List of Papers Browse by Subject Areas Author List
Abstract
Self-supervised pre-training paradigm has gained increasing prominence for learning transferable representations in medical imaging, yet existing methods for ultrasound (US) images operate at the image or frame level, overlooking the anatomical context for clinical-aligned representation learning. In this work, we propose an anatomy-anchored ultrasound Self-Supervision framework (ANAUS) that shifts representation learning from generic visual regions to clinically meaningful anatomical structures. Utilizing a learnable latent prompt engine alongside a one-time domain adaptation on existing public image-mask pairs, we empower the LP-SAM module to achieve annotation-free anatomy delineation at scale. Building upon this anatomical grounding, we propose a dual-policy self-supervised learning paradigm consisting of inter-view semantics-aware anatomy-separating alignment and contextual core-region prediction to enhance representation learning. Specifically, the former enforces feature invariance within identical anatomical regions while promoting discriminability across distinct structures; the latter compels the model to reconstruct corrupted regions, thereby capturing fine-grained structural details. Extensive evaluations on diverse public benchmarks demonstrate that ANAUS consistently outstrips current state-of-the-art methods while maintaining the computational efficiency essential for clinical deployment. Code is available at https://github.com/zhcz328/ANAUS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1980_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zhcz328/ANAUS
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhuChu_AnatomyAnchored_MICCAI2026,
author = { Zhu, Chunzheng AND Wang, Yijun AND Lin, Jianxin AND Wang, Feng AND Wang, Hongwei AND Zhao, Lei AND Li, Shengli AND Li, Kenli},
title = { { Anatomy-Anchored Self-Supervision: Distilling Vision Foundation Models for Invariant Ultrasound Representation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The AnaUS framework is introduced to shift representation learning from general visual regions toward clinically meaningful anatomical structures. LP-SAM, incorporating a learnable potential cueing engine, is also introduced.
- 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.The motivation of leveraging anatomical structures as anchor points is well-justified and forward-looking. 2.The experimental results are strong and cover a wide range of settings. 3.Strong generalization ability: As shown in Figure 4, LP-SAM is able to segment previously unseen anatomical structures during fine-tuning, enabling subsequent comparative learning.
- The availability of open-source code supports the reproducibility of the research.
- 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 training pipeline appears relatively complex. 2.There is room for more comprehensive analysis of hyperparameters.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
The article presents well-motivated objectives, a novel approach, and experimental results that demonstrate certain advantages.
- 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 #2
- Please describe the contribution of the paper
The main contribution of this paper is the proposal of AnaUS, an anatomy-anchored self-supervised learning framework for ultrasound representation learning. Instead of performing self-supervision at the image or generic region level, the method uses anatomically meaningful structures as the basic unit for representation learning. To achieve this, the authors introduce LP-SAM, a SAM-based module adapted to ultrasound via lightweight domain adaptation and a learnable latent prompt engine, enabling automatic anatomical mask generation for unlabeled images. Based on these masks, the paper designs a dual-objective pre-training strategy that combines anatomy-level contrastive alignment across views with contextual prediction over corrupted core regions. The paper further shows that this anatomy-aware pre-training improves downstream performance on both ultrasound classification and segmentation tasks across lung, breast, thyroid, and cardiac benchmarks, outperforming prior general SSL, object-centric SSL, and ultrasound-specific pre-training baselines in the reported experiments.
- 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 has several notable strengths. 1.A meaningful anatomy-guided SSL formulation. The paper reframes ultrasound self-supervised learning around anatomical structures rather than whole images or generic regions, and uses LP-SAM to provide structure-aware masks for unlabeled data. This is an interesting and practically useful way to inject anatomy priors into ultrasound representation learning. 2.Well-designed and complementary pre-training objectives. The pre-training strategy combines anatomy-level contrastive alignment with contextual prediction over corrupted core regions. These two objectives are complementary, encouraging both structure-aware invariance and finer anatomical understanding. The multi-scale design further strengthens the method. 3.A particularly interesting damage design for ultrasound-specific contextual prediction. The damage function used in the contextual prediction branch is a strong aspect of the paper. Instead of relying on generic masking alone, it introduces perturbations such as local pixel shuffling, Gaussian noise, and random occlusion to mimic realistic ultrasound artefacts, including tissue displacement under probe pressure, speckle interference, and acoustic shadowing. This design is interesting because it is not only technically reasonable but also grounded in the physical characteristics of ultrasound imaging, making the pretext task more clinically and modality relevant than standard generic corruption strategies. 4.Broad and fairly strong empirical evaluation. The paper evaluates the learned representations on multiple downstream tasks, including both classification and segmentation, across lung, breast, thyroid, and cardiac ultrasound benchmarks. It also compares against several categories of baselines, including general SSL, object-centric SSL, and ultrasound-specific pre-training methods. This breadth of evaluation makes the empirical section stronger than many works that validate only on a narrow downstream setting. 5.Useful ablations and diagnostic analysis. The paper goes beyond reporting final performance numbers by including ablations on the latent prompt engine, symmetric loss, multi-scale contrast, and contextual prediction. It also analyzes the effect of replacing generic mask sources with LP-SAM masks in object-centric baselines. This is valuable because it helps clarify which components are responsible for the observed gains and supports the central claim that anatomically grounded masks are important for ultrasound SSL.
