Toward CT-Equivalent Image Quality in Low-Dose Radiotherapy Planning: Conditional Diffusion-Based CBCT-to-CT Synthesis and the Impact of CBCT Input Representation

Technical Summary: Conditional Diffusion–Based CBCT-to-CT Synthesis and the Impact of CBCT Input Representation

Problem Statement
In adaptive radiotherapy and patient verification, repeated CT acquisitions are often necessary, leading to increased cumulative X-ray doses. While low-dose Cone Beam CT (CBCT) is routinely used during treatment delivery to mitigate this, its image quality is frequently insufficient for accurate dose calculation and planning. CBCT images suffer from scatter, noise, beam hardening, and reconstruction artifacts. Standard clinical CBCT images (stored as DICOM files) are reconstructed using vendor-specific pipelines optimized for visualization rather than quantitative accuracy, often involving proprietary post-processing (smoothing, intensity normalization) that can distort the physical relationship between projection data and voxel intensities. This study addresses the challenge of synthesizing CT-equivalent images from low-dose CBCT using deep learning, specifically investigating how the choice of CBCT input representation—standard DICOM versus raw-data-based Filtered Back-Projection (FDK)—affects the performance of conditional diffusion models.

Methodology
The study employs a supervised deep learning framework utilizing a Conditional Denoising Diffusion Probabilistic Model (DDPM) with a U-Net backbone and multi-scale self-attention blocks. The model is designed to iteratively denoise a random noise input, guided by a CBCT image, to generate a synthetic CT (sCT).

• Dataset: Data was acquired from a solid anthropomorphic head-and-neck phantom scanned at Ninewells Hospital using matched CT and CBCT protocols. The dataset comprised 920 aligned CT–CBCT slice pairs, split into 80% training (736 pairs) and 20% independent testing (184 pairs).
• Experimental Design: Two experiments were conducted using identical architectures, optimization strategies, and dataset splits. The sole variable was the conditioning input:
  1. Standard Clinical CBCT: Vendor-reconstructed DICOM images.
  2. FDK-Reconstructed CBCT: Images reconstructed directly from raw projection data using the Feldkamp-Davis-Kress (FDK) algorithm, applying minimal post-processing.
• Training: Models were trained for 1500 epochs using a combined Mean Squared Error (MSE) and Structural Similarity (SSIM) loss. CT images were normalized to a fixed Hounsfield Unit window (-1000 to 2000 HU) and resized to 256x256 pixels.

Key Contributions
The primary contribution of this work is the systematic integration and evaluation of FDK-reconstructed CBCT as a conditioning input for diffusion-based CBCT-to-CT synthesis. To the authors' knowledge, this is the first study to quantitatively assess how CBCT reconstruction methodology (specifically using raw projection data) influences the performance of conditional diffusion models for synthetic CT generation. The study isolates the impact of input representation on synthesis fidelity by comparing identical diffusion architectures trained on either standard DICOM or FDK inputs.

Results

• Baseline Characteristics: Before diffusion synthesis, standard DICOM CBCT appeared closer to the ground-truth CT (SSIM ≈ 0.83, PSNR ≈ 25.9 dB) due to vendor smoothing. In contrast, FDK CBCT exhibited lower baseline similarity (SSIM ≈ 0.26, PSNR ≈ 14.6 dB) but retained physically consistent intensity relationships and reduced low-frequency bias.
• Synthesis Performance: Both configurations achieved substantial improvements over their respective baselines.
  • DICOM-conditioned synthesis: Reached SSIM ≈ 0.93 and PSNR ≈ 36.0 dB.
  • FDK-conditioned synthesis: Achieved SSIM ≈ 0.85 and PSNR ≈ 37.4 dB.
• Relative Improvement: While the absolute SSIM was slightly higher for the DICOM baseline, the relative improvement provided by the diffusion model was significantly greater for the FDK input. The FDK-conditioned synthesis showed an average improvement of +0.55 SSIM and +20.1 dB PSNR over the baseline, compared to +0.12 SSIM and +11.8 dB PSNR for the DICOM-conditioned synthesis.
• Qualitative Findings: Visual inspection confirmed that FDK-conditioned synthetic CTs exhibited reduced streak and cupping artifacts, improved soft-tissue uniformity, and enhanced structural consistency at bone-air interfaces. Difference maps indicated lower residual errors for the FDK-conditioned configuration.

Significance and Claims
The paper claims that the representation of CBCT images used during conditional diffusion significantly influences synthesis performance. The key finding is that while standard DICOM CBCT appears superior at baseline due to vendor post-processing, FDK-reconstructed CBCT from raw projection data provides more "physically meaningful conditioning" for diffusion models. Consequently, the diffusion-based denoising process yields greater relative enhancement when conditioned on FDK CBCT.

This supports the hypothesis that reconstruction-aware AI pipelines, which utilize raw-data-based CBCT representations, can better support the suppression of acquisition-related artifacts and improve the fidelity of synthetic CT generation for radiotherapy. The authors conclude that future work should focus on patient-based validation, multi-scanner evaluation, and assessing the dosimetric impact in adaptive radiotherapy workflows, noting that access to clinical raw CBCT projection datasets remains a current limitation.