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

Xiaofeng Liu, Fangxu Xing, Nadya Shusharina, Ruth Lim, C.-C. Jay Kuo, Georges El Fakhri, Jonghye Woo

Abstract

Unsupervised domain adaptation (UDA) has been vastly explored to address domain shifts between source and target domains, by applying a well-performed model in an unlabeled target domain via supervision of a labeled source domain. Recent literature, however, has indicated that the performance is still far from satisfactory in the presence of significant domain shifts. Nonetheless, delineating a few target samples is usually manageable and particularly worthwhile, due to the substantial performance gain. Inspired by this, we aim to develop semi-supervised domain adaptation (SSDA) for medical image segmentation, which is largely underexplored. We, thus, propose to exploit both labeled source and target domain data, in addition to unlabeled target data in a unified manner. Specifically, we present a novel asymmetric co-training (ACT) framework to integrate these subsets and avoid the domination of the source domain data. Following a divide-and-conquer strategy, we explicitly decouple the label supervisions in SSDA into two asymmetric sub-tasks, including semi-supervised learning (SSL) and UDA, and leverage different knowledge from two segmentors to take into account the distinction of the source and target label supervisions. The knowledge learned in the two modules is then adaptively integrated with ACT, by iteratively teaching each other based on the confidence-aware pseudo-label. In addition, pseudo label noise is well-controlled with an exponential MixUp decay scheme for smooth propagation. Experiments on cross-modality brain tumor MRI segmentation tasks using the BraTS18 database showed, even with limited labeled target samples, ACT yielded marked improvements over UDA and state-of-the-art SSDA methods and approached an ``upper bound” of supervised joint training.

Link to paper

DOI: https://link.springer.com/chapter/10.1007/978-3-031-16443-9_7

SharedIt: https://rdcu.be/cVRyd

Link to the code repository

N/A

Link to the dataset(s)

N/A

Reviews

Review #1

  • Please describe the contribution of the paper
    • Introduces the problem of semi-supervised domain adaptation (SSDA) into medical image segmentation
    • Proposes a co-training framework that integrates UDA and SSL (semi-supervised learning)
    • The empirical results show that the proposed approach performs well on the BraTS2018 dataset.
  • 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 problem formulation of semi-supervised domain adaptation (SSDA) is highly relevant for medical imaging
    2. The proposed asymmetric co-training approach provides a framework for making use of both UDA and SSL
    3. The empirical results demonstrate ACT outperforms the baseline UDA and SSL 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.
    1. The methodology lacks motivation from a machine learning perspective. It is unclear how the proposed approach can appropriately integrate the two components given the different assumptions in UDA and SSL.
    2. The empirical evaluation is on one dataset. It would be more convincing to include more datasets in the evaluation to ensure the proposed approach does not overfit to the dataset.
    3. Confusion between ACT and ACT-EMD. If ACT includes EMD, what is ACT-EMD denoting.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • 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

    Code is 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/2022/en/REVIEWER-GUIDELINES.html

    I think the paper needs a stronger motivation for the proposed approach instead of simply stating this is what we propose and it works.

  • 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 paper has good empirical results but lacks justification in the approach relating to the underlying distribution assumptions.

  • Number of papers in your stack

    5

  • What is the ranking of this paper in your review stack?

    2

  • Reviewer confidence

    Somewhat 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 #2

  • Please describe the contribution of the paper

    This paper focuses on semi-supervised domain adaptation (SSDA) in medical image segmentation. It first discusses the unsatisfactory performance achieved by unsupervised domain adaptation (UDA). Inspired by the recent success of SSDA methods on natural images, this paper proposes a new SSDA framework that is tailored to medical image segmentation. The proposed method is termed asymmetric co-training (ACT). Different from traditional co-training methods that describe each example using two different sets of features, ACT decouples SSDA into semi-supervised learning (SSL) and UDA, then applies the co-training strategy. Furthermore, this paper proposes exponential mixUp decay (EMD) to reduce the noise in generated pseudo labels. The proposed framework is evaluated on the BraTS18 database and outperforms previous UDA and SSDA methods by a large margin.

  • 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 idea of asymmetric co-training is interesting and novel. Rather than describe each example using two different sets of features that provide complementary information about the instance, this paper decouples the source data and the target data into two sets which are then fed into an SSL module and a UDA module to achieve the co-training.

  • 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. This paper mentions that previous SSDA methods suffer from “source domain supervision dominates the training”, is any reference or analysis to support this statement?
    2. What is the detailed workflow of SSL and UDA? There are lots of implementations of these two tasks, but this paper does not provide more details. Therefore, it is not clear whether the improvement comes from the new framework or more advanced SSL or UDA implementations.
    3. The meaning of SSDA:1 or SSDA:5 is not clear. Does it mean the number of labeled subjects (as claimed in the second paragraph of Section 3) or the number of samples (as claimed in the fifth sentence of the first paragraph on page 7) in the target domain? What does the “subject” mean?
    4. Since the labeled target samples are quite limited in SSDA: 1 and SSDA: 5, is that enough to train a model? Even so, the generated pseudo labels would be quite noisy which may result in negative transfer problems.
    5. Do other baseline methods share the same labeled target examples with ACT in the SSDA: 1 and SSDA: 5 settings?
    6. This paper is not well written. There are a lot of long sentences which are hard to understand. For example, “In order to prevent a segmentor, jointly trained by both domains, from being dominated by the source data only, we adopt a divide-and-conquer strategy to decouple the label supervisions for the two asymmetric segmentors, which share the same objective of carrying out a decent segmentation performance for the unlabeled data”
  • Please rate the clarity and organization of this paper

