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

Qi Lai; Yutong He

cs.CVOct 5, 2026

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

Authors: Qi Lai, Yutong He

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

Abstract

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.

Figures & tables

Figure 1 : Clinical motivation and conceptual overview of RefineCBCT. Artifact-degraded CBCT and the scarcity of accurately paired training data motivate an unpaired refinement strategy. RefineCBCT couples a cycle-consistent pseudo-label translation module with a short-step diffusion refinement module. The translated reference provides anatomy-aware guidance, while short-step diffusion suppresses artifacts and preserves patient-specific structures with reduced sampling cost.

Dataset	#Patients	#Slices	Slice size
LUNG TCIA O et al. (2012)	20	6,721	512 × 512
PELVIC TCIA Clark et al. (2013)	58	3,850	512 × 512

Table 1 : properties of the public cone-beam CT (CBCT) datasets.

Dataset	Methods	MAE ↓	RMSE ↓	PSNR ↑	SSIM ↑
LUNG TCIA	CycleGAN + Zhu et al. (2017)	36.503	97.512	26.499	0.848
	CUT + Dong et al. (2022)	37.684	101.165	26.048	0.849
	DCLGAN + Wang et al. (2023)	32.503	88.149	27.234	0.876
	SynDiff ∗ Özbey et al. (2023)	34.791	94.493	26.713	0.858
	FGDM ∗ Li et al. (2023a)	31.164	88.817	27.323	0.879
	DiffusionMBIR ∗ Chung et al. (2023)	26.874	89.037	28.819	0.897

Table 2: Quantitative comparison of the image quality produced using various models.

Panel of Figure 2 — Figure 2 : Experimental analysis across various methods on LUNG TCIA and PELVIC TCIA datasets. For each dataset, the left panel shows normalized quality scores for MAE, RMSE, PSNR, and SSIM, and the right panel presents the corresponding inference time on a logarithmic scale.

Panel of Figure 3 — Figure 3 : Qualitative comparison of RefineCBCT against compared methods. For each subfigure, the middle row shows representative axial slices for all methods, the top row provides magnified views of the green boxed regions for fine-structure inspection, and the bottom row shows difference maps with respect to pCT. The pCT is used as the reference for visual comparison. Display window: [-1000, 1000] HU.

Panel of Figure 4 — Figure 4 : Radar-plot comparison of RefineCBCT and baseline methods on two public datasets. Larger values indicate better performance.

Panel of Figure 5 — Figure 5 : Line-profile analysis of HU consistency for RefineCBCT on two public datasets. For each subfigure, the left panel shows matched axial slices of CBCT, sCBCT, and pCT, and the right panel plots HU profiles sampled along the red and yellow lines. (a) On LUNG TCIA, the red line (Line 100) mainly traverses soft tissue, whereas the yellow line (Line 95) crosses both bone and soft tissue. (b) On PELVIC TCIA, the red line (Line 115) mainly traverses soft tissue, whereas the yellow line (Line 120) crosses both bone and soft tissue. In both datasets, sCBCT more closely follows pCT and shows reduced fluctuations relative to CBCT. Display window: [-1000, 1000] HU.

Panel of Figure 6 — Figure 6 : ROI placement and regional HU distribution analysis on two public datasets. (a) and (b) show representative bone and soft-tissue ROIs placed at corresponding anatomical locations in CBCT, sCBCT, and pCT. Green boxes denote bone ROIs and red boxes denote soft-tissue ROIs. From top to bottom, the rows show ROI (Bone), axial slice, ROI (Tissue), and a high-contrast bone window. The display windows for ROI (Bone), axial slice, and ROI (Tissue) are [-1000, 1000] HU, and that for the bottom bone row is [400, 750] HU. (c) and (d) show violin plots of HU distributions within the selected bone and soft-tissue ROIs for LUNG TCIA and PELVIC TCIA , respectively. Compared with CBCT, sCBCT exhibits distributions that are narrower and closer to those of pCT, indicating reduced bias and variance in both bone and soft-tissue regions.

