Extremely Sparse-View Computed Tomography from 2D Projections via Pose-Aware Diffusion Priors

Linrui Dai; Jinxiu Liang; Chu Zhou; Mae Yamaguchi; Ryuya Takahashi; Imari Sato

cs.CVOct 28, 2026

Extremely Sparse-View Computed Tomography from 2D Projections via Pose-Aware Diffusion Priors

Authors: Linrui Dai, Jinxiu Liang, Chu Zhou, Mae Yamaguchi, Ryuya Takahashi, Imari Sato

Organizations: The University of Tokyo · National Institute of Informatics

Abstract

Extremely sparse-view computed tomography (xSV-CT) is critical when scanning time, motion, or radiation tolerance limits each scan to only a few projections. In many such scenarios, the same acquisition limits also constrain dataset construction, making pre-reconstructed 3D volumes from dense scans unavailable. This exposes a gap in current sparse-view approaches: Optimization-based methods, which use the sparse measurements directly but lack data-driven priors to suppress artifacts, and learning-based methods, which use strong priors but usually require densely scanned training volumes acquired outside the xSV-CT regime. We study this train-sparse / infer-sparse setting by learning a pose-aware diffusion prior from one 2D projection per training sample, then guiding diffusion sampling with measured-view fidelity, shared-volume reprojection, and projection-physics consistency for tomographic reconstruction. On two public CT datasets across 1-, 4-, and 8-view settings, the proposed method achieves the best novel-view synthesis scores and best or competitive reconstruction quality compared with baselines that use dense-volume supervision.

Figures & tables

Figure 1: Tomographic reconstruction results from extremely sparse views at $0^{\circ}$ , $45^{\circ}$ , $90^{\circ}$ , $135^{\circ}$ . Current methods cover complementary sides of the extremely sparse view reconstruction problem: a ) optimization-based methods [ 8 ] directly reconstruct from input views but produces severe artifacts when pushed to extreme sparsity at inference. b ) Learning-based methods benefit from data-driven priors, but typically require reconstructed datasets from dense scannings that are difficult to obtain for certain applications. c ) To simultaneously respond to the sparsity limit constraining both the training and inference process, our method learns a pose-aware projection-domain diffusion prior from a single 2D projection per training sample , then uses physics-guided sampling to produce tomographically consistent novel projections for accurate reconstruction.

Figure 2: Overview of our framework. a ) Training phase (Sec. 3.1 ): we learn a pose-aware projection-domain diffusion prior from 2D projections $\mathbf{x}_{0}$ paired with poses $\theta_{i}$ via network $p_{\phi}$ . Specifically, we prepare geometric cues $\theta_{i}$ via sinusoidal pose encoding and inject them with multi-scale pose injection (right). The model is trained with standard denoising loss $\mathcal{L}_{\rm DDPM}$ . b ) Inference phase (Sec. 3.2 ): given $M$ sparse input views $\mathcal{P}$ at poses $\boldsymbol{\theta}$ , we generate $N$ novel views $\hat{\mathbf{x}}_{0}$ at poses $\hat{\boldsymbol{\theta}}$ through consistency-aware guided generation . At each denoising timestep $t$ , a corrector module (right) enforces two complementary constraints: geometric coherency loss $\mathcal{L}_{\rm geo}$ pushes all generated and input views toward a shared 3D explanation via forward projection $\mathcal{A}$ and approximate inverse $\mathcal{A}^{\dagger}$ ; physics-informed consistency loss $\mathcal{L}_{\rm phy}$ further enhances consistency among generated views $\hat{\mathcal{P}}_{t}$ by projecting them onto the consistent range of the Radon transform $\mathcal{A}_{\hat{\boldsymbol{\theta}}}$ .

Figure 3: t-SNE visualizations [ 49 ] of learned features $\mathbf{z}_{l}$ at different layers. Larger dots in the feature space highlight features from the displayed sample projections. $\mathbf{z}_{l}$ at the middle layer (left) aligns features with the general structural distinctiveness of projections to resolve view ambiguity, while $\mathbf{z}_{l}$ at the last layer (right) retains high local sensitivity to individual angles, preserving a separable feature space for accurate pose-conditioned synthesis.

