List of Papers Browse by Subject Areas Author List
Abstract
Sparse-view computed tomography (CT) reduces radiation exposure by subsampling projection views, but conventional reconstruction methods produce severe streak artifacts with undersampled data. While deep-learning-based methods enable single-step artifact suppression, they often produce over-smoothed results under significant sparsity. Though diffusion models improve reconstruction via iterative refinement and generative priors, they require hundreds of sampling steps and struggle with stability in highly sparse regimes. To tackle these concerns, we present the Cross-view Generalized Diffusion Model (CvG-Diff), which reformulates sparse-view CT reconstruction as a generalized diffusion process. Unlike existing diffusion approaches that rely on stochastic Gaussian degradation, CvG-Diff explicitly models image-domain artifacts caused by angular subsampling as a deterministic degradation operator, leveraging correlations across sparse-view CT at different sample rates. To address the inherent artifact propagation and inefficiency of sequential sampling in generalized diffusion model, we introduce two innovations: Error-Propagating Composite Training (EPCT), which facilitates identifying error-prone regions and suppresses propagated artifacts, and Semantic-Prioritized Dual-Phase Sampling (SPDPS), an adaptive strategy that prioritizes semantic correctness before detail refinement. Together, these innovations enable CvG-Diff to achieve high-quality reconstructions with minimal iterations, achieving 38.34 dB PSNR and 0.9518 SSIM for 18-view CT using only \textbf{10} steps on AAPM-LDCT dataset. Extensive experiments demonstrate the superiority of CvG-Diff over state-of-the-art sparse-view CT reconstruction methods. The code is available at https://github.com/xmed-lab/CvG-Diff
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0033_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xmed-lab/CvG-Diff
Link to the Dataset(s)
https://aapm.app.box.com/s/eaw4jddb53keg1bptavvvd1sf4x3pe9h
BibTex
@InProceedings{CheJix_Crossview_MICCAI2025,
author = { Chen, Jixiang and Lin, Yiqun and Qin, Yi and Wang, Hualiang and Li, Xiaomeng},
title = { { Cross-view Generalized Diffusion Model for Sparse-view CT Reconstruction } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {139 -- 149}
}
Reviews
Review #1
- Please describe the contribution of the paper
Unlike most existing Gaussian noise-based diffusion models, this work proposes a novel framework that leverages the different sampling rates to improve the sparse-view CT reconstruction.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
This works shows great quantitative improvements over SOTA methods.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
This work demonstrates that excessive smoothing of local structures can still affect reconstruction results.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
-
Authors state that single-step restoration methods usually produce over-smoothed results. However, the results generated from the proposed CvG-Diff still miss many details, which is opposite to what the diffusion model usually provides, visually improved images.
-
Meanwhile, when the Nv=144, there are some artifacts of VSS, FreeSeed, GlobReDi, DuDoTrans, and CvG-Diff in the liver region, but CoSIGN generates more satisfactory results. Authors should explain this phenomenon.
-
In experiment settings, the geometry configurations are limited. Please give more details, such as pixel size, image size, and the source-to-object distance.
-
Authors should compare some unrolling methods.
-
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I’m a little confused by the results of this article
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The author has addressed my most concerns. However, there are some erros. For example, in Fig. 3, it can be written “the display window is [-250, 300] HU” or “the window level is 50 HU and window width is 500 HU”.
Review #2
- Please describe the contribution of the paper
This paper proposes the Cross-view Generalized Diffusion Model (CvG-Diff) for sparse-view CT reconstruction. Built upon the generalized diffusion model (Cold Diffusion), CvG-Diff models angular sparsity as a deterministic degradation process rather than relying on stochastic Gaussian noise. To solve the artifact propagation issues encountered with conventional generalized diffusion model, CvG-Diff introduces error-propagating composite training strategy (EPCT), a two-step training paradigm to learn artifacts propagated from different levels. To avoid over-smoothing, it presents a semantic-prioritized dual-phase sampling (SPDPS) strategy: the anatomical semantic correction phase to enhance anatomical boundaries followed by detail refinement phase. Experiments on the AAPM Low-Dose CT dataset demonstrate that CvG-Diff quantitatively and qualitatively outperforms both existing feed-forward and diffusion-based methods, achieving high-quality reconstruction with fewer sampling steps.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- This paper introduces a novel idea that models angular sparsity as a deterministic degradation process rather than relying on stochastic Gaussian noise. This approach leverages the inherent correlation between CT images acquired at different sparsity levels, enabling a unified model that can handle a range of sparse-view scenarios effectively.
- This paper provides detailed ablation studies that validate the effectiveness of EPCT and SPDPS.
- CvG-Diff achieves high-quality reconstructions using significantly fewer sampling steps (e.g., 6–10 steps) compared to conventional diffusion models.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The model is only evaluated on the abdominal CT dataset. The model’s generalizability to other CT modalities (such as knee, dental, or head CT) is not explored.
- Certain components of the method are not explained in full detail. An example is the severity level mapping g(t) discussed in Section 2.2. Additionally, a more comprehensive explanation of the pipeline would allow readers to fully understand how the proposed framework is implemented.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
How does the method compare to model-based iterative reconstruction approaches such as diffusionMBIR[a] and DPS[b], which also deal with sparse-view CT problem?
[a] Solving 3D Inverse Problems using Pre-trained 2D Diffusion Models, CVPR’23 [b] Decomposed Diffusion Sampler for Accelerating Large-Scale Inverse Problems, ICLR’24
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper presents a generalized diffusion model for sparse-view CT reconstruction. To my knowledge, it is the first work to reformulate sparse-view CT reconstruction as a deterministic degradation process. The enhancements achieved by the EPCT and SPDPS strategies within the generalized diffusion framework yield improved results for sparse-view CT reconstruction. However, the proposed deterministic degradation process and training paradigm should be further evaluated across additional CT modalities to validate CvG-Diff’s generalizability.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have adequately addressed my concerns. I therefore recommend acceptance.