Method variant	Inference time (s) $\downarrow$	PSNR (dB) $\uparrow$	SSIM $\uparrow$
Diffusion backbone	0.301	21.152	0.451
Diffusion backbone + Generation	0.312	26.394	0.801
Full model	0.342	30.473	0.915

Table 3: Ablation study of each component on LUNG TCIA dataset.

Diffusive	Generation	C.C.L	Inference time (s) $\downarrow$	PSNR (dB) $\uparrow$	SSIM $\uparrow$
$\checkmark$			0.297	25.648	0.539
$\checkmark$	$\checkmark$		0.306	31.259	0.775
$\checkmark$	$\checkmark$	$\checkmark$	0.322	35.941	0.848

Table 4: Ablation study of each component on PELVIC TCIA dataset.

Dataset	Methods	mAP $\uparrow$	$\Delta$ mAP	IoU $\uparrow$	$\Delta$ IoU
LUNG TCIA	Original CBCT	0.529	–	0.189	–
	CycleGAN + Zhu et al. (2017)	0.617	0.088	0.193	0.004
	CUT + Dong et al. (2022)	0.607	0.078	0.279	0.090
	DCLGAN + Wang et al. (2023)	0.648	0.119	0.518	0.329
	SynDiff ∗ Özbey et al. (2023)	0.670	0.141	0.601	0.412
	FGDM ∗ Li et al. (2023a)	0.660	0.131	0.505	0.316

Table 5: Quantitative comparison of downstream object detection performance produced using various models.

Panel of Figure 7 — Figure 7 : Qualitative object detection results on the LUNG and PELVIC TCIA datasets. Red boxes indicate predictions and green boxes indicate the reference targets.

Figure 8 : Representative volumetric results of RefineCBCT on the LUNG TCIA and PELVIC TCIA datasets. Two patient cases are shown for each dataset. For each case, axial, sagittal, and coronal views are presented for the original CBCT, the RefineCBCT-corrected image (sCBCT), and the corresponding planning CT (pCT). The sagittal and coronal views were reconstructed from the independently processed axial slices to qualitatively assess through-plane anatomical coherence.

Panel of Figure 9 — Figure 9 : Hyper-parameter sensitivity analysis of RefineCBCT on two public datasets. PSNR is reported under different combinations of $\lambda_{\mathrm{cyc}}$ and $\lambda_{\mathrm{rec}}$ .

Figure 10 : Technical details of RefineCBCT.

Explore similar work

EPC-3D-Diff: Equivariant Physics Consistent Conditional 3D Latent Diffusion for CBCT to CT Synthesis
May 19, 2026Alzahra Altalib, Chunhui Li, Haytham Ahmad Alewaidat +3Physics-Informed Diffusion Models Cone-Beam CT
School of Science and Engineering, University of Dundee UK · Faculty of Applied Medical Sciences, Jordan University of Science and Technology · Experia Healthcare, Jordan +1
Toward CT-Equivalent Image Quality in Low-Dose Radiotherapy Planning: Conditional Diffusion-Based CBCT-to-CT Synthesis and the Impact of CBCT Input Representation
Aug 9, 2026Alzahra Altalib, Chunhui Li, Christopher Hamill Taylor +2Cone-Beam CT Medical Image Synthesis
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
Field-of-View Extension in Dental Cone-Beam CT via Implicit Neural Representations and Diffusion Model-Based Refinement
Sep 23, 2026Susanne Schaub, Florentin Bieder, Matheus L. Oliveira +5Cone-Beam CT Image Reconstruction
Department of Biomedical Engineering, University of Basel, Allschwil, Switzerland · Piracicaba Dental School, University of Campinas, Piracicaba, Brazil · State Key Laboratory of Oral & Maxillofacial Reconstruction and Regeneration, Key Laboratory of Oral Biomedicine Ministry of Education, Hubei Key Laboratory of Stomatology, School & Hospital of Stomatology, Wuhan University, Wuhan, China +2

May 19, 2026, cs.CVJ/K move · Enter open · S save

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

Alzahra Altalib, Chunhui Li, Haytham Ahmad Alewaidat, Khaled Z. Alawneh+2

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

Open paper AskPDF arXiv

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.

Abstract

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

Field-of-View Extension in Dental Cone-Beam CT via Implicit Neural Representations and Diffusion Model-Based Refinement