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

Abstract

Aortic dissection (AD) is a severe cardiovascular emergency requiring prompt and precise diagnosis for better survival chances. Given the limited use of Contrast-Enhanced Computed Tomography (CE-CT) in routine clinical screenings, this study presents a new method that enhances the diagnostic process using Non-Contrast-Enhanced CT (NCE-CT) images. In detail, we integrate biomechanical and hemodynamic physical priors into a 3D U-Net model and utilize a transformer encoder to extract superior global features, along with a cGAN-inspired discriminator for the generation of realistic CE-CT-like images. The proposed model not only innovates AD detection on NCE-CT but also provides a safer alternative for patients contraindicated for contrast agents. Comparative evaluations and ablation studies against existing methods demonstrate the superiority of our model in terms of recall, AUC, and F1 score metrics standing at 0.882, 0.855, and 0.829, respectively. Incorporating physical priors into diagnostics offers a significant, nuanced, and non-invasive advancement, seamlessly integrating medical imaging with the dynamic aspects of human physiology. Our code is available at https://github.com/Yukui-1999/PIAD.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/0345_paper.pdf

SharedIt Link: https://rdcu.be/dV5Eu

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72104-5_53

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/0345_supp.pdf

Link to the Code Repository

https://github.com/Yukui-1999/PIAD

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Din_Physicalpriorsguided_MICCAI2024,
        author = { Ding, Zhengyao and Hu, Yujian and Zhang, Hongkun and Wu, Fei and Yang, Shifeng and Du, Xiaolong and Xiang, Yilang and Li, Tian and Chu, Xuesen and Huang, Zhengxing},
        title = { { Physical-priors-guided Aortic Dissection Detection using Non-Contrast-Enhanced CT images } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15007},
        month = {October},
        page = {551 -- 561}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper combines model-based computational fluid dynamics and neural networks (convolutional and transformers) to improve (1) segmentation of non-contrast CT for true/false lumen and of aortic dissection (2) contrast-enhanced CT generation from non-contrast CT. The technical contribution seems to be (1) how the CFD results were incorporated into training and (2) multi-task learning for improved performance on both segmentation and image generation tasks.

  • 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 motivation is clear, and the method is mostly well described. There are notable improvements in performance metrics all across the board. Figures and tables are well organized. Amount of validation data and experiments seem sufficient.

  • 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.

    Clarity could be improved for method description. Details in the comments below. The paper should be more critical in terms of the algorithm’s potential clinical use. “Our model enables the early screening of patients with aortic dissection using NCECT” seems to be an overstatement, as Fig. 4 shows visible errors in generated images and segmentation, which can significantly alter clinical decisions. No statistical significance tests on the results.

  • 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?

    Addressing some of the clarifying questions would help increase 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 acronym PIAD is not defined. It is unclear whether the 3 most important CFD features are 3 scalars for the entire image or 3 scalars at each voxel of CE-CT. If it is 3 scalars for the entire image, which the figures seemed to suggest, then how’s the Unet output converted to 3 scalars? What’s the advantage of a Unet as opposed to a convolutional encoder followed by an MLP? Qualitative evaluations show that the generated image’s dissection flap geometry is very different from CE-CT for all methods. Inaccurate image generation is potentially worse than no generation. Segmentation seems better, but still with some visible errors. These findings should be discussed in a more unbiased way.

  • 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?

    Overall, the paper is thorough and presents interesting results. With some clarifications and rephrasing, I believe the paper should be of sufficient quality for MICCAI.

  • 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

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper presents a novel method for the detection of Aortic Dissection (AD) using Non-Contrast-Enhanced CT (NCE-CT) images, which is a significant advancement over the reliance on Contrast-Enhanced Computed Tomography (CE-CT). The key contributions are threefold:

    1. The introduction of a new method that leverages NCE-CT for AD detection, addressing the limitations and risks associated with CE-CT exams.
    2. The integration of biomechanical and hemodynamic physical priors into a 3D U-Net model, which encapsulates the influence of physical factors on AD detection, offering a more comprehensive understanding of the disease.
    3. The effectiveness of the proposed framework was demonstrated through experiments on three datasets, and the effectiveness of the physically prior guided model was demonstrated through ablation experiments.
  • 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 main strengths of the paper are as follows:

    1. The paper introduces a unique approach by integrating biomechanical and hemodynamic physical priors into a 3D U-Net model. This is a novel aspect as it goes beyond traditional pixel-based methods to consider the complex physiological conditions that contribute to aortic dissection, which is particularly interesting because it may lead to more accurate and reliable diagnoses.
    2. The application of Computational Fluid Dynamics (CFD) to calculate hemodynamic parameters and the subsequent use of a neural network to predict these parameters in the absence of direct measurements is an innovative way to use data. This approach enhances the model’s practicality for real-world applications by estimating crucial hemodynamic parameters without the need for direct measurements.
    3. The model is designed to perform multiple tasks simultaneously, including segmentation, generation, and classification. This is a strong aspect of the work as it allows for a more holistic and integrated approach to AD detection, which can potentially lead to better performance compared to single-task models.
  • 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.

