Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

In clinical examinations and diagnoses, low-dose computed tomography (LDCT) is crucial for minimizing health risks compared with normal-dose computed tomography (NDCT). However, reducing the radiation dose compromises the signal-to-noise ratio, leading to degraded quality of CT images. To address this, we analyze LDCT denoising task based on experimental results from the frequency perspective, and then introduce a novel self-supervised CT image denoising method called WIA-LD2ND, only using NDCT data. The proposed WIA-LD2ND comprises two modules: Wavelet-based Image Alignment (WIA) and Frequency-Aware Multi-scale Loss (FAM). First, WIA is introduced to align NDCT with LDCT by mainly adding noise to the high-frequency components, which is the main difference between LDCT and NDCT. Second, to better capture high-frequency components and detailed information, Frequency-Aware Multi-scale Loss (FAM) is proposed by effectively utilizing multi-scale feature space. Extensive experiments on two public LDCT denoising datasets demonstrate that our WIA-LD2ND, only uses NDCT, outperforms existing several state-of-the-art weakly-supervised and self-supervised methods.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/0783_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/0783_supp.pdf

Link to the Code Repository

https://github.com/zhaohaoyu376/WI-LD2ND

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Zha_WIALD2ND_MICCAI2024,
        author = { Zhao, Haoyu and Gu, Yuliang and Zhao, Zhou and Du, Bo and Xu, Yongchao and Yu, Rui},
        title = { { WIA-LD2ND: Wavelet-based Image Alignment for Self-supervised Low-Dose CT Denoising } },
        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

    Authors propose WIA-LD2ND, a novel self-supervised learning method for low-dose CT denoising. WIA-LD2ND incorporates a wavelet-based image alignment (WIA) module and a frequency-aware multi-scale loss (FAM) module. The WIA module aligns normal-dose CT images with low-dose CT images by introducing noise to the wavelet coefficients, while the FAM module directs the network’s focus towards high-frequency image details. Evaluation on the Mayo-2016 and Mayo-2020 datasets demonstrates state-of-the-art performance in both quantitative metrics and visual quality.

  • 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 is well-structured and easy to comprehend.
    2. The method is logically motivated, and the proposed modules are well designed.
    3. WIA-LD2ND introduces innovative components like the WIA module and FAM module, tailored to align images and enhance high-frequency details, respectively.
    4. The method showcases superior performance on both the Mayo-2016 and Mayo-2020 datasets.
  • 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. Comparative Analysis: The paper lacks comparative analysis with existing state-of-the-art methods, such as Equivariant Imaging: Learning Beyond the Range Space (ICCV 2021) and GloRe: Learning to Distill Global Representation for Sparse-View CT (ICCV 2023).
    2. As the method is self-supervised, it begs the question why the authors have not tested it on other datasets such as DeepLesion, especially when trained solely on the AAPM Mayo dataset. Doing so would not only demonstrate the robustness of the method but also provide insights into its generalizability.
    3. Dataset Clarification: Clarity is needed regarding whether the testing images in the Mayo-2016 dataset are from distinct patients.
    4. Inference Time: The paper omits information on inference time and memory consumption, crucial for assessing practical utility compared to existing methods.
  • 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?

    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. Baseline Training Details: Additional information on the training process of baseline methods is necessary to ensure fair comparison.
    2. Method Versatility: Discussing the potential applicability of the proposed method to other medical image restoration tasks would enhance the paper’s impact.
    3. Utilization of Low-frequency Coefficients: Exploring why low-frequency coefficients are not utilized in the Online/Target Encoder despite the presence of a frequency-aware Attention module could improve methodological clarity.
    4. Future Recommendations: Considering the potential superiority of MS-SSIM over L_pixel loss for image restoration tasks, as highlighted in “MS-SSIM: Loss Functions for Image Restoration with Neural Networks” (IEEE TCI), could enhance method effectiveness.
    5. Performance Gain Discrepancy: Clarification regarding the performance gain observed on the Mayo-2020 dataset compared to Mayo-2016 would strengthen the paper’s argument.
  • 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 Reject — could be rejected, dependent on rebuttal (3)

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

    While the identified strengths support a favorable assessment, the highlighted weaknesses, including the absence of critical comparisons, dataset clarification issues, and lack of inference time details, introduce uncertainty. Addressing these weaknesses would significantly enhance the paper’s quality and ensure a more robust contribution to the field.

