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

Abstract

Low-dose computed tomography (LDCT) reduces the risks of radiation exposure but introduces noise and artifacts into CT images. The Feature Pyramid Network (FPN) is a conventional method for extracting multi-scale feature maps from input images. While upper layers in FPN enhance semantic value, details become generalized with reduced spatial resolution at each layer. In this work, we propose a Gradient Guided Co-Retention Feature Pyramid Network (G2CR-FPN) to address the connection between spatial resolution and semantic value beyond feature maps extracted from LDCT images. The network is structured with three essential paths: the bottom-up path utilizes the FPN structure to generate the hierarchical feature maps, representing multi-scale spatial resolutions and semantic values. Meanwhile, the lateral path serves as a skip connection between feature maps with the same spatial resolution, while also functioning feature maps as directional gradients. This path incorporates a gradient approximation, deriving edge-like enhanced feature maps in horizontal and vertical directions. The top-down path incorporates a proposed co-retention block that learns the high-level semantic value embedded in the preceding map of the path. This learning process is guided by the directional gradient approximation of the high-resolution feature map from the bottom-up path. Experimental results on the clinical CT images demonstrated the promising performance of the model.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/3737_paper.pdf

SharedIt Link: https://rdcu.be/dY6fJ

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72390-2_15

Supplementary Material: N/A

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Zho_Gradient_MICCAI2024,
        author = { Zhou, Li and Wang, Dayang and Xu, Yongshun and Han, Shuo and Morovati, Bahareh and Fan, Shuyi and Yu, Hengyong},
        title = { { Gradient Guided Co-Retention Feature Pyramid Network for LDCT Image Denoising } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15012},
        month = {October},
        page = {153 -- 163}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces the Gradient Guided Co-Retention Feature Pyramid Network (G2CR-FPN), which addresses the noise and artifacts in low-dose CT images. The network utilizes three essential paths: bottom-up, lateral, and top-down, incorporating gradient-guided learning to enhance feature maps. Experimental results on clinical CT images demonstrate improved performance.

  • 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 proposed guided-retention operation for inter-feature learning that can emphasize mutual perceptual fields between two feature maps seems interesting to me. This operation should be able to capture interdependencies between high-level and low-level information. Moreover, the experimental results have shown the effectiveness.

  • 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 authors only evaluate the proposed approach by selecting one subject ‘L506’ for testing, and images from the remaining subjects are for training. I am curious about what if to use other settings of data split.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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

    I suggest the authors to do more sets of experiments with different train-test splits in the future.

    In addition, I have a question: what is the difference between the modified Sobel operator introduced by the authors and the standard Sobel operator? What is the specific rationale for using the modified Sobel operator in the approach presented in the manuscript?

  • 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?

    I find the method intriguing and the experimental results are acceptable. I supposed this manuscript could be weakly accepted.

  • Reviewer confidence

    Somewhat confident (2)

  • [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



Review #2

  • Please describe the contribution of the paper

    Authors proposed a novel network for denoising the LDCT introduced in CT images. Specifically author proposed G2CR-FPN network to generate multi-scale features maps and learn better features for LDCT denoising.

  • 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 novelty of the directional gradient approximation. Good results.

  • 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 comparative analysis : the authors should include more recent work for the comparison.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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

    Authors proposed a novel network for denoising the LDCT introduced in CT images. Specifically author proposed G2CR-FPN network to generate multi-scale features maps and learn better features for LDCT denoising.

    1. The method is novel.
    2. the authors should include more recent work for the comparison.
    3. Should add few lines for section e.g. 2. Methods
    4. input images is not clear in abstract.
  • 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. Novelty.
    2. Structure and Write-up.
    3. Results.
  • Reviewer confidence

    Somewhat confident (2)

  • [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



Review #3

  • Please describe the contribution of the paper

    The paper presents the Gradient Guided Co-Retention Feature Pyramid Network (G2CR-FPN) for enhancing LDCT image denoising. This network employs a multi-path architecture: a bottom-up path builds hierarchical feature maps, a lateral path enriches these with gradient-enhanced edge details, and a top-down path integrates high-level semantic content via a co-retention block. This method maintains excellent spatial resolution and semantic integrity, significantly improving image quality. It achieves an RMSE of 7.0516±1.5358 and an SSIM of 0.9602±0.0170 on a clinical CT dataset comprising 900 images, divided into 540 training, 180 validation, and 180 test images. This structured approach optimizes feature extraction and retention, yielding superior denoising performance.