    While the paper presents advancement in the detection of Aortic Dissection (AD) using Non-Contrast-Enhanced CT (NCE-CT) images, there may be several potential weaknesses to consider:

    1. Incompleteness of Experiments: 1.1 The study lacks a comparative analysis using real CFD data inputs to highlight the importance of incorporating physical priors and to assess the modeling error that may stem from the CFD Predictor’s integration into the learning phase of the model. 1.2 The ablation study is not exhaustive. It includes experiments that isolate the effects of physical priors, the Transformer encoder, and the Cascade Classifier, but does not explore every possible pairwise combination of these elements. Furthermore, the experiments for Single Generation, Single Segmentation, and Single Classification also lack a pairwise comparison, which is essential for isolating the impact of each task on the overall model performance.
    2. Writing Clarity and Consistency: The paper’s primary narrative revolves around utilizing NCE-CT for the detection of AD, with a focus on classification. The additional tasks of segmentation and generation, which are presented as auxiliary to classification, might cause some confusion, especially since the visualization results suggest they serve to support the classification task. The caption in Figure 2, which describes the model as designed for all three tasks (segmentation, generation, and classification), could mislead readers regarding the model’s primary objectives and the interplay between these tasks.
    3. Model Design Specificity: The paper lacks specificity regarding the integration of physical priors into the model. There is no detailed explanation of the methods used for this integration or references to previous work that substantiates this approach. The paper only generally cites the concept of attention mechanisms without detailing how it applies to the incorporation of physical priors.

    There are anothers potential weaknesses:

    1. Lack of Justification for Feature Selection: The rationale for choosing only the top three hemodynamic features as identified by XGBoost is not explained, leaving it uncertain if this selection is optimal or if additional features could enhance the model’s performance.
    2. Unsubstantiated Simplification of Physical Parameters: The paper opts for a simplified representation of physical information by using the ratio of extremal values of each parameter, allegedly to decrease the model’s complexity. However, there is a lack of experimental validation to demonstrate that this approach is effective and does not result in a loss of critical information for the model’s predictive accuracy.
    3. Insufficient Training and Ambiguity in Generator Output: There is ambiguity regarding the specific nature of the images generated by the model—whether they are of vascular structures in CE-CT or another aspect of CE-CT imagery. If the generator’s task is not confined to vascular CE-CT images, the
  • 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 provide sufficient information for reproducibility.

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

    Please make the code open-source and validate it using a publicly available dataset (I am not an expert in this field, so if one exists, it should be utilized). The description of the algorithms in the paper is relatively clear. However, as mentioned in the previous discussion of the paper’s weaknesses, a more detailed description of the integration method of physical priors with the network and the selection of the number of hemodynamic features is required.

  • 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. Open-source the code and validate it against a public dataset if possible.
    2. Conduct additional ablation studies to better understand the contribution of each model component.
    3. Clarify the roles of segmentation and generation tasks in the classification process.
    4. Provide a detailed explanation of the fusion method for physical priors.
    5. Discuss the rationale behind the selection of the top three hemodynamic features.
  • 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?

    This paper introduces a method for aortic dissection detection using non-contrast-enhanced CT, addressing the potential risks associated with the use of contrast agents in CE-CT for patients with allergies or acute renal failure. Innovatively, it incorporates physical priors, specifically hemodynamics, to guide the detection of aortic dissection (AD). The effectiveness of the proposed framework is demonstrated through experiments across three datasets, and the utility of the model guided by physical priors is validated through ablation studies.  However, the paper also exhibits several weaknesses: 

    1. The ablation studies are incomplete.
    2. The description of the algorithms is not fully detailed.
    3. The selection of hemodynamic features as physical priors lacks thorough experimental justification.
    4. The clarity and consistency of the writing could be improved to enhance understanding and coherence.
  • 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

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper introduces a physical-priors-guided method (PIAD) for aortic dissection detection using non-contrast-enhanced CT. PIAD is a multi-task network that contains physical parameter prediction, lumen segmentation and contrast-enhanced (CE)-CT generation. This method is evaluated on one internal dataset and two external datasets, showing performance improvement over previous 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.
    1. The motivation of avoiding using contrast agents has strong clinical significance.
    2. The prediction of physical priors and their integration to the network is novel and interesting, which could inspire further exploration of using physical priors in deep learning networks.
    3. The evaluation and experiments seem solid. There are three datasets used for evaluation and ablation studies are provided.
  • 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.