  • 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

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

  • [Post rebuttal] Please justify your decision

    The authors provided clear and detailed responses to my concerns, including explanations on backbone choice, computational efficiency, and dataset suitability. Their commitment to sharing additional results on GitHub shows transparency and dedication to further validation. The responses satisfactorily addressed my questions and justified the design choices and performance improvements claimed in the paper. However, methods such as GloRe (ICCV 2023) could also be effectively applied to this problem, similar to CycleGAN and Noise2Sim. Based on the author responses, I have decided to increase my score to a weak accept.



Review #2

  • Please describe the contribution of the paper

    Authors propose WIA-LD2ND, a self-supervised method aimed to enhance image quality in low-dose CT imaging. This framework combines Wavelet-based Image Alignment (WIA) with Frequency-Aware Multi-scale Loss (FAM) to uniquely address the LDCT denoising task from a frequency perspective. It is a novel approach compared to other self-supervised methods. The efficacy of WIA-LD2ND was validated using two public low-dose CT datasets from AAPM, demonstrating superior performance compared to baseline methods.

  • 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.

    Computational Efficiency: A significant strength of WIA-LD2ND is its minimal increase in training costs, with no additional burden on inference time. The method efficiently utilizes networks trained with Wavelet-based Image Alignment (WIA) and Frequency-Aware Multi-scale Loss (FAM), optimizing inference time. Novelty in Low-Dose CT Denoising: The authors have developed an innovative self-supervised low-dose CT denoising method that exclusively uses normal-dose CT data and analyzes the task from a frequency perspective. They introduce Wavelet-based Image Alignment, which aligns low- and normal-dose CT data, and Frequency-Aware Multi-scale Loss, which captures detailed information effectively. This approach not only enhances image performance in low-dose CT but also eliminates the need for paired data. Consequently, it addresses significant challenges related to dataset construction, such as high costs, privacy issues, and ethical concerns.

  • 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.

    Setting Noise Parameters:The authors mention that the setting of the variance of Gaussian noise in WIA varies depending on the dataset. I suspect that this selection is critical to network performance. However, the approach to parameter selection seems heuristic and may lack a robust, systematic methodology, which could make practical application challenging. Absence of Comprehensive CT Noise Modeling: The method involves simply adding noise to CT images to generate training data, neglecting crucial aspects of denoising in low-dose CT such as streak artifact reduction. This oversight results in notably poor performance in reducing streak artifacts, particularly in areas prone to such issues like the heart and below the spine, as evidenced in Figure 4. The lack of a comprehensive noise model limits the method’s effectiveness in realistic clinical scenarios.

  • 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?

    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

    Consideration of CT Artifacts: Denoising in low-dose CT imaging should address not only noise removal but also artifact reduction, particularly streak artifacts which originate from noise (e.g., Gaussian and Poisson noise) in the projections that are amplified during reconstruction. These artifacts significantly degrade image quality in low-dose CT scans. In the method proposed, there appears to be no mechanism to effectively remove such artifacts. I recommend that future versions of this study focus on integrating solutions specifically aimed at artifact reduction to enhance the overall utility of the method. Detailed Description of Setting Noise Parameters: It is crucial for the reproducibility and validation of your research that a detailed description of how the variance of Gaussian noise parameters is set be included in future iterations. Clear documentation of these parameters would help other researchers more accurately replicate and evaluate the method’s effectiveness. Providing these details is essential to ensuring the method can be rigorously tested and applied in practical settings

  • 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 authors introduce an innovative approach by integrating Wavelet-based Image Alignment (WIA) and Frequency-Aware Multi-scale Loss (FAM), which adds originality to the field of low-dose CT (LDCT) denoising. The unique inclusion of a frequency perspective in a self-supervised method distinguishes this study. However, the paper falls short in addressing streak artifact reduction and lacks detailed documentation of noise parameter settings, which are crucial for reproducibility and effectiveness. Despite these omissions, the novel approach in analyzing LDCT denoising from a frequency perspective merits a Weak Accept recommendation, as it contributes a fresh angle to the existing body of research.

  • 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

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

  • [Post rebuttal] Please justify your decision

    Despite reviewing the authors’ rebuttal, my questions remain unresolved. Additionally, valuable insights from other reviewers reinforce my decision to maintain my current evaluation score.



Review #3

  • Please describe the contribution of the paper

    This paper intrdouces alignig high frequency wavelet decompositions, followed by a frequency aware loss mechanism for LDCT denoising. This is tested on 2 datasets, with promising results.

  • 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.

    I like the idea of adding Gaussian noise to align wavelets, this simplifies the problem. The frequency aware loss is an interesting inclusion, and seems to boost results. The results are demonstrated on two publicly available datasets.

  • 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.

