Abstract

Learning from sparse labels is a challenge commonplace in the medical domain. This is due to numerous factors, such as annotation cost, and is especially true for newly introduced tasks. When dense pixel-level annotations are needed, this becomes even more unfeasible. However, being able to learn from just a few annotations at the pixel-level, while extremely difficult and underutilized, can drive progress in studies where perfect annotations are not immediately available. This work tackles the challenge of learning the dense prediction task of keypoint localization from a few point annotations in the context of 2d carcinosis keypoint localization from laparoscopic video frames for diagnostic planning of advanced ovarian cancer patients. To enable this, we formulate the problem as a sparse heatmap regression from a few point annotations per image and propose a new loss function, called Crag and Tail loss, for efficient learning. Our proposed loss function effectively leverages positive sparse labels while minimizing the impact of false negatives or missed annotations. Through an extensive ablation study, we demonstrate the effectiveness of our approach in achieving accurate dense localization of carcinosis keypoints, highlighting its potential to advance research in scenarios where dense annotations are challenging to obtain.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4234_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZarFar_Learning_MICCAI2025,
        author = { Zarin, Farahdiba and Oliva, Riccardo and Srivastav, Vinkle and Vardazaryan, Armine and Rosati, Andrea and Zampolini Faustini, Alice and Scambia, Giovanni and Fagotti, Anna and Mascagni, Pietro and Padoy, Nicolas},
        title = { { Learning from Sparse Point Labels for Dense Carcinosis Localization in Advanced Ovarian Cancer Assessment } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15970},
        month = {September},

}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a novel loss function for the specific task of dense point localization from few point annotations. The authors specifically focus on the carcinosis keypoint localization from advanced stage ovarian cancer laparoscopic images (video frames). The proposed loss function (called Crag and Tail loss) is built on top of the Hill loss function which the authors adapted to their problem setting. Evaluated on their internal dataset, the proposed loss function outperforms the MSE and other variants of the Hill loss.

  • Please list the major strengths of the paper: you should highlight 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.

    This work falls under the CAI category (even though the authors seem to have gone with MIC) as evident from the dataset and the application. The solution proposed by the authors is also designed specifically to the problem the authors are trying to solve. And accordingly, the work achieves better results compared to the off the shelf approaches that use the MSE or the Hill loss functions. The qualitative results also look promising and given the clinical relevance, this is a good contribution. Additionally, the authors also include a ablation study for the proposed loss function which underscores the specificity of the proposed solution.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    First and foremost, there is a confusion in the name of the proposed solution. In the abstract, the authors call their method Crag and Tail but the rest of the paper goes with Crag and Hill which is inconsistent.

    Secondly, while the authors do a good job with the description of the problem formulation, the description for the Hill loss and the proposed modifications seemed lacking (at least to me). For example, it is not clear what H+ and H- actually represent. Similarly for the proposed baseline loss functions, there isn’t a clear justification for the authors choices in the 0.5maskedMSE and SoftUncertainRegionLoss designs. The main figure (Fig 2), while it gives a good idea of the high level working, it is hard to understand the concepts especially (for example, 6 color dots, 3 color squares, etc it is hard to keep track of the finer details).

    It is also unclear how the multilabel classification problem is forumulated. This isn’t mentioned in the method section as the point localization but it is weighed equally as the point localization problem in Table 1.

  • Please rate the clarity and organization of this paper

    Poor

  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    I recommend the authors to add more details to for the issues I listed under the weaknesses section if the page limit allows for it. Or at least try to modify the descriptions for a better reading experience.

  • 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.

    (3) Weak Reject — could be rejected, dependent on rebuttal

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

    While the problem setting is meaningful and the proposed solution seems to work well, the paper doesn’t describe the methodology to the expectation and there seem to be some oversights (eg. name of the method) and lack of clarity (eg. multilabel classification).

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The authors of the paper propose a novel loss function, Crag and Tail loss, that weights positive sparse labels and down-weights false negative labels for key point localization. The authors apply their loss function for carcinosis localization in ovarian laparoscopic video frames when only sparse training labels (instead of dense labels) are available. Crag and Tail loss shows improved point localization comparing with other common loss functions.

  • Please list the major strengths of the paper: you should highlight 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.

    Obtaining dense labels is impractical in the medical imaging domain. Crag and Tail loss can be significant when many false negative labels are present in a training set. In addition, this novel loss function may be extended beyond carcinosis localization in ovarian laparoscopic video frames.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    (1) The second evaluation scheme, multilabel classification, was unclear. If this means frame-level classification of presence or absence of carcinosis, then it would be binary classification. (2) Instead of a new term, (H-H_hat)^2, what if one simply gives a higher weight on the positive loss term?