Review #3
- Please describe the contribution of the paper
This paper proposes a cross-view generalized diffusion model for sparse-view CT reconstruction. The model first models artifacts caused by angular subsampling as a deterministic degradation operator. Then, an Error-Propagating Composite Training strategy to suppress the propagation of artifacts and a Semantic-Prioritized Dual-Phase Sampling strategy to prioritize semantic correctness before detail refinement are proposed for sparse-view CT reconstruction. Finally, quantitative and qualitative experiments show that the proposed method outperforms the state-of-the-art methods.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1) To adapt the diffusion model to the specific task of sparse-view CT reconstruction, angular subsampling artifacts are encoded as a deterministic degradation operator, replacing the degradation caused by Gaussian noise simulation in existing diffusion-based methods.
2) The novel Error-Propagating Composite Training and Semantic-Prioritized Dual-Phase Sampling strategy successfully address the error propagation problem in multi-step sampling, while preserving computational efficiency.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1) In the experiment evaluating the impact of parameters $\tau$ and $\m$ on Semantic-Prioritized Dual-Phase Sampling strategy, the reasons for both parameters taking only three values are not explained, and the corresponding results are not analyzed.
2) In the SPDPS module of Fig. 1, $\hat{x}_0^s$ should be corrected to $\hat{x}_0^t$.
3) The capitalization of ‘Error-Propagating Composite Training’ and ‘Semantic-Prioritized Dual-Phase Sampling’ should be consistent throughout the paper.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is well-organized and presents novel method, with its superiority convincingly demonstrated through extensive experimental.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
Thank reviewers for their valuable feedback. Overall, reviewers consider the proposed method novel (R1, R2, R3) and efficient (R2, R3), and appreciate its good performance (R1, R2, R3). The major concerns are experiment results (Q1-Q6), and elaboration (Q7-Q9).
[R1]Q1-Q2: Reconstruction results compared to previous diffusion models: We clarify that the qualitative results of CvG-Diff indeed surpass previous diffusion models with higher fidelity. While CoSIGN reconstructs vivid textures that look reasonable and similar to a high-quality CT, its reconstruction often fails to align with patients’ real condition (GT). In contrast, our method prioritizes high-fidelity restoration over generating visually vivid yet low-fidelity results. For example, tips of the spinous process (zoomed in the 2nd row of Fig. 3) are clearly delineated in ours and GT, but they get shortened in the CoSIGN, despite their vivid textures. Visual results of CvG-Diff show the highest fidelity reconstruction over all compared methods. This may not be obvious in the current Fig.3 due to 1) results of GT and CoSIGN are distant, making it difficult to compare; 2) the displayed image size is small, making differences less obvious. We will include error maps in Fig. 3 and increase the image size to better explain this issue.
[R1]Q3: Geometry configurations: The image size is 256x256, with a mean pixel size of 0.73 mm. The source to object distance is 297 mm, detector receives 672 rays with a spacing of 2 mm. We will clarify them in the paper.
[R1]Q4: Comparison to unrolling methods: We agree that it would be better to also compare unrolling methods. However, due to rebuttal regulations, we cannot provide additional experimental results here. We will consider adding a comparison with some unrolling methods in the final version. Notably, we have compared recent SoTA methods (ECCV’24-CoSIGN, TMI’24-VSS, ICCV’23-Glorei), which are highly relevant and widely recognized, thus comprehensively demonstrating the superiority of our method.
[R2]Q5: Generalization to other anatomy: Due to page limit, we follow Freeseed[14] to experiment on AAPM16. Its CT scans already contain multiple anatomical regions, including thorax (Fig. 3 row 1), abdominal (Fig. 3 row 2), and pelvic regions (Fig. 4). AAPM16 is thus more robust than single-anatomy datasets.
[R2]Q6: Comparison to model-based iterative reconstruction methods: The selected TMI’24-VSS[7] and ECCV’24-CoSIGN [25] are already SoTA model-based iterative reconstruction methods. For [a], they are different experimental settings ([a]: 2D projections to 3D volume, ours: 1D projections to 2D slice). Then, it is unfair to compare with [a]. For [b], we observe that VSS/CoSIGN surpassed [b] (as reported in papers of VSS/CoSIGN). Our improvements over VSS/CoSIGN thus implicitly demonstrate superiority to [b].
[R2]Q7: Elaboration on method details: The severity level mapping g(t) assigns each discrete severity level t to a number of views Tt from a sequence T. The most-sparse level is denoted by Tmax, and the least-sparse level is T1. In Sec. 3.1, we explain the implementation of sequence T: [288, 234, 180, 126, 72, 54, 36, 18], with Tmax=18 and T1=288. We will clarify this and polish the elaboration of the pipeline. The code will be released to help understand the implementation.
[R3]Q8: Clarification on ablation: In Tab. 3, $\tau$ (0.97–0.99) and $m$ (2–4) were tested within practical bounds: smaller $\tau$ could make the reset criteria always satisfied and overly triggered, while larger $m$ sacrifices efficiency. Results demonstrate SPDPS is robust to both parameters, with $\tau$ (degradation reset) impacting performance more significantly than $m$ (refinement steps). A finer parameter search was unnecessary due to its robustness. We will clarify this in the revision.
[R3]Q9: Elaboration on details: Thanks for your careful advice. We will carefully check and revise our paper to ensure consistent and precise expression.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A