List of Papers Browse by Subject Areas Author List
Abstract
Recent unsupervised domain adaptation methods in medical image segmentation adopt centroid/prototypical contrastive learning (CL) to match the source and target features for their excellent ability of representation learning and semantic feature alignment. Of these CL methods, most works extract features with a binary mask generated by similarity measure or thresholding the prediction. However, this hard-threshold (HT) strategy may induce sparse features and incorrect label assignments. Conversely, while the soft-labeling technique has proven effective in addressing the limitations of the HT strategy by assigning importance factors to pixel features, it remains unexplored in CL algorithms. Thus, in this work, we present a novel CL approach leveraging soft pseudo labels for category-wise target centroid generation, complemented by a reversed Monte Carlo method to achieve a more compact target feature space. Additionally, we propose a centroid norm regularizer as an extra magnitude constraint to bolster the model’s robustness. Extensive experiments and ablation studies on two cardiac data sets underscore the effectiveness of each component and reveal a significant enhancement in segmentation results in Dice Similarity Score and Hausdorff Distance 95 compared with a wide range of state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1593_paper.pdf
SharedIt Link: pending
SpringerLink (DOI): pending
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/1593_supp.pdf
Link to the Code Repository
https://github.com/MingxuanGu/Soft-Labeled-Contrastive-Learning
Link to the Dataset(s)
https://zmiclab.github.io/zxh/0/mscmrseg19/
https://github.com/FupingWu90/CT_MR_2D_Dataset_DA
BibTex
@InProceedings{Gu_Unsupervised_MICCAI2024,
author = { Gu, Mingxuan and Thies, Mareike and Mei, Siyuan and Wagner, Fabian and Fan, Mingcheng and Sun, Yipeng and Pan, Zhaoya and Vesal, Sulaiman and Kosti, Ronak and Possart, Dennis and Utz, Jonas and Maier, Andreas},
title = { { Unsupervised Domain Adaptation using Soft-Labeled Contrastive Learning with Reversed Monte Carlo Method for Cardiac 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 paper proposes a novel approach to central to central competitive learning by introducing a soft labeling strategy (SLCL) to address misclassification and sparse feature space issues. It also presents a reversed Monte Carlo method (rMC) to achieve a more compact target feature space for improved contrasting learning. Additionally, a centroid norm regulator (CNR) is introduced to ensure consistency between source and target features. The article offers valuable contributions to the field, with experimental results supporting the effectiveness of the proposed techniques in enhancing classification accuracy, efficiency, and alignment during feature mapping.
- 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 reviewed article highlights three key points:
1: Soft Labeling Strategy: This paper converts the binary classification problem into a regression problem by putting soft importance weights on the features, smoothing the calculation of pseudo-labels, which is intuitive
- Reversed Monte Carlo Method: when performing C2C contrastive learning, this paper uses the Reversed Monte Carlo method to provide sufficient gradients for the network to further update pixel features, ensuring steady training of the network and increasing the generalization of the model.
- Centroid Norm Regularizer: the author proposes the Centroid Norm Regularizer to ensure the alignment of the source and target feature spaces by regularizing the size of the target centroid.
- 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 performance of the evaluated method is relatively weaker when compared to the state-of-the-art (SOTA) methods listed. Furthermore, it is essential to include comparisons with more recent top-tier methods to provide a comprehensive evaluation and analysis.
- The manuscript appears to be in draft form, as it leaves room for the inclusion of author names and acknowledgements.
- he author proposes the Centroid Norm Regularizer to ensure the alignment of the source and target feature spaces by regularizing the size of the target centroid.
- The paper mainly focuses on performance improvement, but there is not much discussion on the interpretability of the model or how to interpret the impact of soft labels and the reversed Monte Carlo method on the feature space.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
please refer to the weakness part
- 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 Reject — could be rejected, dependent on rebuttal (3)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The performance is not strong enough, and some details are missing.
- 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
Weak Reject — could be rejected, dependent on rebuttal (3)
- [Post rebuttal] Please justify your decision
I’m referring to comparing against more recent methods, since the latest work included for comparison in Tables 1 and 2 was published in 2022.