    The ablation study does not compare various backbones, this may simply be an artifact of the chosen backbone. Some data on various inferene times could also be included in the ablation study section.

  • 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

    Conduct study with multiple backbones Add data about inference times

  • 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

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

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

    The innovations were clearly explained, well demonstrated, and easy to implement. The results are of good quality, with the ablations clearly showing the impact of the new modules.

  • 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 thank the reviewers for the valuable comments, which help to improve the quality of this paper. We are encouraged that the reviewers found our motivation and idea to be interesting (R1) and novel (R3, R4), and that our comparisons against baseline showed significant improvements (R1, R3, R4). We respond to your concerns as follows. Detail of backbone (R1, R3): To ensure fairness, we use the same Resnet-based generator backbone as CycleGAN[1] and CUT[2]. WIA-LD2ND achieves significant improvements based on this backbone, as shown in Table 2. We conducted experiments on various backbones, all of which demonstrated good performance. Due to page limitations, we have to make a compromise between detailed analysis of each module and presenting the results for different backbones. Due to the rebuttal policy prohibiting new experimental results, we will publish the experimental results on GitHub. Inference time and computational efficiency (R1, R3): WIA-LD2ND increases the parameters by 2.46M over the baseline. Despite this, it does not increase the inference time, maintaining efficiency while significantly enhancing performance, as shown in Table 2. This demonstrates that the additional parameters contribute effectively to improved outcomes without compromising computational efficiency. Detail of noise parameters (R4): Table 1 in supplementary material details the impact of various noise parameters on model performance. Proposing a systematic and robust noise addition module is a promising direction for future research. Utilization of low-frequency components (R3): As the model can effectively reconstruct low-frequency components, as shown in Sec 2.1, the purpose of FAM is to help with the more challenging high-frequency components. Therefore, we do not utilize low-frequency components in the FAM. Additional dataset (R3): The dataset mentioned by R3 is for CT anomaly detection and is not suitable for the low-dose CT denoising task. We are willing to test the WIA-LD2ND on more datasets. Due to the rebuttal policy prohibiting new experimental results, we will provide the results on GitHub in the future. Comparative analysis (R3): The methods mentioned by R3 are designed for sparse-view CT, which is different from low-dose CT. We are willing to compare WIA-LD2ND with other SOTA methods. We conduct comparative experiments against six SOTA self-supervised methods and two SOTA weakly-supervised methods. Additionally, we are open to further validating WIA-LD2ND by comparing it with more SOTA methods. Due to the rebuttal policy prohibiting new experimental results, we cannot include them here. We will provide this comparison on GitHub in the future. Comprehensive CT Noise Modeling (R4): Streak artifacts mainly occur in sparse-view CT. For low-dose CT, the primary issue is a lower signal-to-noise ratio due to reduced radiation dose, leading to increased image noise rather than streak-like artifacts. WIA-LD2ND is designed for denoising. Future recommendations (R3): We will incorporate MS-SSIM into our method in a subsequent study and cite MS-SSIM. Dataset clarification (R3): The test images in Mayo-2016 and Mayo-2020 are all from distinct patients, as mentioned in Sec. 3.1. Method versatility (R3): WIA-LD2ND is designed for CT denoising. Proposing a method for all types of medical image restoration is a promising future research direction. Performance gain discrepancy (R3): The Mayo-2020 dataset is less noisy than the Mayo-2016 dataset, so models typically perform better with it. [1] Zhu, Jun-Yan and Park, Taesung and Isola, Phillip and Efros, Alexei A: Unpaired image-to-image translation using cycle-consistent adversarial networks. ICCV, 2017 [2] Park, Taesung and Efros, Alexei A and Zhang, Richard and Zhu, Jun-Yan: Contrastive learning for unpaired image-to-image translation. ECCV, 2020



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’

    The authors propose WIA-LD2ND, a self-supervised method designed to improve image quality in low-dose CT imaging. This framework integrates Wavelet-based Image Alignment (WIA) with Frequency-Aware Multi-scale Loss (FAM) to uniquely address LDCT denoising from a frequency perspective. Unlike other self-supervised methods, this novel approach has been validated using two public low-dose CT datasets from AAPM, showing superior performance compared to baseline methods.

  • 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).

    The authors propose WIA-LD2ND, a self-supervised method designed to improve image quality in low-dose CT imaging. This framework integrates Wavelet-based Image Alignment (WIA) with Frequency-Aware Multi-scale Loss (FAM) to uniquely address LDCT denoising from a frequency perspective. Unlike other self-supervised methods, this novel approach has been validated using two public low-dose CT datasets from AAPM, showing superior performance compared to baseline methods.



back to top