List of Papers Browse by Subject Areas Author List
Abstract
Graph-based cardiac segmentation with implicit anatomical correspondences provides topological guarantees and population-level analysis capabilities, but models trained on independent frames of image sequences exhibit temporal discontinuities that affect reliable clinical measurements, particularly in cardiac ultrasound. In this work, we introduce self-supervised temporal regularization as a post-training refinement stage that exploits the temporal coherence in image sequences to enforce consistent cardiac segmentation and motion estimation over time, without requiring per-frame annotations. By penalizing velocity and acceleration discontinuities across consecutive frames, our method achieves temporally consistent segmentations while maintaining the learned anatomical correspondences. We further leverage these correspondences to automatically map landmarks to the AHA 17-segment clinical standard, enabling standardized regional assessment and detection of pathological myocardial motion patterns. Validation on CAMUS dataset demonstrates the clinical utility of combining temporal consistency with automatic regional mapping. The code is publicly available at https://github.com/david-montalvoo/MaskHybridGNet-TempReg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5651_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/5651_supp.zip
Link to the Code Repository
https://github.com/david-montalvoo/MaskHybridGNet-TempReg
Link to the Dataset(s)
N/A
BibTex
@InProceedings{MonDav_Selfsupervised_MICCAI2026,
author = { Montalvo-García, David AND Gaggion, Nicolás AND Ledesma-Carbayo, María J. AND Ferrante, Enzo},
title = { { Self-supervised Temporal Regularization for Landmark-Based Cardiac Segmentation with Automatic AHA Regional Mapping } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents work based on the segmentation algorithm proposed in Reference [6]. A self-supervised temporal regularization is introduced as a post-training refinement stage that uses temporal coherence in image sequences to ensure consistent cardiac segmentation and motion estimation over time without requiring annotations for each frame. The proposed method produces temporally consistent segmentations while preserving learned anatomical correspondences. These correspondences are then leveraged to automatically map landmarks to the AHA’s 17-segment clinical standard, which enables a standardized assessment of regional cardiac function.
- 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 algorithm is self-supervised.
- An AHA-based motion computation is obtained.
- 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.
- Temporal regularization appears to be limited to segmentations obtained using the method proposed in Reference [6].
- The assessment of temporal regularization is based on criteria defined by the authors.
- 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
- The following related publications are missing : 1) G. Tziritas, EigenBoundaries for Temporally Regularized Segmentation of Echocardiographic Images, Workshop on Statistical Atlases and Computational Models of the Heart, 2024. 2) Wu, H. , Lin, J. , Xie, W. , Qin, J. , Super-efficient echocardiography video segmentation via proxy- and kernel-based semi-supervised learning, 37th AAAI Conference on Artificial Intelligence, 2023.
- The TED dataset (https://humanheart-project. creatis. insa-lyon. fr/databases. html) contains annotated cardiac ultrasound sequences that cover a complete cardiac cycle. The temporal regularization can be evaluated precisely on this dataset.
- I’m not sure if the comparison to nnUnet is fair.
- Please check the typo of Reference [14].
- 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 proposed method is closely tied to the work referenced in [6], and its assessment is somewhat restricted.
- 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.
Since the authors addressed my comments and replied to the other reviewers’ comments, acceptance is recommended.
Review #2
- Please describe the contribution of the paper
This paper presents a self-supervised temporal regularization strategy for landmark-based cardiac segmentation models with implicit anatomical correspondences, combined with automatic AHA 17-segment regional mapping. The key idea is to fine-tune an already-trained graph-based segmentation model using velocity and acceleration smoothness constraints over image sequences, without requiring additional annotations.
- 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 work addresses a clinically relevant and well-motivated problem at the intersection of segmentation, anatomical correspondence learning, and cardiac motion analysis. The methodology is conceptually clean.
The use of AHA 17-segment mapping is a major strength. Automatic regional mapping from implicit correspondences meaningfully bridges the gap between algorithmic output and clinically interpretable reporting standards.
The work has promise for enable detection of clinically meaningful phenomena (e.g., dyssynchrony, hypokinesis), which aligns well with translational goals.
Additional strengths include better-motivated velocity and acceleration regularization than generic temporal smoothing. While temporal losses themselves and HybridGNet models are available and this work is only a relatively incremental development, the integration into the current context and clinical mapping angle adds meaningful novelty.
- 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.
Assertion of clinical utility (improved reliability of EF, pathological motion detection), but quantitative evaluation w.r.t. clinical standard metrics seems to be missing. Similarly, AHA mapping is visually appealing, but additional clinical adjudication could add confidence and clinical value. Sensitivity analysis/discussion w.r.t. nonideal image quality and/or arrhythmic cycles could strengthen insight for practical echo.
While the regulation of acceleration is argued biologically, an ablation study is desirable to demonstrate the true value.