Review #2
- Please describe the contribution of the paper
The paper proposes a soft-labeling strategy to alleviate the misclassification and sparse feature space in the conventional C2P contrastive learning with a hard-threshold strategy. They also introduce a reversed Monte Carlo method for a more compact target feature space. A centroid norm regularization term is also utilized to force the magnitudes of the source and target features to be consistent.
- 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 paper demonstrates strong writing, organization, and structure. It effectively elucidates the motivation behind the research problem, providing a clear rationale. Furthermore, it tackles a contemporary challenge in contrastive learning. The proposed solution introduces innovative solution utilizing methods such as soft-labeling, regularization, and the reversed Monte Carlo.
- 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 ablation study experiments didn’t thoroughly investigate how the soft-labeling technique affects the final output. Similarly, the influence of the centroid norm regularizer on the final output requires deeper exploration.
In Figure 4-a, the effect of employing multiple partitions on prediction results is depicted. Notably, the optimal count of partitions appears to be 2, prompting inquiry into the true utility of utilizing more partitions. Further investigation is necessary to discern whether additional external factors render partitioning beyond this point ineffective.
- 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
The authors could enhance the ablation study by delving deeper into the impact of the soft-labeling technique on the final output. Likewise, more extensive exploration of the influence of the centroid norm regularizer on the final output would strengthen the analysis.
- 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 paper tackles a contemporary challenge in contrastive learning. The proposed solution introduces innovative solution utilizing machine learning techniques.
- 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
To circumvent the drawbacks of the hard-threshold label strategy in contrastive learning-based UDA methods, this paper proposed a novel approach based on soft pseudo labels with a reversed Monte Carlo method.Extensive experiments and ablation studies on two cardiac data sets were conducted.
- 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.
A novel UDA approach based on centroid-to-centroid contrastive learning with a soft-labeling strategy was proposed and a corresponding centroid norm regularizer was investigated as an extra magnitude constraint to bolster the robustness of proposed model.
- 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 experiments can’t support the title.
- Please rate the clarity and organization of this paper
Satisfactory
- 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
The idea of category-wise centroid CL with reversed Monte Carlo is novel and interesting. One point is confusing,i.e., is the proposed model designed for cardiac segmentation or general medical image segmentation? If it is a task-specific model, the task of cardiac segmentation has not been analysed in the Introduction and Method; if it is a common model, the performance on other segmentation tasks has not been evaluated.
- 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 title indicates that the new method is designed for common segmentation tasks and the authors have not analysed the relation between the characteristics of cardiac segmentation and the proposed method, hence my recommendation is weak 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
My problem has been solved.
Author Feedback
We deeply appreciate the thoughtful and insightful comments provided by the reviewers. Your expertise and attention to detail have greatly enriched our work. In this work, we proposed a novel soft-labeled contrastive learning (SLCL) approach for unsupervised domain adaptation (UDA) tasks. The proposed method alleviates the sparse feature space and incorrect label assignment of current hard-threshold (HT) contrastive learning (CL) solutions. We showed the effectiveness of each proposed component and a significant improvement compared with the other SOTA methods on two cardiac data sets. We are encouraged that the reviewers found our work intuitive (R1) and innovative (R1, R4, R6). We appreciate their commendation of our writing skills and clarity of expression (R4). We are glad they found the work’s open-source spirit (R1). We are pleased that the experiments and results are positively assessed (R6). Top-tier models (R1): Experiments with top-tier models are exciting but fall outside the scope of this paper. Since R1 does not specify, we assume it refers to Transformer or diffusion models. These architectures demand extensive data due to their complexity and high parameter count. However, medical image segmentation often lacks sufficient training data. Although collecting data from multiple data sets is possible, it conflicts with our objective of UDA. Most comparison methods utilized Unet-like architectures or Deeplabs. To keep the consistency in the model architecture and make our results directly comparable, we find it practical to retain the results from experiments with the Unet architecture. Task specification (R6): We acknowledge the omission of the specific objective in the paper, which is designed for cardiac image segmentation. We will change the title to “Unsupervised Domain Adaptation using Soft-Labeled