Abstract

MLP−based networks, while being lighter than traditional convolution− and transformer−based networks commonly used in medical image segmentation, often struggle with capturing local structures due to the limitations of fully−connected (FC) layers, making them less ideal for such tasks. To address this issue, we design a Dual−Path MLP−based network (DPMNet) that includes a global and a local branch to understand the input images at different scales. In the two branches, we design an Axial Residual Connection MLP module (ARC−MLP) to combine it with CNNs to capture the input image’s global long-range dependencies and local visual structures simultaneously. Addi- tionally, we propose a Shifted Channel−Mixer MLP block (SCM−MLP) across width and height as a key component of ARC−MLP to mix information from different spatial locations and channels. Extensive experiments demonstrate that the DPMNet significantly outperforms seven state−of−the−art convolution− , transformer−, and MLP−based methods in both Dice and IoU scores, where the Dice and IoU scores for the IAS−L dataset are 88.98% and 80.31% respectively. Code is available at https://github.com/zx123868/DPMNet.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/2829_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: N/A

Link to the Code Repository

https://github.com/zx123868/DPMNet

Link to the Dataset(s)

http://medicaldecathlon.com/dataaws/

BibTex

@InProceedings{Wan_DPMNet_MICCAI2024,
        author = { Wang, Shudong and Zhao, Xue and Zhang, Yulin and Zhao, Yawu and Zhao, Zhiyuan and Ding, Hengtao and Chen, Tianxing and Qiao, Sibo},
        title = { { DPMNet: Dual-Path MLP-based Network for Aneurysm Image Segmentation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15009},
        month = {October},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This work proposes a dual-path network based on MLP and CNN for image segmentation. Authors propose a new local branch, a shifted channel mixer MLP block (to improve local context understanding of FC), and a modified Axial residual connection MLP block (compared to UNeXt). They applied their network to aneurysms segmentation using one private (?) database and to tumors segmentation using 2 publicly available datasets.

  • 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.
    • usage of a mix of MLP/CNN… maybe some links with transformer.
    • many deep contributions to UNeXt and related architectures
    • Very convincing results. Some std values are very impressive ! (table 2: +/-0.05 and +/- 0.09 for DICE and IoU of DPMNet)
  • 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.
    • it’s not obvious whether the hyperparameters of other networks have been optimized. We regret that nnUnet is not used.
    • For clinical application, in addition to the dice, the number of lesions missed by the algorithms should also be quantified.
  • 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 can consider these two points : Adding nnUNet to tested networks for faire comparison. Change IoU by an accuracy based on IoU to quantified missed lesions (table 2). If authors can discuss on the feasibility to segment multiclass and not only binary?

  • 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?

    very interesting approach and results. Almostly clear paper

  • 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

    Authors proposed a new deep-learning model for aneurysm image segmentation. They proposed to extend previous architecture mixing convolutional with multiple layer perception networks. Authors main contribution is to propose a axial residual connection and shifted channel mixer. By doing so, authors achieved competitive performance compared to state-of-the-art 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.
    • The authors proposed a mechanism for improving the ability of fully convolitional layers to capture local information.
    • Authors compared the results of their method with a large set of different methods.
    • Conducted experiences indicates increase segmetnation accuracy using the proposed methods.
  • 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 comparison with transformer architecture is limited to the TransUnet, however, many new transformer architecture have been proposed (eg, SwinUnet), the comparison should include more transformer to have a better understanding of performance differences.

    • Figure 2. Without proper caption, describing the figure, it is complicated to understand what each color encodes.

  • 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
    • It is not clear from the ablation study how DPMNet enables to produce more accurate segmentation compared to Local Branch and Global Branch. One of the biggest difference seems to be the downsampling in between Local and Global. Wouldn’t using another degree of down-sampling in the ARC-MLP of the Global Branch results to the same improvement?
  • 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?
    • Authors proposed a new method
    • This method has been compared to multiple others methods using a rigorous evaluation framework
    • Results indicates competitive performances
  • 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



Review #3

  • Please describe the contribution of the paper

    This paper introduces DPMNet, a dual-path MLP-based network that enhances aneurysm image segmentation. Key contributions include the integration of an Axial Residual Connection MLP (ARC-MLP) with CNNs to capture both global and local image features, the development of a Shifted Channel-Mixer MLP (SCM-MLP) to improve channel and spatial interactions, and the demonstration of superior performance, outperforming seven state-of-the-art methods in Dice and IoU metrics across five datasets.

  • 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. Exceptional Clarity and Organization: The paper is well-structured and articulately written, ensuring that readers can easily follow and understand the content. The inclusion of clear and informative figures further aids comprehension, illustrating complex concepts and the workings of the proposed network effectively.