- 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 paper has several important weaknesses. 1.I am not fully convinced that the “annotation-free” or “expert-free” claim is entirely appropriate. LP-SAM is first adapted using public image-mask pairs in a supervised stage before guiding SSL on unlabeled data. Therefore, while the SSL pre-training itself does not require manual annotation, the overall framework still relies on annotated segmentation data. This claim should be phrased more carefully. 2.It is not fully clear whether the gains come from the SSL formulation or mainly from better masks. The paper shows that simply replacing generic masks with LP-SAM masks already improves object-centric baselines substantially, and it also notes that precise anatomical masks may be a key factor in ultrasound pre-training quality. Therefore, it remains unclear how much of the gain comes from the proposed SSL objectives versus the mask generator itself. Stronger control experiments would help clarify this. 3.The comparison is not entirely fair because the proposed method benefits from extra supervised anatomical priors. AnaUS does not rely only on unlabeled ultrasound data; it also uses supervised LP-SAM adaptation on segmentation datasets such as DDTI, TG3K, and CAMUS. This provides additional structured supervision that many competing SSL baselines do not use in the same way. As a result, the comparison may not be fully apples-to-apples, especially when improvements are attributed mainly to the SSL framework itself. 4.Some methodological details are still under-explained. Some methodological details need clearer explanation. For example, Fig. 2 shows that the latent prompt refinement is repeated N times, and the implementation sets N=4, but this choice is not discussed. Similarly, the shuffle-based corruption with parameter ε is interesting and well motivated, but the paper only studies whether each corruption type is used or not. More sensitivity analysis, especially for the shuffle strength, would make this part more convincing. 5.The related-work positioning is not fully convincing. In the introduction, prior ultrasound SSL is mainly grouped into frame-level and region-level approaches, but some cited methods are not fully representative, and the categorization feels somewhat mixed. In particular, the discussion shifts between ultrasound-specific SSL and generic region/object-centric SSL from natural images, which weakens the clarity of the literature positioning. A more precise review of prior ultrasound SSL would strengthen the motivation. 6.The choice of pre-training data is insufficiently justified. The SSL pre-training is conducted mainly on the combined Butterfly + CAMUS dataset, but the paper does not explain clearly why these two sources were selected. It is unclear whether they were chosen because they are representative in terms of anatomy, acquisition diversity, or data scale, or simply because they were available. Since the paper makes broad claims about invariant ultrasound representation learning, the rationale for using only these two datasets should be better justified.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
My overall score reflects that, despite some limitations, I find this paper to make a meaningful and sufficiently strong contribution to ultrasound self-supervised learning. The anatomy-anchored formulation is well motivated, the overall framework is technically coherent, and the ultrasound-specific corruption design in the contextual prediction branch is particularly interesting. The paper also provides fairly broad evaluation across multiple downstream classification and segmentation tasks, with consistently strong empirical results. Overall, I believe the work is novel enough, practically relevant, and sufficiently complete for acceptance. The main reasons I did not rate it higher are related to claim calibration and experimental clarity. In particular, the “annotation-free” framing should be stated more carefully, since LP-SAM still relies on supervised adaptation with image–mask pairs. It is also not entirely clear how much of the gain comes from the proposed SSL objectives versus the improved LP-SAM masks, and some methodological choices could be better explained. However, these issues do not outweigh the overall contribution of the paper. In my view, they are limitations that should be clarified, but not major flaws that affect acceptability.