Figure 4: Qualitative comparison of tomographic reconstruction and novel view synthesis from extremely sparse (4) views on CT-RATE and SinoCT datasets. Input angles are marked in red dashed and novel angles in green solid lines in the inset polar plots.

Metric			PSNR $\uparrow$						SSIM $\uparrow$						LPIPS $\downarrow$
Dataset			CT-RATE [ 61 ]			SinoCT [ 62 ]			CT-RATE			SinoCT			CT-RATE			SinoCT
Inference Views			1	4	8	1	4	8	1	4	8	1	4	8	1	4	8	1	4	8
Projection-	based	FBP [ 52 ]	7.2	10.1	11.3	10.6	11.5	12.2	0.15	0.13	0.29	0.28	0.39	0.42	0.50	0.48	0.48	0.52	0.48	0.43
		OS-SART [ 25 ]	7.4	8.9	14.7	12.1	11.9	19.6	0.14	0.16	0.31	0.35	0.44	0.53	0.50	0.46	0.36	0.54	0.43	0.40
		SAX-NeRF [ 10 ]	15.0	19.7	20.5	11.4	12.3	15.1	0.32	0.56	0.61	0.28	0.42	0.49	0.43	0.42	0.39	0.52	0.43	0.40
		R 2 -Gaussian [ 11 ]	14.1	15.4	25.8	11.2	13.6	16.8	0.46	0.52	0.59	0.21	0.31	0.44	0.45	0.45	0.40	0.58	0.51	0.46

Table 1: Quantitative comparison for novel view synthesis on a stratified subset of 50 test cases per dataset. Metrics are averaged across 180 synthesized novel projections covering the entire pose space. We highlight best , second best and third best results.

Metric			PSNR $\uparrow$						SSIM $\uparrow$						LPIPS $\downarrow$
Dataset			CT-RATE [ 61 ]			SinoCT [ 62 ]			CT-RATE			SinoCT			CT-RATE			SinoCT
Inference Views			1	4	8	1	4	8	1	4	8	1	4	8	1	4	8	1	4	8
Projection-	based	FBP [ 52 ]	7.4	8.0	9.6	11.7	11.7	13.0	0.12	0.19	0.23	0.34	0.38	0.39	0.55	0.53	0.54	0.54	0.45	0.43
		OS-SART [ 25 ]	7.9	7.8	11.2	12.7	17.8	19.0	0.15	0.19	0.27	0.33	0.37	0.44	0.50	0.48	0.56	0.53	0.43	0.41
		SAX-NeRF [ 10 ]	6.3	11.4	20.2	9.4	15.0	17.9	0.15	0.37	0.54	0.29	0.40	0.46	0.66	0.52	0.40	0.49	0.42	0.36
		R 2 -Gaussian [ 11 ]	9.8	15.9	24.8	9.3	12.0	19.1	0.21	0.40	0.56	0.23	0.28	0.42	0.55	0.45	0.29	0.56	0.47	0.41

Table 2: Quantitative comparison for tomographic reconstruction on the same 50-case stratified subset as Table 1 .

Figure 5: Performance scaling with the number of input views, metrics averaged over 30 cases.

Task	Novel View Synthesis						Tomographic Reconstruction
Dataset	CT-RATE			SinoCT			CT-RATE			SinoCT
Metric	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
No pose	11.9	0.21	0.56	12.5	0.23	0.50	9.4	0.17	0.61	10.7	0.25	0.53
Naive CrossAttn	24.5	0.40	0.48	20.2	0.45	0.47	11.5	0.33	0.41	15.6	0.39	0.50
w.o. $\mathcal{L}_{\rm phy}$ , $\mathcal{L}_{\rm geo}$	15.6	0.32	0.55	14.6	0.32	0.49	9.2	0.19	0.59	10.0	0.18	0.53
w.o. $\mathcal{L}_{\rm geo}$	28.6	0.78	0.30	21.4	0.50	0.40	16.4	0.35	0.44	18.9	0.51	0.47

Table 3: Ablation study on key components evaluated with 8 input views on 50 test cases per dataset.

Figure 6: Parameter sensitivity analysis with 8 input views. Blue and red curves show PSNR on CT-RATE and SinoCT datasets respectively. Gray dashed lines indicate our default settings.

