CBCT-Based Synthetic CT Image Generation Using Conditional Denoising Diffusion Probabilistic Model

1. Introduction

• CBCT is widely used in IGRT, but artifacts and inaccurate HU values hinder quantitative applications and CBCT-based ART.The cited artifacts include streaking, shading, cupping, and scatter contamination.Source: §1 · p. 3
• Current approaches either correct CBCT artifacts or generate sCT images intended to approach pCT image quality.The introduction presents these as two major routes toward direct CBCT-based ART.Source: §1 · p. 3
• Hardware modifications can reduce scatter artifacts but may lower system quantum efficiency and reconstructed-image signal-to-noise ratio.Examples include anti-scatter grids, primary-modulation beam filters, and lattice-shaped beam stoppers.Source: §1 · p. 3
• Monte Carlo scatter estimation is accurate but computationally intensive, while analytical kernels are faster but less effective for heterogeneous structures; learning-based correction is artifact-source-specific.The source-specific limitation is described as common to artifact-reduction methods.Source: §1 · p. 3
• Paired CBCT–dpCT translation can suffer anatomical mismatch; proposed alternatives include unpaired Cycle-GAN consistency and added structure-preserving modules or losses.Gradient-difference loss and attention mechanisms are cited as examples of added strategies.Source: §1 · pp. 3–4
• DDPM and other diffusion models have been used for medical-image reconstruction, MRI–CT conversion, denoising, and generation.The introduction characterizes diffusion models as analytically principled, easy to train, and capable of state-of-the-art image quality.Source: §1 · p. 4
• The proposed conditional DDPM trains on CBCT–dpCT pairs and gradually transforms Gaussian noise into an sCT conditioned on its CBCT counterpart.The dpCT serves as the target distribution and CBCT as the condition.Source: §1 · p. 4
• Brain and H&N experiments found that conditional DDPM generated sCT with accurate HU values and reduced artifacts, supporting direct CBCT-based ART.The authors state that these results make direct CBCT-based ART implementation possible.Source: §1 · p. 4

2.1. Image acquisition and preprocessing

• The study included 41 brain and 47 H&N patients; brain training used 4682 slices from 30 patients and testing used 500 slices from 11 others.For H&N, 4314 slices from 37 patients were used for training and 500 slices from 10 other patients for testing.Source: §2.1 · p. 4
• CT images were deformably registered and resampled to match corresponding CBCT images, with identical data volume and voxel size; all images were cropped to 256 × 256 and normalized to [-1, 1] before network input.Merged body contours extracted from CBCT and CT images were applied to both images.Source: §2.1 · p. 4

2.2. Denoising diffusion probabilistic models (DDPM)

• DDPM gradually corrupts target images with Gaussian noise, trains a network to predict that noise, and reverses the process from Gaussian noise to recover an image.The noise-prediction objective supports estimating the reverse transition at each step.Source: §2.2 · p. 5

2.3. The proposed conditional DDPM for synthetic CT from CBCT images

• The proposed model concatenates the input CBCT with the noisy CT sample at each step, conditioning the noise estimator and reverse process on CBCT.This makes the task conditional image-to-image translation rather than generation of a random sample from the CT distribution.Source: §2.3 · p. 6

2.4. Compared conditional DDPM approaches

• Uncond-1 applies ILVR during inference, while Uncond-2 refines unconditional sampling using a progressively weakened CBCT condition.ILVR uses a low-pass-filtered CBCT constraint to preserve high-level semantics; both approaches use an unconditionally trained model.Source: §2.4 · p. 8
• Adap-Cond-1 and Adap-Cond-2 use adaptively noised CBCT conditions during training and sampling instead of the proposed method’s constant condition.Adap-Cond-1 conditions prediction on y_(t-1), whereas Adap-Cond-2 uses y_t so the CBCT and CT conditions have the same noise level.Source: §2.4 · p. 8

