Generative Priors for CBCT Reconstruction: A 3D Conditional Diffusion Model for Prostate Cancer CT
Abstract
Purpose
Inferior CBCT quality from artifacts or incomplete data can compromise anatomy visualization during Image-Guided Radiotherapy (IGRT), increasing uncertainty in target localization and organ-at-risk positioning. Improving CBCT reconstruction can enable more reliable daily guidance and adaptive decisions, potentially improving target coverage and normal tissue sparing. This study develops a 3D conditional patch diffusion model trained on planning Computed Tomography (PCT) to generate robust priors for CBCT reconstruction.
Methods
A total of 267 PCT scans were resampled to 256×256 pixels axially (2 mm isotropic). The model used 64×64×64 3D patches, randomly sampled and provided as a second channel alongside initial random noise. The diffusion model was trained to iteratively denoise the input, conditioned on patch data, with (noisy) and without (clean) added Gaussian noise, to learn realistic prostate anatomy priors. Evaluation was performed on an independent set of 20 clinical cases (10 PCTs and 10 CBCTs) under clean and noisy conditions. Image quality was quantified using Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), Mean Squared Error (MSE), and Mean Absolute Error (MAE) against corresponding clean ground truth.
Results
The model demonstrated high fidelity and noise robustness across modalities. For clean images, the PCT cohort achieved SSIM 0.959 ± 0.012 and PSNR 24.57 ± 4.52 dB, while CBCT achieved SSIM 0.960 ± 0.014 and PSNR 26.09 ± 4.76 dB. Under noisy conditions, PCT priors yielded SSIM 0.959 ± 0.018 and PSNR 26.90 ± 3.64 dB; CBCT priors achieved SSIM 0.931 ± 0.022 and PSNR 28.05 ± 4.31 dB. MSE and MAE remained consistently low.
Conclusion
We developed a 3D conditional patch diffusion model that generates high-quality prostate anatomical priors with consistent performance on PCT and CBCT, including under noise. This approach is well-suited for regularizing iterative CBCT reconstruction to improve image guidance in prostate radiotherapy, supporting more reliable localization and adaptation.