Views	mean $\times 10^{-3}$	std. $\times 10^{-4}$
1-view	5.48	4.64
4-view	1.31	1.74
8-view	0.96	0.91

Table 4: Cross-seed

\ell_{2}

residual on generated views.

Appendix figures & tables35 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Projection Geometry				No. Training	No. Validation
Dataset	Detector Shape	Volume Shape	Detector Pixel Size	Voxel Size	Cases	Cases
CT-RATE [ 61 ]	$256\times 256$	$256\times 256\times 256$	$2\times 2$	$2\times 2\times 2$	4200	600
SinoCT [ 62 ]	$64\times 512$	$64\times 512\times 512$	$4\times 1$	$4\times 1\times 1$	2800	400

Appendix

Table 5: Projection geometry and dataset splits (training, validation and testing) for CT-RATE and SinoCT. Our method uses only a single 2D projection per training sample, while learning-based baselines require the full 3D volumes reconstructed from dense angular sampling.

Method	Supervision	Retraining
DiffusionMBIR [ 37 ]	Reconstructed axial slices	Per dataset (view-count agnostic)
C 2 RV [ 13 ]	Reconstructed 3D volumes	Per dataset and per view count
DSDF [ 14 ]	Reconstructed 3D volumes	Per dataset and per view count (both stages)
Ours	X-ray projections	Per dataset (view-count agnostic)

Appendix

Table 6: Training paradigms of learning-based baselines and ours.

Figure 7: Tomographic reconstruction across different xSV-CT settings comparing our method against two best-performing methods [ 14 , 37 ] . One axial and one coronal slice is shown for each case. Our method produces more plausible reconstructions with fewer artifacts across all sparsity levels.

Figure 8: Consistency analysis via $\ell_{2}$ residuals under the 3 $\times$ SV-CT setting across 4 random seeds. Top: residuals of novel view synthesis at different angles, showing reduced error near input angles and consistent behavior across seeds. Bottom: average residuals of tomographic reconstruction overlayed on ground-truth axial slices, showing that reconstruction error decreases as more views become available. The low cross-seed variance suggests that the guided generation produces similar outputs across different initializations.

Figure 9: Performance under fixed radiation budgets: each vertically aligned group of cells corresponds to a constant radiation budget (inference views $M$ $\times$ per-view dose $D$ ). Red outlines indicate the best-performing combination of $M$ and $D$ under each budget. Numbers denote PSNR values for either Novel View Synthesis (left) or Tomographic Reconstruction (right) on CT-RATE. The optimal operating point consistently involves moderate sparsity paired with moderate per-view dose, suggesting that our framework could provide flexibility in balancing view count and per-view exposure under radiation constraints.

#Train Views/Sample	#Test Views
#Train Views/Sample	1	4	8
1 (Ours)	23.1 / 21.1	32.8 / 24.5	36.3 / 28.6
4	23.0 / 21.1	33.2 / 24.8	36.9 / 28.6
32	26.0 / 21.3	33.6 / 24.8	38.1 / 29.3

Appendix

Table 7: Effect of training-time data availability: PSNR for novel view synthesis / tomographic reconstruction on CT-RATE dataset with more views per sample seen at training. With a single projection per training sample, our method achieves competitive results. Additional training views provide gains, particularly for single-view inference, while multi-view inference is less sensitive to the number of views per training sample.

Figure 10: Qualitative evaluation under noisy input projections. Each row shows reconstruction from 8 noisy projections (one example shown in the first row) under simulated 75%, 50%, 25% and 10% of original X-ray dosage to guide the inference process.

Equiv.	Novel View Syn.			Reconstruction
Dosage	P $\uparrow$	S $\uparrow$	L $\downarrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$
100%	36.3	0.95	0.17	28.6	0.69	0.27
75%	34.3	0.90	0.20	26.1	0.66	0.28
50%	33.6	0.89	0.23	25.3	0.66	0.30
25%	30.6	0.85	0.24	25.0	0.65	0.31
10%	31.2	0.85	0.24	24.3	0.61	0.30

Appendix

Table 8: Robustness to noisy input projections (left) and angular uncertainty (right). Metrics ( P SNR/ S SIM/ L PIPS) are averaged on 50 CT-RATE cases with 8 input views.