    2. Thorough Analysis of the Novel Approach: The paper provides an in-depth analysis of the novel dual-path MLP-based network. Each component, including the Axial Residual Connection MLP and the Shifted Channel-Mixer MLP, is thoroughly explained and detailed. This meticulous detailing underscores the innovative aspects of the approach and their contributions to the field.

    3. Outstanding Performance: The proposed method demonstrates excellent performance, outperforming existing state-of-the-art models in both Dice and IoU scores across multiple datasets. This significant achievement highlights the effectiveness and potential of DPMNet in advancing image segmentation technologies across multiple areas.

  • 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 visualizations of segmentation results are too small and lack clarity, making it difficult to effectively assess the qualitative performance of the proposed method. Enhancing the size and readability of these visualizations would better demonstrate the advantages of the proposed approach and strengthen the paper’s visual impact.

    The ablation study, which is crucial for understanding the contribution of individual components of the proposed framework, shows that the improvements in segmentation metrics are marginal, at most 1%. Considering the increased parameter size of the model, these minor improvements do not compellingly justify the added complexity.

  • Please rate the clarity and organization of this paper

    Excellent

  • 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. Improvement of Segmentation Visualizations (Fig. 3): The current segmentation visualizations present masks that are too small relative to the size of the images, which weakens the ability to compare the differences between models’ predictions and assess the effectiveness of the proposed method. I recommend cropping the images to focus more closely on the masked regions. This adjustment would make the masks a larger portion of the image and allow for a more detailed comparison.

    2. Diversification of Segmentation Metrics: The paper primarily utilizes Dice and IoU metrics, which are both similarity measures. Therefore IoU may provide redundant information. To offer a more comprehensive evaluation of the segmentation performance, I suggest incorporating additional types of metrics, such as the Average Hausdorff Distance. This metric could provide insights into the geometric differences between the predicted and actual segmentations. For reference, consider the comprehensive overview of segmentation metrics in the following paper: Taha, A.A., Hanbury, A. “Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool,” BMC Med Imaging 15, 29 (2015). https://doi.org/10.1186/s12880-015-0068-x.

    3. Reevaluation of Ablation Study Results: The ablation study shows only marginal improvements in segmentation metrics (at most 1%) with a significant increase in model parameters (from 9 million for the Global Branch to 31 million for DPMNet). This raises concerns about the practicality and efficiency of the added complexity. I recommend revising this section to provide a more compelling justification and explaining clearly why those numbers are significant. For example, an increase in IoU from 80.07±0.16 for the Global Branch to 80.31±0.17 for DPMNet, given the substantial growth in parameters, needs a stronger rationale to convincingly demonstrate the benefits of the proposed method.

  • 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 introduces a novel approach that effectively incorporates several recent innovations, resulting in impressive segmentation performance. The manuscript is exceptionally well-organized and clearly written, making it accessible and informative. However, there are a few areas for potential improvement that prevent a perfect score. Notably, enhancing the clarity of visualizations and expanding the diversity of evaluation metrics could further strengthen the paper. Considering these points, I rate this paper as a 5 - Accept

  • 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 thank the reviewers for their valuable feedback. We are pleased to see that the reviewers agree on the novelty and efficacy of the proposed model for aneurysm image segmentation. Some major questions are addressed below. (1)Visualisation: R4. For the segmentation results in Fig. 3, we will crop the images and adjust the masks to take up a larger proportion of the images, thus enhancing the visualisation of the differences between the comparative models’ predictions. (2)Ablation experiments: R1, R4. The core innovation of this paper is the proposed SCM-MLP and ARC-MLP, they are the fundamental reason for the excellent performance of the DPMNet in this paper. So as can be seen from Table 3, even global branching containing only three layers of the encoder-decoder structure achieves good results. In addition, the dual-path strategy using global and local branches allows the model to capture both the long-range spatial dependencies between image patches and the high-level semantic details of the image. The adoption of the dual-path strategy logically solves the problem of FC layers’ lack of localisation; as a result, it improves certain segmentation performance. However, it does bring about a growth in the number of model parameters, and we will consider optimising the network to reduce the number of model parameters in future work to achieve a more lightweight model with better segmentation performance. (3)Colour in Fig. 2: R1. The colours used in Fig. 2 are intended to provide a more intuitive and aesthetically pleasing representation of the process of shifting in width and height and mixing the different channels. (4)Multiclassification: R5. Data labelling of medical images is a very time-consuming and complex task. The aneurysm dataset used in the paper was manually labelled with aneurysms by skilled medical professionals as a ground truth. Since the labels obtained are binary, only binary segmentation of the aneurysm images is performed in this paper. (5)Comparison experiments: R1, R5. All hyperparameters of the other models compared used the default values from the original paper. Once again, we sincerely thank the reviewers for their valuable feedback, and we will take all of their comments into consideration for the camera-ready version of our paper.




Meta-Review

Meta-review not available, early accepted paper.



back to top