List of Papers Browse by Subject Areas Author List
Abstract
Establishing accurate point-wise correspondences between different images is crucial for various tasks in medical image computing, such as motion tracking, shape analysis and spatial normalization. Existing approaches mainly rely on deformable registration or (semi-) supervised landmark detection and typically require dedicated training or annotated correspondences. In this work, we propose a training-free framework for correspondence estimation that leverages internal representations from pretrained segmentation networks. We hypothesize that deep learning segmentation models implicitly encode rich anatomical structure within their feature maps. By computing voxel-wise similarities in this feature space, interpretable point-wise correspondences are obtained without need for additional training or landmark annotations. We evaluate our approach and the impact of design options, such as the granularity of the segmentation, the choice of feature maps and the similarity measure, for one-shot landmark detection in different applications involving different anatomies and modalities, demonstrating competitive performance in terms of mean target registration error (mean TRE: intra-subject cardiac MRI: 2.85mm; intra-subject lung CT: 2.76mm; inter-subject brain MRI: 3.08mm). We further illustrate the potential of the method for joint estimation of coupled landmarks by incorporating geometric constraints and for unsupervised salient keypoint discovery across a sequence of images. Our results show that accurate, interpretable, label- and training-efficient correspondence estimation is feasible by repurposing segmentation-derived representations in a generic manner.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5323_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{DuvEli_Uncovering_MICCAI2026,
author = { Duval, Eli AND Maes, Frederik},
title = { { Uncovering Interpretable Point-Wise Image Correspondences from Deep Segmentation Features } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents an approach to detecting anatomical landmarks based on the activations of a segmentation network. Specifically, the authors evaluate the similarity between pixels of two images by computing similarity between their multi-channel, multi-layer representations within a segmentation network. When normalized to sum to one, such a map is treated as a probability distribution over all pixels in one image as potential matches to a selected pixel in another image. The authors also present an extension that detects distinct landmarks in a set of images via “chain rule” that combines pairwise similarity maps. An empirical evaluation shows that the proposed method outperforms baseline registration methods in 3 different modalities: cardiac MRI, Lung CT, and Brain MRI.
- 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.
The paper is well written and the method is clearly explained. The empirical study shows the advantages of combining the features across layers.
- 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.
The authors present the method as if establishing sparse correspondences is the final goal. But I can’t think of a setting where it is. Instead, the correspondences are used to answer clinically relevant questions. What are these? What motivated this development? The paper feels unfinished in the present form as it does not articulate the use of correspondences nor tests the proposed method in such a context. One could imagine using correspondences to drive registration, but then the evaluation would focus on alignment of organs, not just distances between estimated matching landmarks. The comparison to registration methods is incomplete – many numbers in the table are missing for the registration methods.
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Interesting idea, but is not fully developed to demonstrate potential uses that lead to better image analysis.
- 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
Authors propose a point-wise correspondence estimation method used for one-shot landmark detection in pairs of images and zero-shot keypoint detection in image series. They use multi-scale features - from UNet networks trained on a segmentation task for a specific anatomy and modality - to measure the similarity of point pairs belonging to two images subsequently producing a similarity map representing correspondence likelihoods. The method is evaluated on cardiac MRI, lung CT and brain MRI showing competitive performance with respect to the state-of-the-art.
- 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.
The proposed method has three key strengths : 1.The straightforward probabilistic formulation, in practice only requiring the definition of a distance metric and an interpolation method. 2.It can leverage feature maps of pre-trained convolutional networks without any training or annotated data.
- Correspondence likelihood distributions allow interpretability and uncertainty analysis.
- 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.
A minor drawback is perhaps the lack of comparison with a state-of-the-art method on cardiac MRI.
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method and the findings are of good interest for the MICCAI readership and deserve to be brought to the attention of the community.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Additional clarifications brought by the authors’ comments solidify the acceptance decision I recommended earlier.