Figure 11: Systematic analysis of failure modes across four conditions, comparing our method with the best-performing optimization-based (3DGR-CT [ 12 ] ) and learning-based methods (DSDF [ 14 ] ) using 8 input views (0 ∘ , 22.5 ∘ , 45 ∘ , 67.5 ∘ , 90 ∘ , 112.5 ∘ , 135 ∘ , 157.5 ∘ ). The model of ours and the learning-based method are trained on CT-RATE dataset. Sample input projections as well as input angles are shown on the left. We provide tomographic reconstruction results under 4 different settings: a ) Standard xSV-CT setting with 8 input views: our method preserves fine-grained anatomical details. b ) xSV-CT with a simulated 5% dose of the 8 input views: our diffusion prior maintains structural integrity while baselines introduce severe artifacts. c ) xSV-CT with the 8 input view angles randomly perturbed within $\pm 10^{\circ}$ : our method stays anchored to the available projections, while baselines develop geometric artifacts (3DGR-CT) or hallucinate structure (DSDF). d ) xSV-CT on an out-of-distribution (OOD) target with respect to the training datasets: our method exhibits graceful blurring along the superior-inferior axis, yet still produces more faithful reconstructions than both baselines (note that for optimization-based methods there are no OOD targets in this sense).

Variant	PSNR $\uparrow$
FBP, $M{=}360$ , corrupted projections	20.7
FBP, $M{=}360$ , corrupted projections, non-metal region	21.2
Ours, $M{=}8$ , corrupted projections	26.4 ( $-2.2$ )
Ours, $M{=}8$ , corrupted projections, non-metal region	27.7 ( $-0.9$ )

Appendix

Table 9: Reconstruction under simulated metal-induced beam hardening. PSNR is computed against the clean, metal-free volume; values in parentheses give the change relative to reconstructing from clean projections with our methods with

M=8

Case No.		Case 1			Case 2
No. Inference Views		1-view	4-view	8-view	1-view	4-view	8-view
CT-RATE	NVS	Fig. 12	Fig. 13	Fig. 14	Fig. 15	Fig. 16	Fig. 17
CT-RATE	TR	Fig. 18	Fig. 19	Fig. 20	Fig. 21	Fig. 22	Fig. 23
SinoCT	NVS	Fig. 24	Fig. 25	Fig. 26	Fig. 27	Fig. 28	Fig. 29
SinoCT	TR	Fig. 30	Fig. 31	Fig. 32	Fig. 33	Fig. 34	Fig. 35

Appendix

Table 10: Index of qualitative results. Each entry links to the corresponding figure showing novel view synthesis (NVS) or tomographic reconstruction (TR) for a specific case and sparsity level.

Figure 12: Novel View Synthesis results (45 ∘ , 90 ∘ ) from 1 view (0 ∘ ) on Case 1 from the CT-RATE dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph. We provide zoomed-in details for results with plausible anatomy compared to their corresponding ground truths.

Figure 13: Novel View Synthesis results (22.5 ∘ , 112.5 ∘ ) from 4 views (0 ∘ , 45 ∘ , 90 ∘ , 135 ∘ ) on Case 1 from the CT-RATE dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph. We provide zoomed-in details for results with plausible anatomy compared to their corresponding ground truths.

Figure 14: Novel View Synthesis results (11.25 ∘ , 101.25 ∘ ) from 8 views (0 ∘ , 22.5 ∘ , 45 ∘ , 67.5 ∘ , 90 ∘ , 112.5 ∘ , 135 ∘ , 157.5 ∘ ) on Case 1 from the CT-RATE dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph. We provide zoomed-in details for results with plausible anatomy compared to their corresponding ground truths.

Figure 15: Novel View Synthesis results (45 ∘ , 90 ∘ ) from 1 view (0 ∘ ) on Case 2 from the CT-RATE dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph.

Figure 16: Novel View Synthesis results (22.5 ∘ , 112.5 ∘ ) from 4 views (0 ∘ , 45 ∘ , 90 ∘ , 135 ∘ ) on Case 2 from the CT-RATE dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph. We provide zoomed-in details for results with plausible anatomy compared to their corresponding ground truths.