    Satisfactory

  • 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 manuscript misses some technical details, for example, the implementations of SSL and UDA in the proposed ACT framework. Fortunately, the code is 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/2022/en/REVIEWER-GUIDELINES.html
    1. The writing needs to be substantially improved. A suggestion is to avoid using very long sentences which are unnecessary in most cases.
    2. Page 6: change “Without loss generality” to “Without loss of generality”
    3. Page 7: “ACT-EMD” in “The better performance of ACT over ACT-EMD demonstrated the effectiveness of our EMD scheme for smooth adaptation with pseudo-label” is confusing. Suggest using ACT (w/o EMD).
    4. It is suggested to briefly introduce “Target Only” and “Supervised Joint Training ” in Table 1 and 2. It is not clear how they are implemented. For example, what is training data used in these two settings?
      Table 1: change “SSAD” in Table 1 to “SSDA”
  • 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?

    This paper proposes an interesting and novel idea to apply the co-training strategy to the SSDA problem. The performance is impressive. My concerns are (i) the experimental settings need more justification, (ii) some important technical details are missing in the manuscript, and (iii) the writing is not clear and needs to be improved.

  • Number of papers in your stack

    7

  • What is the ranking of this paper in your review stack?

    2

  • 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 authors propose an asymmetric co-training~(ACT) strategy to decouple labeled data into SSL and UDA for semi-supervised domain adaptation (SSDA). Two models are trained with labels from different domains and then boosted together with pseudo labels generated from each other. And exponential MixUp decay is proposed for gradual co-training. They validate the effectiveness of the proposed method on cross-modality brain tumor segmentation tasks, outperforming other SSDA and UDA 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 idea is novel and works well. It is reasonable to generate pseudo labels for each other and boost each other for SSDA.

    2. The methodology is easy to follow. Gradual co-training is suitable and helpful for asymmetric co-training.

    3. The experiments are well-designed, which demonstrates the effectiveness of ACT. Ablation studies and sensitivity analysis can help further evaluate the feasibility of decoupling labels to SSL and UDA for SSDA.

  • 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. Unlike the statement “no SSDA for medical image analysis (MedIA)”There are many works about SSDA for MedIA [1]. And decoupling labels to different parts, such as SSL and UDA, is not new to SSDA, e.g., [2] [3]. The authors should discuss more recent SSDA works in the introduction section.

    2. Accordingly, the authors should compare the proposed method with more recent UDA and SSDA works such as [2-4], besides these UDA and SSDA methods for natural images.

    [1] Guan H, Liu M. Domain adaptation for medical image analysis: a survey[J]. IEEE Transactions on Biomedical Engineering, 2021. [2] Zhao Z, Xu K, Li S, et al. MT-UDA: Towards Unsupervised Cross-modality Medical Image Segmentation with Limited Source Labels[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2021: 293-303. [3] Li K, Wang S, Yu L, et al. Dual-teacher++: Exploiting intra-domain and inter-domain knowledge with reliable transfer for cardiac segmentation[J]. IEEE Transactions on Medical Imaging, 2020, 40(10): 2771-2782. [4] Chen C, Dou Q, Chen H, et al. Unsupervised bidirectional cross-modality adaptation via deeply synergistic image and feature alignment for medical image segmentation[J]. IEEE transactions on medical imaging, 2020, 39(7): 2494-2505.

  • 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 author submitted the code as the support material. There are no reproducibility concerns.

  • 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/2022/en/REVIEWER-GUIDELINES.html

    1. The authors should discuss more recent UDA and SSDA works for medical image segmentation, and compare with these methods in the experiment section to further prove the effectiveness of the proposed method.

    2. More cross-modality datasets, such as MM-WHS and Chaos, can be implemented for further evaluation.

  • 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 idea of the divide-and-conquer strategy for SSDA is well-motivated. The methodology and experiments are well-designed, and the results are promising for cross-modality segmentation. I highly suggest that the authors include more discussions on recent UDA and SSDA methods. Also, the authors should give more comparisons with these methods.

  • Number of papers in your stack

    4

  • What is the ranking of this paper in your review stack?

    1

  • 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 received consistently positive comments on the idea of integrating unsupervised domain adaptation (UDA) and semi-supervised learning (SSL) for asymmetric co-training. However, the reviewers also pointed out that the clarity of the paper and the writing need to be significantly improved, e.g., the motivation from an algorithm viewpoint should be clearly explained, technical details should be provided (such as implementation of UDA and SSL modules), experimental setup and result interpretation should be described in detail (see comments from Reviewer #2), and recent relevant work should be discussed. Please improve the manuscript by considering the reviewers’ comments when preparing the final version.

  • What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    2



Author Feedback

N/A



back to top