Review #3
- Please describe the contribution of the paper
This paper proposes a training-free framework for correspondence estimation by leveraging internal feature representations from pretrained segmentation networks. The central hypothesis is that segmentation models implicitly encode rich anatomical structure, which can be exploited to compute voxel-wise similarity in feature space for one-shot landmark detection.
- 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.
1- Relevant and important problem: The paper addresses the challenging task of anatomical landmark detection, which is highly relevant to many medical imaging applications.
2-Interesting and novel direction: The idea of leveraging pretrained segmentation networks in a training-free manner for landmark detection is conceptually appealing and offers a potentially efficient alternative to fully supervised approaches.
3- Clarity of presentation: The paper is generally well-written and easy to follow, with a clear explanation of the proposed framework.
4-Promising empirical results: The results are competitive and supported by ablation studies that help validate different components of the method.
- 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-Incomplete evaluation metrics: The evaluation would benefit from reporting Success Detection Rate (SDR) at standard thresholds (e.g., 2 mm, 3 mm, 6 mm), which is a commonly used metric in landmark detection. This would provide a clearer picture of whether predictions fall within clinically acceptable error margins.
2- Overstated claims on landmark discovery: The paper suggests that the method enables anatomical landmark discovery; however, this claim is not sufficiently validated. While qualitative visualizations are provided, they are not enough to support such a strong claim. More rigorous validation is needed, for example: a) Anatomical significance: evaluating whether discovered landmarks correlate with clinically meaningful variables (e.g., disease classification or clinical scores). b) Consistency analysis: assessing stability across subjects (e.g., mapping landmarks to a common template and measuring inter-subject variability or inter-landmark distances). Without such analyses, the claim of “landmark discovery” appears somewhat overstated.
3- Figure quality and presentation: Figure 1 could be improved. The ground truth and predicted landmarks are difficult to visualize due to their small size, and the figure occupies excessive space relative to its informational content. A more compact and clearer visualization would improve readability.
4-Method illustration: The methodology section would benefit from a more informative schematic or pipeline diagram to better convey the workflow and key components of the approach.
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, this paper presents a novel and well-motivated framework for training-free one-shot landmark detection by leveraging pretrained segmentation features. The idea is interesting, and the empirical results are promising across multiple datasets.
However, the evaluation could be strengthened by including standard metrics such as SDR, and the claims regarding landmark discovery require more rigorous validation beyond qualitative visualization. Additionally, improvements in figure clarity and methodological illustration would enhance the presentation.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reviewing the rebuttal alongside the other reviewer comments and responses, I still consider the submitted work worthy of publication, as it addresses an interesting topic of relevance to the community.
Author Feedback
We thank the reviewers and the area chair for their valuable and insightful feedback, which we take to heart to strengthen our work further. Below we address the main concerns:
End goal of sparse correspondences (R1): We agree that establishing sparse correspondences is not usually the end goal in medical image analysis, but rather enables a range of down-stream tasks in various applications, such as intra- and inter-subject registration, segmentation, shape and motion analysis. This broad purpose is our motivation to develop a generic method for correspondence detection in a training-free and interpretable way, which is the main contribution of our paper. To assess the potential of our approach, we perform quantitative evaluation at the voxel-level using annotated landmarks, rather than on downstream task performance since this would introduce task-specific modelling choices beyond the scope of this work. To illustrate the downstream applicability of our proposed framework, we have included a proof-of-concept on cardiac wall thickness tracking across the cardiac cycle (Figure 1c). We will clarify the contribution of our work more explicitly in the paper.
Missing numbers in Table 1 (R1): We understand that Table 1 with its current design may seem incomplete and we will remove this confusion from the paper. Table 1 is actually not missing values: we report only results from the original, peer‑reviewed publications as evaluated on the same datasets, landmarks and metrics. Rather than providing an exhaustive benchmarking, the purpose thereof is to contextualize our proposed approach with respect to representative state-of-the-art methods in both registration and landmark detection. What constitutes state-of-the-art is application dependent: landmark detection methods for instance require a fixed set of predefined landmarks across subjects and are therefore not directly applicable to the lung CT dataset with its highly patient-specific correspondences. We will better motivate the scope and intent of the comparison in the paper.
