Abstract

Metallic implants in X-ray Computed Tomography (CT) scans can lead to undesirable artifacts, adversely affecting the quality of images and, consequently, the effectiveness of clinical treatment. Metal Artifact Reduction (MAR) is essential for improving diagnostic accuracy, yet this task is challenging due to the uncertainty associated with the affected regions. In this paper, inspired by the capabilities of diffusion models in generating high-quality images, we present a novel MAR framework termed Dual-Domain Conditional Diffusion (DCDiff). Specifically, our DCDiff takes dual-domain information as the input conditions for generating clean images: 1) the image domain incorporating raw CT image and the filtered back project (FBP) output of the metal trace, and 2) the sinogram domain achieved with a new diffusion interpolation algorithm. Experimental results demonstrate that our DCDiff outperforms state-of-the-art methods, showcasing its effectiveness for MAR.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1608_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: N/A

Link to the Code Repository

N/A

Link to the Dataset(s)

https://nihcc.app.box.com/v/DeepLesion/

BibTex

@InProceedings{She_DCDiff_MICCAI2024,
        author = { Shen, Ruochong and Li, Xiaoxu and Li, Yuan-Fang and Sui, Chao and Peng, Yu and Ke, Qiuhong},
        title = { { DCDiff: Dual-Domain Conditional Diffusion for CT Metal Artifact Reduction } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15007},
        month = {October},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces DCDiff, a novel Metal Artifact Reduction (MAR) framework based on diffusion models, designed to improve the quality of CT scans by reducing artifacts caused by metallic implants. It combines dual-domain information from both image and sinogram domains to generate cleaner CT images.

  • Please list the main strengths of the paper; you should write about 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.

    The paper introduces a diffusion interpolation algorithm for sinogram generation, which helps in accurately reconstructing clean CT images. DCDiff incorporates both image and sinogram domains, the approach provides a more comprehensive perspective for artifact reduction, which enhances performance.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    1. The paper does not provide direct access to code or data, which may hinder reproducibility.
    2. The author claims that the proposed method is pioneering diffusion-based framework for MAR, marking the first exploration of leveraging supervised conditional diffusion models tailored for MAR tasks. However, we can search the relative paper from the google scholar: Choi Y, Kwon D, Baek S J. Dual Domain Diffusion Guidance for 3D CBCT Metal Artifact Reduction[C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024: 7965-7974. 3.The formulas in the article should be listed instead of being placed in paragraphs to enhance readability. 4.The compared visualization in Fig.2 is not the SOTA methods in the table, such as DICDNet. Moreover, more new MAR networks should be included in the table. 5.The result lacks of clinical dataset validation. The time consumption should be discussed.
  • 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 submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    1. Consider providing the source code to enable reproducibility and further research in the field. This would also help validate the results independently.
    2. Improve the expression of the article.
    3. Add new comparative methods and clinical data validation.
  • 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

    Reject — should be rejected, independent of rebuttal (2)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents a novel approach to MAR using diffusion models, showcasing significant performance improvements over existing methods. However, the limited reproducibility, article expression and insufficient experimental results prevent a higher rating.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Strong Reject — must be rejected due to major flaws (1)

  • [Post rebuttal] Please justify your decision

    I am confident that OSCNet is not the current state-of-the-art method. Secondly, not only DICIDNet, but also many new MAR methods, the results given in the visualization of the paper are not convincing. The WACV paper we have listed is just an example, and the author’s claim that they are the first dual domain diffusion model is not valid.



Review #2

  • Please describe the contribution of the paper

    This paper proposed a method that performs CT metal artifact reduction in both sinogram and image domain. Compared to comparison methods the author listed, the best results were achieved.

  • Please list the main strengths of the paper; you should write about 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. The author proposed a diffusion-based method that leveraging dual-domain optimization, exploring diffusion models equipped with various conditional images.
    2. Good comparison experiments and clear description about dataset.
  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    1. Regarding the Diffusion Interpolation (DI), in the ablation study section, the author states that they compare the diffusion model, with or without DI, in sinogram domain, and it shows the effectiveness of DI: PSNR/SSIM of 31.59 dB/0.9124 without DI, and 35.47 dB/0.9390 within DI. However, the effectiveness of DI in the dual domain diffusion method, which is proposed by author, is not clear.
  • Please rate the clarity and organization of this paper

    Very 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 submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    Please provide an analysis of the effectiveness of Diffusion Interpolation (DI) in the dual domain diffusion model.

  • 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

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
    1. Using dual domain diffusion models for metal artifacts reduction is a novel idea.
    2. Comparative experiments and ablation experiments are relatively detailed.
    3. Clear description on dataset and details for reproducibility.
    4. The flaw is that the effectiveness of Diffusion Interpolation (DI) is only demonstrated on a single diffusion model;
  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Accept — should be accepted, independent of rebuttal (5)

  • [Post rebuttal] Please justify your decision

    The author has resolved my confusion and I think this is a publishable article.



Review #3

  • Please describe the contribution of the paper

    The paper introduces DCDiff, a novel dual-domain conditional diffusion-based framework designed for reducing metal artifacts in CT images. By integrating information from both the image domain (raw CT image, filtered back project of metal trace) and sinogram domain (obtained via a new diffusion interpolation algorithm), DCDiff generates metal-free images. Utilizing the DeepLesion dataset, the framework showcases its capability and surpasses existing techniques in Metal Artifact Reduction (MAR).