2.5. Implementation and evaluation

• A single time-embedded U-net with attention modules and residual blocks predicts noise across 1000 steps, with variance scheduled linearly from β_1 = 10^-4.The same noise-prediction model is used at every diffusion step.Source: §2.5 · p. 11
• MAE, PSNR, and NCC were calculated between sCT images and CT images taken as ground truth.Source: §2.5 · p. 11
• MAE measures voxel-based HU difference between the original CT and sCT, while PSNR measures whether predicted sCT intensity is evenly or sparsely distributed.Source: §2.5 · pp. 11–12
• NCC measures similarity between CT and sCT as a function of displacement.Source: §2.5 · p. 12

3.1. Visual quality improvement

• Brain CBCT beam-hardening artifacts were significantly suppressed in sCT while fine structures, including contrast agent and maxillary sinus, were preserved.The paper notes that preserving these structures can improve organ-segmentation accuracy.Source: §3.1 · p. 12
• In H&N images, sCT appeared artifact-free for two slices and suppressed artifacts better than dpCT in extremely metal-degraded slices.The CBCT images had severe streaking artifacts from motion and metal implants.Source: §3.1 · p. 13
• Across 500 testing slices per study, sCT HU-histogram shapes and peaks were closer to dpCT than CBCT, indicating improved HU fidelity.Source: §3.1 · p. 13

3.2. Quantitative analysis

• 25.99 HU MAE for brain sCT versus 40.63 HU for CBCT; H&N MAE fell to 32.56 HU from 38.99 HU.Brain sCT PSNR and NCC were 30.49 dB and 0.99, compared with 27.87 dB and 0.98 for CBCT; H&N PSNR and NCC improved slightly.Source: §3.2 · p. 14

3.3. Comparison studies

• ILVR preserved a CBCT-like outline but showed distortions and artifacts, whereas Uncond-2 performed better and preserved more image detail.The text attributes Uncond-2's advantage to directly enforcing the image condition during sampling.Source: §3.3 · p. 15
• Weak adaptive conditioning produced different failures: Adap-Cond-1 retained most anatomy with some distortion, while Adap-Cond-2 failed to preserve the body contour.Source: §3.3 · p. 15
• Residual artifacts remained among the four compared methods, and weak guidance caused artifacts in some cases.Source: §3.3 · pp. 15–16

4. Discussion

• To the authors’ knowledge, this was the first reported use of a diffusion model for CBCT–CT translation.Source: §4 · p. 16
• Supervised medical image translation commonly lacks exactly matched training pairs, and matched CBCT–CT images are described as impossible to obtain.Source: §4 · p. 16
• In the H&N study, conditional DDPM significantly reduced metal artifacts, producing sCT superior to the artifact-corrupted dpCT counterparts used for training.Source: §4 · pp. 16–17
• Because dpCT is not anatomically matched to sCT, sCT–dpCT quantitative metrics are unfair to some extent, while sCT seems closer to artifact-free ground truth.Source: §4 · p. 17
• The discussion proposes further study of unsupervised diffusion for sCT generation, noting that its conditional form requires unconditional training and conditional sampling.In this setup, standard conditional-probability condition enforcement is inaccessible, and possible solutions are left for future work.Source: §4 · p. 17
• Stochastic Gaussian noise makes it difficult to incorporate prior information during training, leaving sampling quality dependent on network manifold learning and insufficient for inter-channel correlations such as DECT.Source: §4 · p. 17
• Sampling time is a major obstacle to clinical use, although diffusion-model training time is comparable to one-step U-net models.With T steps, generation takes approximately T times as long as with U-net or GAN models, making online ART impractical.Source: §4 · p. 17
• The method uses 2D slices without patch extraction; 3D patch-based models could better preserve structure and capture spatial relationships, while other diffusion alternatives may outperform DDPM.Source: §4 · p. 17

• The conditional DDPM generates synthetic CT (sCT) from CBCT images.Source: §5 · p. 18
• Improved HU accuracy makes the generated sCT suitable for dose calculation.Source: §5 · p. 18
• Fewer artifacts in sCT than in CBCT could improve CBCT-based online segmentation and replanning.Source: §5 · p. 18
• The method supports further quantitative applications of CBCT and may enhance the potential for clinical CBCT-based adaptive radiotherapy.Source: §5 · p. 18