  • 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.

    • Novelty of Method: The introduction of a co-retention block that handles intra- and inter-feature relationships is innovative, enhancing the preservation of edge details and semantic content. This approach differentiates it from previous works by integrating gradient-guided learning within a pyramid structure, which is not extensively explored in existing literature.

    • Robust Evaluation: The paper presents a thorough evaluation against five state-of-the-art denoising methods. The proposed method not only outperforms these in terms of RMSE and SSIM but also demonstrates superior qualitative results through visual assessments conducted by radiologists.

  • 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 paper does not discuss the real-time processing capabilities of the model, nor does it provide details on the computational efficiency or potential for real-time deployment in clinical settings.

    • While the paper discusses the structure and theoretical underpinnings of feature learning within the network, it lacks a detailed examination or visualization of how features are specifically learned and what features are preserved or enhanced.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

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

    The source code availability significantly aids reproducibility.

  • 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

    The paper presents a promising method for LDCT image denoising. Recommendations for future work include:

    • The authors might consider discussing the potential real-time processing capabilities of the model within the existing framework of the study. Including an analysis or discussion section on computational efficiency, based on the model’s architecture and operational requirements, would enhance the paper’s applicability to clinical settings. This could involve a theoretical discussion on the model’s complexity, inference time, and resource requirements based on its structure and previously reported performance metrics.

    • The authors could expand on the theoretical discussion by describing how specific features are emphasized or preserved through the network’s layers. Visualizations such as feature maps at various stages of the network could be included to illustrate this process.

  • 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 innovative integration of gradient guidance and semantic retention within a pyramid network structure provides significant enhancements in LDCT denoising, warranting the positive rating. The technical merit and demonstrated improvements justify acceptance.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 all three reviewers for their constructive comments and appreciations for our strengths. Our point-by-point rebuttals are as follows” Rebuttals to Reviewer #1:

  1. Do more experiments with different train-test splits. Rebuttal: In fact, we did extensive experiments and had more results. However, due to the 8-page limitation, we can not compress all of them into the manuscript. We will show more results in the presentation and extend the conference paper to a full journal paper.
  2. The difference between the modified Sobel operator and the standard Sobel operator. Rebuttal: The standard Sobel operator is generally used as an edge detection operator on natural images. It computes image intensity gradients through fixed isotropic 3x3 kernels. While the modified Sobel operator is proposed with learnable factors on the isotropic kernels. It learns the directional gradients on feature maps, and it is a data-driven kernel for better denoising performance.
    Rebuttals to Reviewer #3:
  3. The paper does not discuss the potential real-time processing capabilities of the mode. Rebuttal: The computational cost for the proposed algorithm heavily depends on the hardware configuration. Since the resource requirements and the source codes regarding the model size and inference time are provided through the anonymized link, the potential real-time processing capabilities can be easily tested by the readers on their own hardware platform.
  4. Visualization of the feature maps. Rebuttal: Due to 8-page limitation, we did not put the visualized feature maps into the manuscript. We will show the visualized feature maps in the presentation and put them into a follow-up extended journal paper to include more details.
    Rebuttals to Reviewer #4:
  5. More recent work for the comparison. Rebuttal: We have compared with more recent works. We will show more comparison results with respect to the sate-of-the-art in the conference presentation and put them into a follow-up extended journal paper.
  6. Add few lines for section, e.g. 2. Methods. Rebuttal: Due to 8-page limitation, we deleted many details in our first 14-page draft. In a follow-up full journal paper, we will recover the deleted details in section 2.
  7. Input images is not clear in abstract. Rebuttal: The input images in our approach are low-dose CT images. We implicitly wrote that: It is highly challenging to denoise LDCT images while preserving essential diagnostic features.



Meta-Review

Meta-review not available, early accepted paper.



back to top