- Reviewer confidence
Very confident (4)
- [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 paper proposes an anatomy-anchored self-supervised learning framework for ultrasound representation learning. Instead of treating the whole frame or generic image regions as the contrastive unit, the method uses automatically discovered anatomical structures as the basis for representation learning. To enable this, the paper introduces LP-SAM, a lightly adapted SAM-based module with a latent prompt engine for annotation-free anatomy delineation in ultrasound. On top of this anatomical grounding, the method combines anatomy-level cross-view contrastive matching with a contextual prediction objective over overlapping core regions. The paper reports improvements over general SSL, object-centric, and several ultrasound-specific pretraining baselines across multiple downstream classification and segmentation benchmarks.
- 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.The paper addresses a meaningful and practical problem. Learning transferable ultrasound representations without dense manual annotation is important, and the attempt to ground SSL in anatomical structures is clinically better motivated than purely frame-level or generic region-level strategies. 2.The empirical evaluation is relatively broad. The paper reports results on multiple public downstream benchmarks spanning both classification and segmentation tasks, and the overall performance trends are consistently favorable. 3.The study includes ablation experiments that at least partially support the design choices, especially the importance of the LP-SAM component and the contribution of the different objective terms. 4.The idea of using anatomically guided regions rather than generic proposals is intuitively sensible for ultrasound, where boundaries are noisy and clinically meaningful structures matter more than arbitrary visual patches.
- 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 main weakness is clarity and presentation. The manuscript is harder to follow than it should be. The introduction is overly dense, the method section starts with notation before a clear verbal overview, and several core components are not explained in sufficiently direct language. As a result, the paper’s central idea is less clear than the authors likely intend. 2.The technical novelty appears moderate. Many ingredients, including SAM adaptation, prompt-based refinement, BYOL-style learning, contrastive matching, and contextual prediction, are existing ideas. The contribution seems to lie mainly in the anatomy-anchored combination and adaptation to ultrasound, rather than in a fundamentally new method component. 3.The paper does not sufficiently foreground what seems to be its key technical selling point, namely the latent-prompted anatomy discovery module. Although LP-SAM is emphasized in the contribution list and ablation study, its role, intuition, and distinction from more standard adaptation pipelines are not communicated sharply enough. 4.The writing contains presentation issues that reduce confidence and readability. For example, the baseline taxonomy is introduced somewhat abruptly, some terminology is insufficiently motivated, and there is at least one broken citation in the experimental section.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
I lean toward a weak accept. The paper addresses an important and practical problem in ultrasound representation learning and presents consistently strong empirical results across multiple public benchmarks, which makes the work relevant and potentially useful to the community. I also find the anatomy-anchored motivation reasonable and well aligned with the characteristics of ultrasound imaging, where clinically meaningful structures are often more informative than generic visual regions. The manuscript proposes a fairly complete framework, and the overall study is supported by broad experiments, ablation analyses, and generally solid figures and tables.
That said, I do have some reservations. The presentation is at times dense, and several parts of the method could be explained more directly and intuitively. In particular, the paper does not always highlight its key technical ideas as clearly as it could, and some components appear to build on existing techniques rather than introducing fundamentally new methodological elements. There are also some presentation issues, such as abrupt terminology and minor reference-formatting problems. However, I do not view these concerns as fatal. In my opinion, they mainly affect accessibility and sharpness of presentation rather than the overall validity or usefulness of the work.
Overall, although the paper could be written more clearly and the novelty is not absolute in every component, I believe the problem is important, the method is reasonably well motivated, and the empirical study is sufficiently strong and complete to justify acceptance at the weak-accept level.
- 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
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 submission proposes AnaUS, an anatomy-anchored self-supervised learning framework for ultrasound imaging. Three reviewers (R1, R2, R3) vote for weak accept. The paper presents a well-motivated and timely contribution by incorporating anatomical priors into self-supervised learning for ultrasound images, addressing a key limitation of existing frame-level approaches. The work is above the acceptance threshold and is suitable for acceptance. Several points should be addressed in the final manuscript. R1 points out that the training pipeline is relatively complex and lacks sufficient analysis. R2, It is currently unclear how much of the performance gain comes from the SSL objectives versus improved anatomical masks from LP-SAM. R3 highlights issues with clarity and presentation.