EPC-RawDC: Equivariant Physics Constrained 3D Diffusion with Raw Data Consistency for CBCT to CT Image Synthesis

Alzahra Altalib; Chunhui Li; Alessandro Perelli

cs.CVOct 29, 2026

EPC-RawDC: Equivariant Physics Constrained 3D Diffusion with Raw Data Consistency for CBCT to CT Image Synthesis

Authors: Alzahra Altalib, Chunhui Li, Alessandro Perelli

Organizations: Faculty of Science, Engineering and Business, University of Dundee, DD1 4HN, UK. · Faculty of Applied Medical Sciences, Jordan University of Science and Technology, Irbid, 21410, Jordan · School of Cardiovascular and Metabolic Health, University of Glasgow, G12 8TA (UK).

Abstract

Cone beam computed tomography (CBCT) is widely used for image guided and adaptive radiotherapy, but scatter, noise, reconstruction artifacts, and inaccurate Hounsfield units limit its quantitative use. Recent diffusion approaches for CBCT to CT synthesis remain largely image driven, with limited use of the underlying acquisition physics. We propose EPC-RawDC, a conditional 3D latent diffusion framework that combines volumetric generative modeling with two complementary constraints derived from the same scanner specific CBCT forward model. First, geometry equivariant projection constraint exploits the rotational equivariance of the CBCT forward operator by supervising the synthesized CT through correspondingly transformed reference projections. Second, raw data consistency uses noisy projection measurements of the individual acquisition as scan specific supervision during training. The framework was evaluated on two independent clinical cohorts with distinct scanner geometries and projection sources. For JUST, the physics constraints used measured clinical raw CBCT projections acquired with an Elekta system, whereas for SynthRAD2025 the projection data were simulated from the reference CT using a Varian geometry with Poisson photon statistics. EPC-RawDC achieved an SSIM of 0.97 in both cohorts and improved PSNR over DDPM by 2.66 dB on SynthRAD2025 and 2.10 dB on JUST, while reducing mean MAE by 20.3% and 19.0%, respectively. On SynthRAD2025, projection consistency and equivariant errors were reduced by 55.0% and 41.4% relative to DDPM. These results demonstrate that combining a volumetric diffusion prior with scanner geometry and scan specific projection measurements improves quantitative CBCT to CT synthesis across both simulated and measured projection domains. Code available after publication at https://github.com/ALZAHRAALTALIB/EPC-RawDC.

Figures & tables

Panel of Figure 1 — Fig. 1: Overview of the proposed EPC-RawDC framework. The reference CT $\mathbf{x}_{0}$ and input CBCT $\mathbf{x}_{c}$ are encoded into volumetric latent representations, and conditional 3D diffusion produces the synthesized CT $\hat{\mathbf{x}}_{0}$ . Geometry equivariance links an image domain rotation $\mathbf{R}_{\phi}$ to the corresponding angular translation $\mathbf{T}_{\phi}$ in projection space, while RawDC constrains $\hat{\mathbf{x}}_{0}$ using the measured CBCT projections.

	SynthRAD2025	JUST
Source	Multicenter clinical CT	Clinical CT and CBCT
Patients	96 train / 24 test	32 train / 8 test
3D grid	$50\times 256\times 256$	$50\times 256\times 256$
Voxel spacing	$1\times 1\times 3$ mm	$1\times 1\times 3$ mm
CBCT geometry	Varian	Elekta
Detector	$768\times 1024$	$512\times 512$

TABLE I: Summary of the clinical datasets and CBCT acquisition models used in the study.

Panel of Table 3 — TABLE II: Average and Standard Deviation quantitative comparison at patient level: DDPM, EPC-3D-Diff, RawDC, EPC-RawDC, 3D ResUNet, 3D CycleGAN, on Single Domain SynthRAD2025 and JUST Test Sets for axial orientation.

