Paper Info Reviews Meta-Review Author Feedback Post-rebuttal Meta-Reviews Back to top
List of Papers By topics Author List
Paper Info Reviews Meta-review Author Feedback Post-Rebuttal Meta-reviews

Authors

Zhiwei Deng, Songnan Xu, Jianwei Zhang, Jiong Zhang, Danny J. Wang, Lirong Yan, Yonggang Shi

Abstract

The automated segmentation and analysis of small vessels from \emph{in vivo} imaging data is an important task for many clinical applications. While current filtering and learning methods have achieved good performance on the segmentation of large vessels, they are sub-optimal for small vessel detection due to their apparent geometric irregularity and weak contrast given the relatively limited resolution of existing imaging techniques. In addition, for supervised learning approaches, the acquisition of accurate pixel-wise annotations in these small vascular regions heavily relies on skilled experts. In this work, we propose a novel self-supervised network to tackle these challenges and improve the detection of small vessels from 3D imaging data. First, our network maximizes a novel shape-aware flux-based measure to enhance the estimation of small vasculature with non-circular and irregular appearances. Then, we develop novel local contrast guided attention(LCA) and enhancement(LCE) modules to boost the vesselness responses of vascular regions of low contrast. In our experiments, we compare with four filtering-based methods and a state-of-the-art self-supervised deep learning method in multiple 3D datasets to demonstrate that our method achieves significant improvement in all datasets. Further analysis and ablation studies have also been performed to assess the contributions of various modules to the improved performance in 3D small vessel segmentation.

Link to paper

DOI: https://doi.org/10.1007/978-3-031-43901-8_34

SharedIt: https://rdcu.be/dnwDI

Link to the code repository

https://github.com/dengchihwei/LCNetVesselSeg

Link to the dataset(s)

N/A

Reviews

Review #4

  • Please describe the contribution of the paper

    This paper proposed a deep learning-based flux filter for segmenting small blood vessels in 3D MRI images. The flux filter predicted 128 radii on 128 sampling directions, where a local contrast attention and a local contrast enhancement module were calculated by integrating over these directions. Self-supervised losses were calculated using the vesselness and direction continuity over the estimated principal directions.

  • 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 method in the paper is novel in the following ways:

    1. the proposed flux filter is novel in the way that it uses a deep neural network to predict the radii on all 128 directions;
    2. the proposed local contrast attention and enhancement modules are novel in the way that they aggregates information over the estimated radii channels to calculate vessel contrast;
    3. the proposed self-supervised losses are novel in the way that a principal direction-based negative vesselness loss and a path-wise loss were calculated.
  • 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 figure illustrations of the model architectures are not very clear, which makes it hard to reproduce the work;
    2. The self-supervised losses rely on the calculation of the principal directions, which seem not to be differentiable;
    3. The only ablation study was performed over sampling schemes and the local contrast attention, but local contrast enhancement module was missing.
    4. Some hyperparameters are very important to the calculation of the principal direction and the self-supervised loss, e.g., the number of directions m, and the length of the path for calculating L_path, and more detailed experiments for choosing these hyper parameters should be given.
  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

    Overall the idea and the network seem to be valid and reproducible. But there is a discontinuity between the illustration in Figure 3 and Figure 2, which is confusing to the readers and might be a blocker for reproducibility.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
    1. The illustration colors and arrows don’t match between Figure 2 and Figure 3, e.g., it seems the LCA and LCE modules take inputs from the dashed arrows in Figure 2, but they only take single inputs in Figure 3. Besides, the caption of Figure 2 mentioned scalar fields R and vector field P, but there is no illustration in the figure for these two items.
    2. The actual value for the small constant in Equation (5) is not mentioned anywhere in the experiment section.
    3. For the self-supervised flux loss, since it depends on the principal direction, which can only be indirectly estimated using the predicted radii - would it make the loss not differentiable to the network weights?
    4. Some of the hyperparameters are crucial to the principal direction and self-supervised loss calculation, e.g., the number of directions m and the length of the path calculating L_path. A detailed experiments choosing the hyperparameter values should be given.
  • Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

    6

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

    The novelties of the paper are prominent, including a novel deep learning-based flux filter, local contrast attention and enhancement modules, and self-supervised losses. The only downside of the work are the illustrations and the hyperparameter choices, which don’t impede the originality of the work.

  • Reviewer confidence

    Very confident

  • [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 addresses a very relevant and difficult issue of segmenting small vessels from 3D imaging data with the aid of a self supervised network. The authors put forth a novel network that first reconstructs the irregular cross-section profiles of small vessels. Then, the local contrast attention (LCA) module improves the small vessel pattern while simultaneously suppressing the background clutter. Finally, the model determines the properties of small vessel structures in relatively limited resolution. The proposed method shows promising improvements in 3D datasets of multiple modalities compared to previous unsupervised approaches.

  • 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 formulates a novel self-supervised network to detect and analyse small vessels in non-invasive imaging data. This is particularly important for studying cerebral small vessel disease, which is a common cause of dementia.