While the data is public and references existing architecture, some implementation details, \lambda_v, \lambda_a, length of temporal sequence, number of refinement epochs, etc are underspecified for reproducibility. Unclear if code will be made public. Also, no reproducibility safeguards in multiple runs or variance reporting.
- 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?
The paper presents a clinically motivated, technically sound, and well-aligned MICCAI contribution that extends implicit correspondence learning into the temporal domain and meaningfully connects segmentation to standardized cardiac reporting. However, the work would benefit from stronger validation, clearer clinical endpoints, and additional ablation studies to fully justify its claims of clinical utility.
- 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.
Dataset limitations are a reality across the board beyond this work. While nnU-Net has its limitations as a comparison benchmark, there is also admittedly a lack of a truly good candidate. Overall I feel the strength in the clinical-method connection outweighs the limitation.
Review #3
- Please describe the contribution of the paper
In this paper, the authors propose a self-supervised temporal regularization method as a post-training fine-tuning stage, which exploits temporal coherence in image sequences to achieve more temporally consistent cardiac segmentation and motion estimation without requiring frame-by-frame annotations. Specifically, the method penalizes discontinuities in velocity and acceleration across consecutive frames, enabling temporally coherent predictions while preserving the learned anatomical correspondences.Furthermore, the paper leverages these learned correspondences to automatically map landmarks to the clinical AHA 17-segment standard, thereby supporting standardized regional assessment and the detection of pathological myocardial motion patterns. Experiments on the CAMUS dataset demonstrate the potential clinical value of combining temporal consistency with automatic regional mapping.
- 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 proposes a self-supervised temporal regularization framework that improves the temporal consistency of cardiac segmentation and motion estimation without requiring frame-by-frame annotations. 2.The method leverages the learned anatomical correspondences to automatically map landmarks to the AHA 17-segment clinical standard, which supports standardized regional analysis and improves the clinical interpretability of the model. 3.The paper introduces a jitter metric to quantify the temporal regularity of landmark trajectories, which is used to capture high-frequency perturbations in motion trajectories.
- 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 evaluation is conducted on a single dataset (CAMUS) only, which is insufficient to fully assess the generalization ability of the proposed method. 2.The comparison is somewhat limited, as the method is mainly compared with nnUNet and lacks evaluation against other strong temporal or motion-aware baselines. 3.As shown in Figure 2, nnUNet achieves better clinical results on several metrics. Therefore, the paper should further clarify why temporal regularization is necessary and what additional clinical benefit it provides beyond conventional segmentation accuracy. 4.The paper emphasizes that the method learns from temporal behavior itself as a supervisory signal, yet the training procedure still includes a first stage using T independently annotated frames in each iteration.
- 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.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1.The comparative experiments are insufficient. The paper only compares the proposed method with nnUNet and does not include other strong baseline methods of a similar level, which makes the evaluation less convincing. 2.The study is conducted on a single dataset, which is not sufficient to thoroughly validate the generalization ability of the proposed method. 3.Although the paper introduces a temporal regularization loss, the part that automatically maps landmarks to the clinical AHA 17-segment standard based on learned anatomical correspondences appears to be more of an engineering application than a fundamentally novel methodological contribution.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Thank you to the authors for the detailed response. The authors’ clarification of T is helpful, especially in distinguishing the spatial supervision on annotated ED/ES frames from the temporal regularization on consecutive unannotated frames. This partially addresses my previous concern about the self-supervised formulation. The authors also further explain the relationship between AHA mapping and implicit anatomical correspondences, which makes this part appear not merely as simple engineering post-processing, but as having certain clinical interpretability. The paper is still evaluated only on the single CAMUS dataset. The authors’ explanation that other datasets lack myocardial masks can justify the choice of CAMUS, but it cannot sufficiently demonstrate the generalizability of the proposed method. The comparison with nnU-Net is also explained, but comparison only with nnU-Net still cannot prove the superiority of the method. Therefore, the rebuttal improves the clarity of the paper, but does not significantly change the strength of the evidence in the current submission. I acknowledge that this work has certain clinical potential, but I believe that the current experimental validation and method comparison are still insufficient to support a higher score.