Figure 17: Novel View Synthesis results (11.25 ∘ , 101.25 ∘ ) from 8 views (0 ∘ , 22.5 ∘ , 45 ∘ , 67.5 ∘ , 90 ∘ , 112.5 ∘ , 135 ∘ , 157.5 ∘ ) on Case 2 from the CT-RATE dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 18: Tomographic Reconstruction results from 1 view (0 ∘ ) on Case 1 from the CT-RATE dataset, the input angle is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 19: Tomographic Reconstruction results from 4 views (0 ∘ , 45 ∘ , 90 ∘ , 135 ∘ ) on Case 1 from the CT-RATE dataset, the input angles is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 20: Tomographic Reconstruction results from 8 views (0 ∘ , 22.5 ∘ , 45 ∘ , 67.5 ∘ , 90 ∘ , 112.5 ∘ , 135 ∘ , 157.5 ∘ ) on Case 1 from the CT-RATE dataset, the input angles is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 21: Tomographic Reconstruction results from 1 view (0 ∘ ) on Case 2 from the CT-RATE dataset, the input angle is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 22: Tomographic Reconstruction results from 4 views (0 ∘ , 45 ∘ , 90 ∘ , 135 ∘ ) on Case 2 from the CT-RATE dataset, the input angles is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 23: Tomographic Reconstruction results from 8 views (0 ∘ , 22.5 ∘ , 45 ∘ , 67.5 ∘ , 90 ∘ , 112.5 ∘ , 135 ∘ , 157.5 ∘ ) on Case 2 from the CT-RATE dataset, the input angles is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 24: Novel View Synthesis results (45 ∘ , 90 ∘ ) from 1 view (0 ∘ ) on Case 1 from the SinoCT dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph. We provide zoomed-in details for results with plausible anatomy compared to their corresponding ground truths.

Figure 25: Novel View Synthesis results (22.5 ∘ , 112.5 ∘ ) from 4 views (0 ∘ , 45 ∘ , 90 ∘ , 135 ∘ ) on Case 1 from the SinoCT dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph. We provide zoomed-in details for results with plausible anatomy compared to their corresponding ground truths.

Figure 26: Novel View Synthesis results (11.25 ∘ , 101.25 ∘ ) from 8 views (0 ∘ , 22.5 ∘ , 45 ∘ , 67.5 ∘ , 90 ∘ , 112.5 ∘ , 135 ∘ , 157.5 ∘ ) on Case 1 from the CT-RATE dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph. We provide zoomed-in details for results with plausible anatomy compared to their corresponding ground truths.

Figure 27: Novel View Synthesis results (45 ∘ , 90 ∘ ) from 1 view (0 ∘ ) on Case 2 from the SinoCT dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph. We provide zoomed-in details for results with plausible anatomy compared to their corresponding ground truths.

Figure 28: Novel View Synthesis results (22.5 ∘ , 112.5 ∘ ) from 4 views (0 ∘ , 45 ∘ , 90 ∘ , 135 ∘ ) on Case 2 from the SinoCT dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph. We provide zoomed-in details for results with plausible anatomy compared to their corresponding ground truths.

Figure 29: Novel View Synthesis results (11.25 ∘ , 101.25 ∘ ) from 8 views (0 ∘ , 22.5 ∘ , 45 ∘ , 67.5 ∘ , 90 ∘ , 112.5 ∘ , 135 ∘ , 157.5 ∘ ) on Case 2 from the SinoCT dataset, the input (red) and novel (green) angles are also indicated in the inset polar graph. We provide zoomed-in details for results with plausible anatomy compared to their corresponding ground truths.

Figure 30: Tomographic Reconstruction results from 1 view (0 ∘ ) on Case 1 from the SinoCT dataset, the input angle is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 31: Tomographic Reconstruction results from 4 views (0 ∘ , 45 ∘ , 90 ∘ , 135 ∘ ) on Case 1 from the SinoCT dataset, the input angles is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 32: Tomographic Reconstruction results from 8 views (0 ∘ , 22.5 ∘ , 45 ∘ , 67.5 ∘ , 90 ∘ , 112.5 ∘ , 135 ∘ , 157.5 ∘ ) on Case 1 from the SinoCT dataset, the input angles is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 33: Tomographic Reconstruction results from 1 view (0 ∘ ) on Case 2 from the SinoCT dataset, the input angle is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 34: Tomographic Reconstruction results from 4 views (0 ∘ , 45 ∘ , 90 ∘ , 135 ∘ ) on Case 2 from the SinoCT dataset, the input angles is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Figure 35: Tomographic Reconstruction results from 8 views (0 ∘ , 22.5 ∘ , 45 ∘ , 67.5 ∘ , 90 ∘ , 112.5 ∘ , 135 ∘ , 157.5 ∘ ) on Case 2 from the SinoCT dataset, the input angles is also indicated in the inset polar graph. We zoom in on methods with plausible details with the ground truths.