Fig. 2: Qualitative comparison after single domain training for the SynthRAD2025 and JUST clinical test sets using (a) the full HU range and (b) the soft tissue window. Columns show the reference CT ( $\mathbf{x}_{0}$ ), CBCT ( $\mathbf{x}_{c}$ ), synthesized CT ( $\hat{\mathbf{x}}_{0}$ ) for the ablation and comparison methods, and the corresponding absolute error maps.

Method	SSIM $\uparrow$	PSNR (dB) $\uparrow$	MAE (HU) $\downarrow$
CBCT vs CT	$0.63\pm 0.02$	$27.88\pm 0.66$	$38.84\pm 3.11$
DDPM	$0.87\pm 0.01$	$34.80\pm 0.56$	$16.70\pm 0.88$
RawDC	$0.92\pm 0.02$	$35.65\pm 0.64$	$13.51\pm 1.27$
EPC-RawDC	$\mathbf{0.92\pm 0.02}$	$\mathbf{36.12\pm 0.63}$	$\mathbf{12.75\pm 1.31}$

TABLE III: Cross dataset generalization from SynthRAD2025 training to the JUST test cohort using Varian geometry.

Method	SSIM $\uparrow$	PSNR (dB) $\uparrow$	MAE (HU) $\downarrow$
CBCT vs CT	$0.57\pm 0.06$	$25.74\pm 0.82$	$47.89\pm 7.16$
DDPM	$0.90\pm 0.01$	$33.98\pm 1.39$	$16.23\pm 3.64$
RawDC	$0.94\pm 0.01$	$33.98\pm 1.48$	$15.51\pm 3.91$
EPC-RawDC	$\mathbf{0.95\pm 0.01}$	$\mathbf{34.02\pm 1.51}$	$\mathbf{15.13\pm 4.04}$

TABLE IV: Cross dataset generalization from JUST training to SynthRAD2025 test cohort using Elekta geometry.

	Method	SSIM $\uparrow$	PSNR (dB) $\uparrow$	MAE (HU) $\downarrow$
RAD2025	CBCT vs CT	$0.66\pm 0.03$	$27.65\pm 0.96$	$37.53\pm 3.50$
	DDPM	$0.94\pm 0.02$	$34.08\pm 1.03$	$16.86\pm 3.36$
	EPC-RawDC	$\mathbf{0.95\pm 0.02}$	$\mathbf{34.57\pm 1.06}$	$\mathbf{15.76\pm 3.16}$
JUST	CBCT vs CT	$0.60\pm 0.02$	$25.50\pm 0.54$	$49.26\pm 3.73$
	DDPM	$0.94\pm 0.01$	$34.19\pm 1.19$	$16.65\pm 2.60$
	EPC-RawDC	$\mathbf{0.95\pm 0.01}$	$\mathbf{34.54\pm 1.41}$	$\mathbf{15.68\pm 2.76}$

TABLE V: Performance of the jointly trained models on the held out SynthRAD2025 and JUST test cohorts. The models were trained using both datasets and evaluated on each domain.

Fig. 3: Coronal comparison of 2D and 3D CBCT to CT synthesis on representative SynthRAD2025 slices using full HU and soft tissue window. Columns show the reference CT, CBCT, 2D DDPM, 2D EqDiff, 2D CycleGAN, 3D DDPM, EPC-3D-Diff, 3D CycleGAN, and EPC-RawDC, and the absolute HU error for 2D DDPM, 3D DDPM, and EPC-RawDC. PSNR and SSIM are reported for the synthesis images and MAE for the error maps.

Test	Training	SSIM $\uparrow$	PSNR $\uparrow$	MAE $\downarrow$
RAD	JUST + RAD2025	$0.95\pm 0.02$	$34.57\pm 1.06$	$15.76\pm 3.16$
RAD	RAD2025 only	$0.97\pm 0.02$	$39.18\pm 1.82$	$8.67\pm 2.81$
JUST	JUST + RAD2025	$0.95\pm 0.01$	$34.54\pm 1.41$	$15.68\pm 2.76$
JUST	JUST only	$0.97\pm 0.01$	$38.00\pm 3.68$	$9.28\pm 2.91$

TABLE VI: Comparison of single domain and joint domain EPC-RawDC training across the SynthRAD2025 and JUST geometries.