  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    (1) On page 2, the authors mention two challenges: few positive labels and misleading incorrect labels. After reading the entire manuscript, it made sense. If the authors can elaborate more about the two challenges on page 2, it will help readability. (2) On page 3, there are two occasions where it is written as “Crag and Hill loss” that need to be edited. (3) On page 4, describe more on an HRNet model and why it is used. (4) In Table 2, subgrouping various changes may help readability.

  • 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.

    (4) Weak Accept — could be accepted, dependent on rebuttal

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

    The topic is significant in medical domain and the proposed loss function is novel. Additional clarifications for readability would be needed.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    The main contribution of the paper is the proposal of a novel loss function designed to estimate regions from keypoint annotations that may contain missing labels. Using a benchmark dataset, the method achieves higher accuracy compared to existing approaches.

  • Please list the major strengths of the paper: you should highlight 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.

    One major strength of the paper is the logical and mathematical development of a novel loss function that enables stable training of models to recognize regions based on keypoint annotations. This approach is particularly valuable as it helps reduce the annotation effort typically required for creating dense segmentation labels.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    One potential weakness of the paper is that, although the method is described as performing dense prediction, the output images from the model may appear to simply be enlarged keypoints. This could lead some to argue that the results do not fully reflect true dense region estimation.

  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • 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.

    (6) Strong Accept — must be accepted due to excellence

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

    I recommend this paper because the burden of region annotation is a common and significant challenge across many tasks. The proposed method offers a promising solution to this problem and is supported by sound mathematical reasoning.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A




Author Feedback

We thank the reviewers for their time and efforts in evaluation of our manuscript and for providing their valuable feedback. We appreciate the following being pointed out and attempt to elaborate upon them as follows:

  • Multilabel Classification (R1, R4): We are indeed performing binary classification for presence/absence of carcinosis. However, we do this not at frame level, but at the anatomical station level for the clinical task as detailed in the Introduction and in the Experimental Setup sections. To elaborate further upon this, the clinical task requires assessment of presence/absence of carcinosis in the 6 anatomical stations. These stations contain specific organs established in prior clinical works. For each frame analyzed, we take the outputs of our model which is post-processed to obtain the spatial coordinates of the carcinosis points. We then use the ground truth organ segmentation mask of the organs present in that frame to check if any carcinosis points are present inside the organ masks or not. If any carcinosis is found to be present, then the anatomical station that organ belongs to is positive for presence for carcinosis. This is performed for all 6 anatomical stations, resulting in independent yes/no for the multilabel classification task detecting presence of carcinosis. The addition of this multilabel metric is to highlight how tackling the problem of learning from sparse point annotations contributes to improving the given clinical task. Determining the specific organs affected by carcinosis is reliant on the point localization task, as each anatomical station in the image is checked for carcinosis points. Improvement of the point localization task thus results in improvement of the multilabel classification, as carcinosis detection improves, leading to improvements in identifying the anatomical stations affected.

  • Methodology Clarification (R1, R4): We thank the reviewers for mentioning the inconsistency in naming and will change all instances of ‘Crag and Hill’ to ‘Crag and Tail’. We built upon Hill loss for tackling our problem of learning from incorrect labels, where H+ are the logits rescaled for semi-hard mining positives, while H- are the default logits. While it is possible that other weighting mechanisms might also be successful, we opted for a component which adaptively weights difficult samples and have not explored the addition of another hyperparameter. However, we thank R1 for the suggestion and would like to expand upon this in future works. Our choice of HRNet is based on its success in heatmap regression in other domains, as high resolution feature maps are crucial for point localization. Since we limit our focus to tackling the learning from sparse points specifically with analysis of the hill loss, we limit the choice of backbone to HRNet.

  • Additional Proposed Losses (R4): The proposed 0.5maskedMSE and SoftUncertainRegionLoss are chosen to apply less weightage or Hill loss in estimated false negative regions only. The first loss is designed to showcase how manually placing less weightage in estimated false positive regions can improve results compared to using MSE as is, identifying the missing annotations as one of the key challenges in Table 1. The second loss places a combination of MSE+Hill for the same purpose and essentially adds Hill loss to 0.5maskedMSE with more weightage for Hill in potential false negative regions. This loss is to analyze how another potential combination of MSE+Hill may be beneficial to the task, like the naive combination of just adding MSE+Hill in Table 1.

We thank the reviewers R1, R3, R4 for finding our contribution in identifying and tackling the key challenges when learning from sparse annotations significant. We will focus on expanding upon certain sections to improve the clarity of the work as has been suggested.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    Although most reviewers mentioned certain lack of clarity in the method description, there seems to be a consensus that this is a good paper and should eventually be accepted. I am recommending acceptance, and I will trust the authors to take the advice of reviewers and improve the clarity of their paper in the points that the reviewers mentioned.



back to top