    2. The advantage is that the method can achieve a high dice score at areas of small vessels with very small radius and low contrasts, and this is an improvement over the conventional filters and the flow-based DL method. Additionally, in contrast to the flow-based DL, the proposed model can produce sharper and clearer vesselness maps.

    3. The authors proposed a flux-based loss as the negative average vesselness score over the whole image spatial space. This cost function takes advantage of the clear edge separation of flux-based filter and robustness to irregular-shaped vessels.

    4. The ablation studies show that for the 7T MRA dataset, the proposed modules improve the AUC metric of the baseline by 4.87% on the test set as compared to the flow-based DL method.

    5. The experimental set-up is well-devised across four different datasets (two public and two in-house) and CT, MRA, and MRI modalities.

    6. I appreciate that the authors compared their work against multiple methods. The paper is well-organized.

  • 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. Some of the improvements listed in Table 1 are actually quite small, mostly within 1-2 AUC points (in some cases this even includes for accuracy, sensitivity, and specificity). The authors did not perform a statistical evaluation to assess the significance of the reported improvement.

    2. The authors neither gave an intuitive nor an empirical explanation for selected values of the network’s hyperparameter, such as m, λ1, λ2 and λ3. Moreover, it is unclear how these hyperparameters were selected (within a single execution of the train and validation steps, or the best cross-validation performance).

    3. The paper fails to shed light on the overall duration of the hyperparameter tuning and training steps. Also, the authors do not discuss the time and memory footprint needed for the proposed DL-approach to segment (inference) the small vessels when compared to the conventional filters.

  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

    The paper uses two public and two private datasets, and the source-code is not supplied in the manuscript. Therefore, it is only possible to completely reproduce the main results if all the data and scripts are provided.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
    1. The authors should perform paired tests (maybe Wilcoxon signed-rank test) to determine the statistical significance of their performances.
    2. The implementation subsection lacks the description of the network’s hyperparameters. Do the hyperparameters vary for each dataset? Please elaborate on the selection processes of the hyperparameters and the hard threshold.
    3. Additionally, please provide sufficient documentation (resampling, normalisation method, augmentation strategies, etc.) for researchers to confidently replicate the results.

    Minor Issues:

    1. I found a few typographical and grammatical errors, for example : (a) in the Introduction, add a space after the number and before the SI unit (0.5 mm instead of 0.5mm) (b) certain words, such as suboptimal, should not be hyphenated. (c) in the Implementation Details subsection there is a sentence, “We reported five metrics in Table 2, namely, the area-under-curve(AUC), accuracy, sensitivity, specificity, and dice score.” The Table should be 1, not 2.
    2. I request the authors to enlarge Figure 4 should for a better view.

    Possible future work:

    1. It would be good to include an ablation study of the network’s hyperparameters to properly evaluate their impact.
    2. Can the conventional filters (eg. Frangi or OOF) be used in conjunction with the proposed model as an inductive bias, or perhaps as a pre-processing step to refine the image? How might this impact the overall performance?
  • Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

    5

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

    The manuscript has presented explorative ideas with a lot of potential across the field. Even though there are minute flaws such as small improvements and tuned hyperparameters, the strengths out weigh them. I believe that this detailed analysis is important for the community, and therefore, I recommend a weak accept.

  • Reviewer confidence

    Confident but not absolutely certain

  • [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

    This paper proposed an unsupervised method for 3D vessel segmentation, targeting at small vessels with irregular appearances and low contrast. By adaptively estimating the oriented fluxes and introducing attention mechanism, this paper improves the vessel segmentation 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 research is of great practical significance. The writng of this paper is logical and accurate in general. The method is relatively innovative. Plenty of datasets are used. Experimental results gain reasonable improvement.

  • 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 title does not indicate “self-supervision”. If you don’t highlight it in this paper, you should compare your method with other existing supervised methods. (2) In Fig1, it is better to add some words and pictures of current methods’ poor performance on small vessels to show the truth that small vessel segmentation is the main challenge. Only the characteristics of small vessels are not enough. “These characteristics are generally not well modeled by existing methods for vessel detection” needs to be visualized. (3) In Fig2, point out the necessary notations in picture, such as raw image I, the output: P, R, reconstructed image I ̂ (4) In analysis of small vessel segmentation, the experiment lacks quantitative comparison results. Since images in some datasets include large vessels and small vessels, Table1 is supposed to presented in more detail according to thickness. (5) In Fig4, only performance on TubeTK and 7T MRA are visualized, which is not sufficient. the other two datasets should be included to show some segmentation maps. And why false negatives are considered other than false positives? (6) The descriptions of “CS” and “AS” are not clear, do they both use the three proposed loss or “CS” only use some of them? MSE loss seems to have nothing to do with sampling strategy.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

    Yes

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html

    Please refer to my comments in the “weaknesses” section.

  • Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

    5

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

    In general, the work has some novelty and the writting is good. The experimental results can be further improved by including comparisons with recent SOTA supervised methods.

  • Reviewer confidence

    Very confident

  • [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



Primary Meta-Review

  • Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    This paper is about the segmentation of blood vessels in 3D MRI images. The method uses a deep learning-based flux filter, local contrast attention module, and self-supervised loss. The proposed method is novel and shows promising results.



Author Feedback

N/A



back to top