality and artifact suppression without modifying the reconstruction algorithm.