Method	SSIM $\uparrow$	PSNR (dB) $\uparrow$	MAE (HU) $\downarrow$
CBCT vs CT	$0.66\pm 0.04$	$26.91\pm 0.91$	$43.02\pm 6.50$
2D DDPM	$0.66\pm 0.12$	$21.81\pm 12.69$	$58.35\pm 26.91$
2D EqDiff	$0.73\pm 0.05$	$24.28\pm 1.03$	$41.27\pm 7.57$
2D CycleGAN	$0.78\pm 0.05$	$26.55\pm 1.50$	$35.44\pm 9.50$
3D DDPM	$0.89\pm 0.02$	$36.37\pm 1.19$	$15.28\pm 1.92$
3D CycleGAN	$0.94\pm 0.02$	$35.14\pm 1.37$	$14.27\pm 3.45$

TABLE VII: Patient level comparison of 2D and 3D CBCT to CT synthesis across the SynthRAD2025 test subjects under the coronal evaluation protocol.

Fig. 4: Patient level distributions of SSIM, PSNR, and MAE for (a) SynthRAD2025 and (b) JUST. DDPM, EPC-3D-Diff, RawDC, and EPC-RawDC are compared using matched patient level measurements. Boxes indicate the median and interquartile range, whiskers extend to 1.5 times the interquartile range, and individual points denote test patients. The annotation “EPC-RawDC vs DDPM” indicates the number of patients for which EPC-RawDC outperformed DDPM for the corresponding metric.

Panel of Table 12 — TABLE VIII: Direct evaluation of the physical consistency constraints across the 24 SynthRAD2025 test cases.

Medical Image Synthesis

Medical image synthesis creates, translates, reconstructs, and edits clinical images to fill missing data and support analysis.

See all 312 papers

May 19, 2026

EPC-3D-Diff: Equivariant Physics Consistent Conditional 3D Latent Diffusion for CBCT to CT Synthesis

Cone-beam CT (CBCT) is routinely acquired during radiotherapy for patient setup, but its quantitative reliability is degraded by scatter, noise, and reconstruction artifacts, limiting Hounsfield Unit (HU) accuracy. We propose EPC-3D-Diff, a novel conditional 3D latent diffusion framework for volumetric CBCT to CT synthesis that introduces a projection domain equivariance loss derived from acquisition physics. Unlike common image domain equivariance, we exploit the fact that an in plane rotation of the volume corresponds to an angular shift in its projections. During training, we enforce this relationship by forward projecting rotated synthesized CT volumes and matching them to appropriately angle shifted projections of the paired target CT, yielding a physics consistent equivariance constraint integrated into the diffusion objective. To capture full 3D context efficiently, conditional diffusion is performed in a compact latent space learnt by a lightweight 3D autoencoder, preserving axial depth while downsampling in plane resolution for stable training. We validate on a paired head CBCT/CT phantom dataset, including repeat scans, and paired clinical data using patient wise splits, and perform single and mixed domain training, ablations, and comparisons with diffusion and CycleGAN. EPC-3D-Diff generalizes well and achieved substantial improvements, +7.4 dB (phantom) and +1.8 dB (clinical data) in PSNR compared to state of the art methods, alongside improved SSIM and HU accuracy, within tissue boundaries. Overall, EPC-3D-Diff improves robustness and physics consistency, supporting HU aware synthesis for downstream radiotherapy workflows. The open source code for EPC-3D-Diff is available at https://github.com/ALZAHRAALTALIB/EPC-3D-Diff.