Sparse-View CT Reconstruction

Sparse-view computed tomography (CT) reconstruction recovers 3D images from limited X-ray projections while suppressing artifacts and preserving anatomy.

See all 52 papers

Conditional Diffusion Posterior Alignment for Sparse-View CT Reconstruction
Apr 23, 2026Luis Barba, Johannes Kirschner, Benjamin BejarCone-Beam CT 3D Medical Imaging
Swiss Data Science Center (SDSC) in Paul Scherrer Institute (PSI), Villigen, Switzerland. · Swiss Data Science Center (SDSC) and ETH Zurich, Switzerland. · Swiss Data Science Center (SDSC) and Paul Scherrer Institute (PSI), Villigen, Switzerland.
Progressively Texture-Aware Diffusion for Contrast-Enhanced Sparse-View CT
Apr 13, 2026Tianqi Wang, Wenchao Du, Hongyu YangSparse-View CT Reconstruction Diffusion-Based Inverse Problems
The College of Computer Science, Sichuan University, Chengdu 610065, China
Geometry-Aware Diffusion Approximate Posterior Sampling for Sparse-View and Limited-Angle CT
Oct 5, 2026Honglei Brinkmann, Carola-Bibiane Schönlieb, Ander BiguriSparse-View CT Reconstruction Computed Tomography
Department of Applied Mathematics and Theoretical Physics, University of Cambridge, Cambridge, UK.

Apr 23, 2026J/K move · Enter open · S save

Conditional Diffusion Posterior Alignment for Sparse-View CT Reconstruction

Luis Barba, Johannes Kirschner, Benjamin Bejar

Swiss Data Science Center (SDSC) in Paul Scherrer Institute (PSI), Villigen, Switzerland. · Swiss Data Science Center (SDSC) and ETH Zurich, Switzerland. · Swiss Data Science Center (SDSC) and Paul Scherrer Institute (PSI), Villigen, Switzerland.

Cone-Beam CT 3D Medical Imaging Sparse-View CT Reconstruction Conditional Diffusion Models Diffusion Posterior Sampling

Open paper AskPDF arXiv

Computed Tomography (CT) is a widely used imaging modality in medical and industrial applications. To limit radiation exposure and measurement time, there is a growing interest in sparse-view CT, where the number of projection views is significantly reduced. Deep neural networks have shown great promise in improving reconstruction quality in sparse-view CT, especially generative diffusion models. However, these methods struggle to scale to large 3D volumes due to several reasons: (i) the high memory and computational requirements of 3D models, (ii) the lack of large 3D training datasets, and (iii) the inconsistencies across slices when using 2D models independently on each slice. We overcome these limitations and scale diffusion-based sparse-view CT reconstruction to large 3D volumes by combining conditional diffusion with explicit data consistency. We propose Conditional Diffusion Posterior Alignment (CDPA) to enable scalable 3D sparse-view CT reconstruction. A 2D U-Net diffusion model is conditioned on an initial 3D reconstruction to improve inter-slice consistency, combined with data-consistency alignment to match measured projections. Experiments on synthetic and real Cone Beam CT (CBCT) data show state-of-the-art performance, with ablations that confirm the synergistic effects of the proposed pipeline. Finally, we show that the same principles also strengthen fast denoising U-Nets, yielding near-diffusion quality at a fraction of the computational cost.

Abstract

Sparse-View CT Reconstruction

Explore similar work

Conditional Diffusion Posterior Alignment for Sparse-View CT Reconstruction

Progressively Texture-Aware Diffusion for Contrast-Enhanced Sparse-View CT

Geometry-Aware Diffusion Approximate Posterior Sampling for Sparse-View and Limited-Angle CT