Lack of comparison with a state-of-the-art method on cardiac MRI (R3): To provide a more complete comparison in the paper, we will compare our results on cardiac MRI with the supervised landmark detection results reported in the original landmarks’ dataset publication ([12]).
Success Detection Rate (SDR) as evaluation metric (R4): We agree that SDR is a valuable metric and will include it in the empirical study.
Claim on landmark discovery (R4): We would like to clarify that the paper does not claim anatomical landmark discovery in an unsupervised manner. We explicitly distinguish between two tasks in the paper: one-shot landmark detection and unsupervised keypoint discovery. The experiments on the one-shot task quantitatively confirm that pre-defined anatomical landmarks can be consistently detected across subjects, modalities and landmarks. In contrast, the unsupervised keypoint discovery experiment (Figure 1de) does not aim to validate anatomical landmarks: instead, it aims to identify stable correspondences across a sequence of images, which reflect salient image features but not anatomical landmarks necessarily. We include this experiment as a proof‑of‑concept to illustrate the interpretable uncertainty information provided by the proposed framework. To prevent this ambiguity in the paper, we will emphasize the exploratory nature of this experiment and avoid terminology that could be interpreted as anatomical landmark discovery.
Figures (R4): We agree that Figure 1 could be improved for clarity and we will increase the visibility of the ground truth and predicted landmarks. At the same time, we want to keep a visual example for each application with their interpretable probability maps, since they are essential to the proposed framework. Moreover, as suggested by the reviewer, we will include a schematic diagram of the methodology of framework for a better understanding.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. 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 presents a novel and clear methodological contribution for anatomical landmark detection using pretrained segmentation features. The idea of leveraging multi-scale feature representations in a training-free manner is conceptually appealing, and the unsupervised key point detection enhances interpretability via probabilistic correspondence maps. Empirical results across multiple modalities (cardiac MRI, lung CT, brain MRI) are promising and indicate competitive performance.
However, several important concerns are raised.
- A primary concern is the seemingly missing values in Table 1, which is incomplete.
- Another key limitation is the lack of clear practical grounding: the paper frames correspondence estimation as the end goal, but does not convincingly demonstrate downstream utility (e.g., for registration or clinical tasks), leaving the contribution somewhat incomplete. Relatedly, the claim of landmark discovery is not sufficiently validated and should be complemented with quantitative analysis.
- Third, reproducibility seems to be improvable.
- Finally, I agree with reviewer 3 that this paper could gain in clarity by introducing an overview figure.
Overall assessment: While the good initial ratings highlight the quality of this paper, I cannot recommend acceptance for a paper with missing results. Thus I recommend rebuttal, which will also give the opportunity to the authors to clarify some concerns about practical grounding for downstream tasks, reproducibility, and softening of the “landmark discovery” claim.
- 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.
The authors have satisfyingly answered most concerns raised by the reviewers, particularly they have justified why Table 1 might seem incomplete (and will provide clarifications in the revised paper). Moreover, I appreciate the addition of another metric, another baselines for cardiac applications, and the promise to revise figures to improve clarity. Therefore, I will side with all reviewers and recommend acceptance for this paper.
Meta-review #2
- 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.
This paper received positive recommendations both before and after the rebuttal. The reviewers acknowledged the novelty and insightfulness of the work. Although some limitations in the evaluation were noted, particularly by Reviewers 1 and 3, the overall strengths of the paper outweigh these weaknesses.
In light of the post-rebuttal discussion and the overall reviewer feedback, AC therefore recommend a clear accept.
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.
The rebuttal has addressed the major concerns of the reviewers. The proposed method is promising and demonstrates applicability in an arrange of medical scans.