Alzahra Altalib, Chunhui Li, Haytham Ahmad Alewaidat +3

School of Science and Engineering, University of Dundee UK · Faculty of Applied Medical Sciences, Jordan University of Science and Technology · Experia Healthcare, Jordan +1

Physics-Informed Diffusion Models Cone-Beam CT Latent Diffusion Models Physics-Informed ML Medical Image Synthesis Conditional Diffusion Models

Ask PDF

Aug 9, 2026

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

During standard radiotherapy planning, repeated CT acquisitions are often required for patient registration, verification, and adaptive planning, resulting in increased cumulative X-ray dose. To mitigate this, low-dose cone-beam CT (CBCT) is routinely acquired during treatment delivery. However, CBCT image quality remains insufficient for accurate dose calculation and adaptive radiotherapy planning due to increased scatter, noise, beam hardening, and reconstruction related artifacts. This study develops a supervised deep learning based CBCT to CT synthesis framework using a conditional denoising diffusion probabilistic model (DDPM), where the generation of a CT-based planning for accurate positioning and dose calculation is obtained using generative models with low dose CBCT imaging. Beyond demonstrating CBCT to CT synthesis, the primary objective is to investigate how the representation of CBCT input data, either standard clinical DICOM CBCT images or filtered back-projection (FDK) reconstructions from raw projection data, affects the performance of diffusion based CT synthesis. The overarching aim is to assess whether physics aware CBCT representations better support CT-equivalent image quality while maintaining reduced imaging dose in radiotherapy workflows.

Alzahra Altalib, Chunhui Li, Christopher Hamill Taylor +2

Faculty of Applied Medical Sciences, Jordan University of Science and Technology, Irbid, Jordan · School of Science and Engineering, University of Dundee, Scotland, UK · NHS Ninewells Dundee, Scotland, UK +1

Cone-Beam CT Medical Image Synthesis Conditional Diffusion Models

Ask PDF

Oct 5, 2026

Anatomy-preserving unpaired cone-beam CT refinement for image-guided radiotherapy using pseudo-label guided diffusion

Cone-beam computed tomography (CBCT) is widely used in image-guided radiotherapy, but scatter, beam hardening, noise, truncation, and other artifacts limit image quality and CT number accuracy. Paired CBCT and CT data are difficult to obtain clinically because of motion, anatomical changes, and acquisition mismatch. We present RefineCBCT, an unpaired CBCT refinement framework that uses pseudo-label guidance and short-step diffusion to reduce artifacts while preserving patient-specific anatomy. RefineCBCT was trained and evaluated on unpaired CBCT and planning CT data from public LUNG TCIA and PELVIC TCIA datasets and compared with representative GAN and diffusion based methods. On LUNG TCIA, it achieved the best results across all metrics, with MAE 19.411, RMSE 62.758, PSNR 30.845 dB, and SSIM 0.931. On PELVIC TCIA, it achieved the best MAE, PSNR, and SSIM, with values of 14.905, 36.671 dB, and 0.876. The refined images showed fewer streaking and shading artifacts, clearer anatomical boundaries, and improved soft tissue uniformity, with line profile and ROI analyses showing closer agreement with planning CT. These results suggest that RefineCBCT provides efficient and effective CBCT refinement under clinically realistic unpaired training conditions and may support more reliable CBCT use in image-guided radiotherapy workflows. Code is publicly available on GitHub, and the evaluated datasets are available from The Cancer Imaging Archive.

Qi Lai, Yutong He

Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, 518055,China. · Southern university of science and technology, Shenzhen, 518055,China.

Cone-Beam CT Diffusion Guidance Unpaired Image-to-Image Translation

Ask PDF

Abstract

Medical Image Synthesis

Explore similar work

EPC-3D-Diff: Equivariant Physics Consistent Conditional 3D Latent Diffusion for CBCT to CT Synthesis

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

Anatomy-preserving unpaired cone-beam CT refinement for image-guided radiotherapy using pseudo-label guided diffusion