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

Abstract

In this paper, we address the segmentation of tumor subtypes in whole slide images (WSI) by utilizing incomplete label proportions. Specifically, we utilize partial' label proportions, which give the proportions among tumor subtypes but do not give the proportion between tumor and non-tumor. Partial label proportions are recorded as the standard diagnostic information by pathologists, and we, therefore, want to use them for realizing the segmentation model that can classify each WSI patch into one of the tumor subtypes or non-tumor. We call this problem `learning from partial label proportions (LPLP)’’ and formulate the problem as a weakly supervised learning problem. Then, we propose an efficient algorithm for this challenging problem by decomposing it into two weakly supervised learning subproblems: multiple instance learning (MIL) and learning from label proportions (LLP). These subproblems are optimized efficiently in the end-to-end manner. The effectiveness of our algorithm is demonstrated through experiments conducted on two WSI datasets. This code is available at https://github.com/matsuo-shinnosuke/LPLP.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1667_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/1667_supp.zip

Link to the Code Repository

https://github.com/matsuo-shinnosuke/LPLP

Link to the Dataset(s)

https://www.cs.toronto.edu/~kriz/cifar.html http://ufldl.stanford.edu/housenumbers/ https://zenodo.org/records/1214456

BibTex

@InProceedings{Mat_Learning_MICCAI2024,
        author = { Matsuo, Shinnosuke and Suehiro, Daiki and Uchida, Seiichi and Ito, Hiroaki and Terada, Kazuhiro and Yoshizawa, Akihiko and Bise, Ryoma},
        title = { { Learning from Partial Label Proportions for Whole Slide Image Segmentation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15011},
        month = {October},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a weakly supervised MIL-based approach, leveraging bag-level labels and positive label proportions to directly learn instance-level classes. Unlike a two-step sequential process where an MIL model is first trained on bag-level labels followed by a label proportion model, this method adopts an end-to-end approach, even demonstrating improved performance compared to the sequential approach. Promising results are demonstrated across two datasets, with comparisons drawn against both a two-stage pipeline and a strongly supervised model trained on instance-level labels.

  • 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.
    • Paper is well written with easy to read notations to understand the end-to-end learning process
    • The paper proposes a clever way to utilize the proportions of tumors reported as an additional supervision to get instance level tumor labels from a WSI level label.
    • The proposed method is also end-to-end removing the need to train the model in multiple stages. This is done by computing the proportions of tumors from the outputs of the MIL model in a differential way and minimizing a cross entropy style loss on the label proportions
  • 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.
    • As mentioned by the authors, the label proportion labels are with respect to the tumor regions and not with respect to the whole WSI, this could lead to potential degenerate solution which is not addressed by the authors. For.eg. Let us consider two cases, Case-1: The tumor region (consisting of two tumor types) occupies 50% of the WSI and the proportion of tumor-1 is 20% and tumor-2 is 80%, Case-2: The tumor region occupies only 10% of the WSI and proportion of the tumor-1 & tumor-2 are the same as in Case-1. LLP loss is not incentivized to distinguish between the two conditions, in other words, if the network assigns 2 patches to tumor-1 and 8 patches to tumor-2 or 20 patches to tumor-1 and 80 patches to tumor-2, the LLP loss is not going to change as long as the ratio remains the same.
    • The authors did not include a reason for a significant performance gap between the two-stage approach and the proposed method. For.eg. Fig.3 shows that Two-stage approach highlight more regions to be tumorous compared to the proposed method which is counterintuitive to the optimization problem of MIL where finding a single positive instance is enough to minimize the MIL loss
  • 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?

    The proposed method is clearly explained and would be relatively easy to reproduce

  • 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
    • A great ablation would be to train the proposed model on synthetic data by keeping the proportions of tumors constant but varying the proportion of positive instances (see my comment above) between WSIs to really get an understanding of the combined loss MIL + LLP
  • 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?

    End-to-end training of the WSIs with tumor-positive/negative labels and tumor proportion labels leads to predictions of instance-level tumor labels, which appears to be better than training an MIL model followed by training a tumor proportions model

  • 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 #2

  • Please describe the contribution of the paper

    The paper proposes a weakly-supervised learning method for whole slide image segmentation, using bag-wise label proportion as supervision. The framework uses mi-Net as multiple instance learning predictor, and a multi-class classifier for learning the label proportion. Both networks are trained in an end-to-end fashion.

  • 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 overall solution for segmentation with partial label proportion is novel, using a multiple instance learning model to learn the binary mask first, and then using the multi-class classifier to learn the label proportion.
    2. The proposed network is trained in an end-to-end fashion.
    3. The label proportion loss is novel, using the MIL predicted binary mask to normalize the classification scores.
  • 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 the paper has no major weakness regarding novelty. The main issue is on the writing and details for reproducibility.

  • 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?

    Some of the details of the implementation is not clear:

    1. for the feature extractor, is it a single fixed resnet 18, or different modules trained separately for MIL and LLP?
    2. the details of the LLP classifier is not mentioned in the text.
    3. the superscripts and subscripts in the denotations are not clearly defined, and have to be inferred from the context, which might cause confusion.
  • 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

    Overall, the language of the whole paper should be further improved for readability purposes. For example, sentences such as “use them for realizing the segmentation model”, and “Assume (0.8, 0.2), “ are not clear or grammatically correct. We strongly recommend that the authors use a professional editing service to improve the clarity of their writing.

  • 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 novelty of the paper is clearly presented. But the language and the writing of the paper should be largely improved to meet the publication standards, and the implementation details should be thoroughly explained for reproducibility purposes.

  • 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

    The paper has defined the problem and found the solution for it. Especially for two-stage solving and mixing and combining MIL and LLP methods. They developed the new loss function to have better learning and have an improvement over related works in their working area.

  • 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 greatly split the problem into two parts and solved them together to have an end-to-end solution for the problem. They combined MIL and LLP methods to reach their final solution greatly. The authors have developed a new loss function that better finds the 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.

    For the evaluations, there are not enough comparisons with methods. The paper has results that are lower than the Oracle but better than a method that is not very comprehensive and thorough. There are some mistakes in the writing of the paper. For example, in the fig1 the “tumor” is written as “tumoe”. The gray arrow in the fig2 does not have informative information and the concept transfer to the reader is not very well in this part. In general, the description could be better and more aligned.

  • 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?

    The authors provided the details of the implementation of neural network configuration and dataset and provided the mathematical equations. But they have not mentioned any directed clue for reproducibility like code.

  • 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 author can compare more baselines with the current method to have a better comparison of their method and other related works.

  • 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 overall score for this paper comes from the idea that the authors have to combine those methods to make a novel solution regardless of their result in my opinion. They have creatively made a great idea for this problem.

  • 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

N/A



Meta-Review

Meta-review not available, early accepted paper.



back to top