Contrastive Learning with Reversed Monte Carlo Method for Cardiac Image Segmentation”. We will revise the first sentence of the Introduction to: “Accurate cardiac segmentation is crucial for various medical applications. In clinical settings, multi-modality medical images are extensively utilized to aid diagnosis. However, automatic cardiac image segmentation often suffers from performance degradation due to a lack of labels.” We will analyze the connection between cardiac image segmentation and the proposed method and include the following paragraph in the Method section: “The HT criteria used in current CL methods may introduce incorrect pseudo-labels, resulting in scattered classification in cardiac image segmentation and inaccurate diagnosis in clinical applications. In contrast, the soft-labeling (SL) strategy smooths the impact of ambiguous pseudo-labels and conducts better consistency within each category of the cardiac anatomy (Fig. 3).” Investigation of SL and centroid norm regularizer (CNR) (R1 & R4) In Fig. 4b, we presented extensive parameter studies for SL and CNR. We compared the performance of the proposed method under SL and HT strategies with different thresholds (0.0 to 0.9), demonstrating the robustness of SLCL and interpreting the trends in Section 3.2. Additionally, we examined SLCL with and without CNR, showing consistent improvement with CNR across various thresholds. Impact of the reversed Monte Carlo Method (rMC) (R1 & R4) We did abundant experiments (Fig. 4a) to show the impact of rMC, ranging the partition number P from 1 to 512. In Section 3.1, we interpreted the value of P as a trade-off between centroid stability (P=1) and feature compactness (P=512). The ablation study with the t-SNE plot (Fig. 5) for rMC shows more compact feature space and better class separation for SLCL with rMC, which confirms the interpretation and experiment results above. Due to the rebuttal rules, we are not allowed to add more experiments, results, or analyses. Nevertheless, we deeply appreciate all the comments regarding further interpretation and investigation.
Meta-Review
Meta-review #1
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This paper presents a novel contrastive learning approach leveraging soft pseudo labels for category-wise target centroid generation, complemented by a reversed Monte Carlo method to achieve a more compact target feature space. The reviewers acknowledge the technical contributionand the effectiveness of the proposed method. However, they raised significant concerns about experiments, investigation of SL & centroid norm regularizer, task specification, and impact of the reversed Monte Carlo Method. After rebuttal, the main concerns can be addressed. Therefore, the submission is considered acceptable for publication.
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/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).
This paper presents a novel contrastive learning approach leveraging soft pseudo labels for category-wise target centroid generation, complemented by a reversed Monte Carlo method to achieve a more compact target feature space. The reviewers acknowledge the technical contributionand the effectiveness of the proposed method. However, they raised significant concerns about experiments, investigation of SL & centroid norm regularizer, task specification, and impact of the reversed Monte Carlo Method. After rebuttal, the main concerns can be addressed. Therefore, the submission is considered acceptable for publication.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This paper received 2 positive scores and 1 negative score from the reviewers. The work has some merits, however, the performance improvements compared to state-of-the-art methods are too limited, with only a 0.002 increase reported in Table 2.
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/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).
This paper received 2 positive scores and 1 negative score from the reviewers. The work has some merits, however, the performance improvements compared to state-of-the-art methods are too limited, with only a 0.002 increase reported in Table 2.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The paper receives borderline recommendation, with one weak reject and two weak accept. The reviewers all recognize the novelty of the proposed method, but have concerns on the experimental evaluation (R1: the latest work included for comparison in Tables 1 and 2 was published in 2022. R4: Enhance the ablation study by delving deeper into the impact of the soft-labeling technique and the centroid norm regularizer. ) I would encourage the authors to seriously consider the comments from the reviewers and incorporate them into the final version.
my major concern is also with the experimental evaluation of the paper.
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/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).
The paper receives borderline recommendation, with one weak reject and two weak accept. The reviewers all recognize the novelty of the proposed method, but have concerns on the experimental evaluation (R1: the latest work included for comparison in Tables 1 and 2 was published in 2022. R4: Enhance the ablation study by delving deeper into the impact of the soft-labeling technique and the centroid norm regularizer. ) I would encourage the authors to seriously consider the comments from the reviewers and incorporate them into the final version.
my major concern is also with the experimental evaluation of the paper.