Author Feedback
We thank the reviewers for recognizing the conceptual clarity [R2], translational potential [R2, R3], and the relevance of our self-supervised temporal regularization [R1, R3]. All reviewers highlighted the AHA regional mapping as a major strength [R1, R2, R3], alongside our Jitter metric [R3] and the potential to quantify clinical biomarkers [R2]. SELF-SUPERVISION [R3]: Apologies for the confusion caused to R3 about the first training stage using T independently annotated frames (R3 comment 7.4). Confusion was caused by a typo duplicating the variable T in Sec. 2.2 Training Procedure. The refinement objective L_refine is composite. Each iteration first processes N independent annotated ED/ES frames to compute the spatial loss L_base, which anchors the model and prevents collapse to a trivial zero-velocity solution. In a second stage, T consecutive unannotated frames are used to compute temporal losses L_vel and L_accel self-supervisedly, since no GT masks are used in this temporal regularization step. METHODOLOGICAL NOVELTY & CLINICAL UTILITY [R2, R3]: [R3] describes AHA mapping as an “engineering application.” We clarify that it is enabled by our methodological contribution: integrating implicit anatomical correspondences with temporal regularization. Unlike pixel-based models such as nnU-Net, which lack inherent anatomical consistency, our graph-based approach tracks the same anatomical points throughout the cardiac cycle. The node-to-AHA mapping is computed once on a population-averaged reference atlas and propagated across sequences (Sec. 2.3). Mask-based methods require ad-hoc frame-wise partitioning, which reduces temporal consistency and produces noisier motion traces. As [R2] notes, our mapping bridges algorithmic outputs and clinically interpretable standards. BASELINES & EVALUATION METRICS [R1, R3]: [R1, R3] question the nnU-Net comparison. Because temporal consistency is difficult to compare between landmark- and pixel-based methods, we generated “pseudo-landmarks” from nnU-Net masks via post-processing; otherwise, temporal quality could not be assessed. Direct comparison is limited by the lack of echo-video segmentation methods with implicit anatomical correspondences. We will cite [Tziritas 2024, Wu 2023] for context and correct the typo in Ref. [14]. Regarding metrics [R1], standard Dice/HD95 assess static quality but cannot quantify temporal jitter or motion smoothness. In the absence of standard temporal-consistency metrics, we use Jitter, FTD, and Jerk to quantify high-frequency disturbances and physics-based motion, clinically critical aspects common metrics ignore. DATASET AND GENERALIZATION [R1, R2, R3]: [R3] notes the use of a single dataset. CAMUS is currently the only public 2D echo dataset with both A2C/A4C views and epicardial/endocardial contours, which are required to use AHA segments. Other datasets, like EchoNet-Dynamic, lack myocardial masks, limiting their usefulness. We are working with the TED dataset [R1] as a CAMUS subcohort to further validate temporal consistency. Regarding robustness [R2], Fig. 3 includes both low-EF and normal patients; the final manuscript will clarify this and stratify existing results by CAMUS quality scores and risk. We will mention in future works that we will consider temporally labeled data (as TED) and extend regional strain analysis to hypokinesis and dyssynchrony detection. REPRODUCIBILITY [R2]: For hyperparameters [R2], a linear scheduler (0.01 to 0.001) was used for λ_v and λ_a, training over 500k iterations (~600 epochs). T was set to 20 random continuous frames per batch due to 24GB NVIDIA RTX 3090 limits. Code, optimized hyperparameters, and a demo will be released with camera-ready version to ensure full reproducibility. SUMMARY: Our self-supervised, temporally stable tool provides anatomically consistent regional tracking that pixel-wise methods cannot, supporting MICCAI’s translational goals through clinically interpretable myocardial motion analysis.
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.
The paper has a real strength in connecting technical output to clinical standards via AHA mapping. However, the evaluation is conducted with only a single dataset, tight coupling to one prior method, and a lack of comparisons to similar temporal/motion-aware baselines limit the assessment of novelty and generalizability. The clinical promise is asserted but not yet proven.
- 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 paper addresses the issue of temporal discontinuities cardiac ultrasound segmentation models and propose a self-supervised temporal regularization framework applied during the post-training refinement stage. By introducing velocity and acceleration regularization terms into the loss function, the method enforces smooth, physiologically plausible landmark trajectories without requiring per-frame temporal annotations. Additionally, the paper proposes an automatic mapping procedure that projects learned implicit landmarks onto the standard AHA 17-segment cardiac model using a population-averaged atlas, enabling automated regional motion analysis. This significantly improves the clinical value of the method. While concerns about the single dataset, and comparison with nnUnet still exist given the authors’ response, I would like to recommend ACCEPT for this work due to it clear novelty and clinical relevance.
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.
The concern about the single dataset of CAMUS is fair, but I think the authors present a quite clinical relevant study on not soly segment chambers on echo but also the AHA segment which are quite helpful for cardiologists.
Meta-review #3
- 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.
The paper presents a clinically motivated idea that connects landmark-based temporal regularization with AHA regional mapping, and the rebuttal clarified the self-supervised training procedure and some reproducibility details. However, the experimental evidence remains insufficient, as the method is evaluated only on CAMUS and lacks comparison with directly relevant temporal echocardiography baselines or full-cycle temporal evaluation resources. The clinical value of the AHA-based regional motion analysis is promising, but it is not yet quantitatively validated against clinical references.