    Overall, I think this is a good paper. However, statistical significance is not tested. I wonder if the differences between the proposed method and the second-best methods are statistically significant.

  • 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
    1. Please add statistical significance tests.
    2. Is using CE-CT for AD detection considered as upper bound? If so, what is the classification performance for the methods that use CE-CT?
    3. Why the image generation results (columns 3-6) in Fig. 4 look a lot different from the ground truth CE-CT (column 2)? Does the network only learn to generate the aortic region?
  • 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?

    This paper proposes a novel method to integrate physical priors and improves the diagnostic performance on non-contrast CT.

  • 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

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Author Feedback

Thanks for the valuable comments and suggestions. Below we respond to the comments. Code available: According to MICCAI’s requirements, our code will be provided in the final version of the paper.

  1. To reviewer 1: ①PIAD is an acronym for Physically Informed Aortic Dissection Detection.②We believe that Unet enhances the segmentation assistance of the model better compared to CNN or MLP, which is more helpful for aortic dissection detection. This architecture is also better adapted to the embedding of physical information (similar to the DDPM paper). ③ We will discuss the results of generation and segmentation more reasonably and fairly if allowed.
  2. To reviewers 1&4: Physical information selection: After performing CFD calculations on CE-CT, we obtain several scalar features at each voxel within the aorta’s true lumen. We selected three indicators based on their importance ranking determined by XGBoost for predicting aortic dissection (AD). These indicators primarily represent blood flow velocity and mesh quality. In cases of AD, the true lumen’s shape is often irregular, leading to lower mesh quality and localized maxima in blood flow velocity (as shown in Figure 3a of the submitted paper). We used the ratio of the maximum to minimum values of each voxel feature as our final feature due to initial unsuccessful experiments. Our original plan to use Physics-Informed Neural Networks (PINNs) to calculate blood flow parameters for all voxels in the aorta proved to be highly challenging and error-prone. Therefore, we adopted the simplified method presented in the paper. Selecting three indicators instead of ten does not significantly impact the results and can reduce the difficulty of forecasting.
  3. To reviewers 1&4: The way physical information is combined with Unet: Our method was inspired by the way U-Net combines with timestep t in the diffusion model. In our approach, at each layer of the U-Net, we assume the shape of the image data is [batch, channel, depth, height, width], and the shape of the physical information data is [batch, 3]. We reshape the image data to [batch, channel, depth × height × width] and the physical information parameters to [batch, 1, 3]. Then we perform cross-attention where Q is the physical information data, and K and V are the image data. This enhances the model’s ability to distinguish based on specific physical information. We hope this explanation helps
  4. To reviewer 3: ①Per MICCAI’s requirements, we cannot add significance test results here. If allowed, I will include them in the final version of the paper. Preliminary results are promising. ②The classification performance of CE-CT is widely recognized as an upper bound (clinical gold standard), with sensitivity and specificity exceeding 95%, as reported in related papers.
  5. To reviewers 3&4: Model region of interest: Our model uses nnUnet to segment the aortic region, focusing on the true and false lumen. Research shows that accurate diagnosis of aortic dissection mainly requires attention to the aorta, with minimal contribution from other CT scan regions.
  6. To reviewer 4: Experimental adequacy and clarity of writing: ①While our results validate the effectiveness of physical information, the control experiment with real CFD data you mentioned is necessary to improve confidence and provide an upper bound for model predictions. If allowed, we will include this in the revised paper. ②In the ablation study, the effects of the physical priors, the isolation of the Transformer encoder, and the classifier cascade should be considered as combination experiments of the remaining two. The single-task ablation study is intended to highlight the effectiveness of multi-task joint learning, and further ablating each part in the single-task experiment may not be particularly meaningful. ③ We will improve the text’s clarity and consistency in the final version to emphasize the importance of classification over segmentation and generation.



Meta-Review

Meta-review not available, early accepted paper.



back to top