  • Please list the main strengths of the paper; you should write about 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) the paper offers a novel approach to metal artifact reduction in CT images which may has potential for significant impact in clinical applications. 2) The experimental results on the DeepLesion dataset show superior performance compared to state-of-the-art methods, demonstrating the model’s effectiveness 3) The ablation study provides insights into the impact of different components within the framework, contributing to a better understanding. 4) Overall, the paper is clearly written and well-structured.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    1) The paper reports good performance on the DeepLesion dataset, however more extensive testing on diverse datasets (e.g. Dental, and Clinical CLINIC-metal) are needed to strengthen the claim of generalizability. 2) The authors don’t explicitly state the total number of parameters in the model. However, given the complexity of the dual-domain framework with the conditional diffusion-based models, it is likely to require a significant number of parameters​.

  • 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 submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Do you have any additional comments regarding the paper’s reproducibility?

    The paper outlines the proposed method with detailed explanations and diagrams, including the architecture of the dual-domain conditional diffusion models and the diffusion interpolation algorithm.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    1) While the paper shows success with the DeepLesion dataset, the question remains: How well does the proposed method generalize to other CT datasets? 2) It would be helpful to include details about the computational cost and the total number of parameters to provide a clearer picture of the model’s efficiency and scalability. 2) Additional visualization examples would help demonstrate the effectiveness of the framework more clearly.

  • 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

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents a novel dual-domain framework using conditional diffusion, however, the experiments in the paper are limited to the DeepLesion dataset, raising questions about the method’s generalizability to other datasets and scenarios.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A




Author Feedback

We sincerely appreciate all reviewers’ comments. We’ll address the weaknesses mentioned.

R1Q1. Effectiveness of DI in the whole framework: Thank you for your suggestion. Our results in the main submission have demonstrated that the performance without DI is significantly worse in the sinogram domain, highlighting the effectiveness of DI. Given the sinogram domain output is a crucial input prior for image generation within the entire framework, it is expected that the poorer quality of the sinogram domain output without DI will similarly degrade the performance across the whole framework. Unfortunately, due to the rebuttal policy, we cannot include these results now. We appreciate your suggestion and will incorporate these results in our future work.

R3Q1. Access to code or data: We apologize for being unable to provide an external link at this time due to the rebuttal policy. We will make the code and data publicly available upon acceptance.

R3Q2. Relative paper [WACV2024]: 1) Our work differs from WACV24 in the following aspects: a) Methods: The two conditional diffusion models in WACV24 are both trained in the image domain (one for metal-free prior image learning and the other for metal artifact prior image learning). In contrast, our approach involves training two diffusion models in both the sinogram and image domains, and we introduce a novel DI model for the sinogram stream and new conditioning methods for image generation; b) Tasks: WACV24 aims to remove metal artifacts from 3D CBCT, whereas our focus is on general 2D CT; c) Datasets: We used the widely recognized public dataset, DeepLesion, while WACV24 primarily utilized a private dataset. 2) This paper was published in January 2024, which coincides with the period when we had completed most of our work. During the preparation of our manuscript, we were unaware of any supervised diffusion dual-domain methods for MAR. We apologize for overlooking this concurrent work. We’ll include a discussion of this paper.

R3Q3. Formulas should be listed: Thank you for your suggestion. We’ll list the formulas separately instead of embedding them within paragraphs in the final version.

R3Q4. Not SOTA methods in Fig.2, such as DICDNet: a) As shown in Table 1, OSCNet is the current SOTA method. We have included the results of OSCNet in Fig.2. b) Fig.4 of the OSCNet paper compares with other methods, such as DICDNet, demonstrating that OSCNet performs better. c) We’ll include DICDNet in the final version, as suggested.

R3Q5&R4Q1&R4Q3. Lack of clinical validation & additional visualization result: 1) We’ve evaluated our model on CLINIC-Metal and a private clinical foot CT dataset. However, due to the page limitation, we decided to follow the papers of DuDoNet [DuDoNet: Dual Domain Network for CT Metal Artifact Reduction, CVPR2019] and MEPNet [MEPNet: A Model-Driven Equivariant Proximal Network for Joint Sparse-View Reconstruction and Metal Artifact Reduction in CT Images, MICCAI2023], both of which evaluated their models exclusively on synthesized DeepLesion data. This allows us to focus on method interpretation and comparison using the most widely recognized dataset, synthesized DeepLesion. 2) Although we can’t include results due to space constraints and rebuttal policy, we will provide detailed evaluations on additional datasets alongside the publicly released code. These results will also be incorporated into our future work.

R3Q5. Time consumption: As mentioned in the implementation details, we have used DDIM to improve efficiency. The time is about 3.1 seconds/image for training and 4.4 seconds/image for testing, around 15 times faster than DDPM.

R4Q2. Parameter number: The parameter number is the sum of parameter numbers in the two denoising networks (UNet in our proposed method), similar to other methods that consist of two UNets as well (eg, DuDoNet, DSCMAR, DuDoNet++), which is around 26M.




Meta-Review

Meta-review #1

  • 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

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    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

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    N/A



back to top