List of Papers Browse by Subject Areas Author List
Abstract
Cardiac phase identification is important for assessing heart function in echocardiography, yet existing automatic approaches typically rely on supervised learning or strong structural assumptions. We show that self-supervised learning of smooth latent dynamics can enable phase-conditioned geometric analysis of cardiac motion. In this work, we present a Spline Autoencoder that learns temporally smooth and consistent evolutions of latent embeddings, capturing cardiac motion as low-rank dynamical structures. We discover a cardiac phase variable using a label-agnostic algorithm based on the spectral embedding of normalized latent trajectory velocities, enabling precise temporal localization of end-diastole (ED) and end-systole (ES) events. Evaluated on the EchoNet-Dynamic dataset, our method achieves a localization mean absolute error (MAE) of 2.3 frames (ED/ES MAEs of 2.7/3.6 frames after label-informed semantic assignment with 94.0% accuracy), which is competitive with supervised methods. Zero-shot evaluation on the EchoNet-Pediatric dataset demonstrates robust generalization across populations and imaging views, with a localization MAE of 3.0 frames. These results suggest that temporal smoothness and consistency are effective inductive priors for learning cardiac motion dynamics, yielding a latent geometry that supports precise phase-event localization, with lightweight calibration for semantic ED/ES assignment. Our code is available at https://github.com/LCJKwan/SplineAE-Cardiac.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1703_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/LCJKwan/SplineAE-Cardiac
Link to the Dataset(s)
EchoNet-Dynamic: https://github.com/echonet/dynamic
EchoNet-Pediatric: https://echonet.github.io/pediatric/
BibTex
@InProceedings{KwaLeo_Selfsupervised_MICCAI2026,
author = { Kwan, Leong Chit Jeff AND Chung, Albert C. S.},
title = { { Self-supervised Spline Autoencoder for Smooth Latent Dynamics in Echocardiography } },
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
The paper proposes an unsupervised framework for cardiac phase discovery in echocardiography based on smooth latent dynamics. Its main contribution is a spline-based autoencoder that learns temporally smooth and consistent latent trajectories from echo videos without phase labels, combined with a dynamical phase discovery method that uses normalized latent velocities, spectral embedding, and phase-aligned projection to localize ED and ES.
- 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 main strengths of the paper are the relevance of the problem, the coherent methodological pipeline, and the zero-shot evaluation on pediatric echocardiography. In particular, the idea of combining smooth spline-based latent trajectories with a dynamical phase discovery procedure is conceptually clear and reasonably well motivated. The zero-shot transfer results across populations and views are also a positive aspect of the experimental evaluation.
- 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.
While the latent trajectory learning and phase discovery stages are unsupervised, the final ED/ES semantic assignment uses a global ED–ES alignment axis computed from training-set ED and ES embeddings, which relies on label information.
The paper’s technical description could also be improved. For example, the encoder architecture is not described clearly enough in the main text, which weakens reproducibility.
Finally, although the results are competitive, the gain over the strongest prior unsupervised baseline is limited. On EchoNet-Dynamic, the pooled localization MAE improves only modestly over LMP, and the ES result is in fact clearly worse than LMP (3.6 vs. 2.0 frames), even though the overall pooled metric is slightly better. This makes the practical advantage less convincing than the headline result suggests.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
The idea of learning smooth latent trajectories for phase discovery is reasonable, and the zero-shot pediatric evaluation is a positive aspect, but the overall contribution feels closer to an incremental refinement of existing unsupervised phase-analysis approaches than to a clearly new formulation.
A major factor in my score is that the “unsupervised” claim should be stated more carefully. Although the representation learning and phase discovery stages are unsupervised, the final ED/ES semantic assignment relies on a global ED–ES alignment axis computed from training-set ED and ES embeddings, which uses label information. In my view, this weakens the end-to-end unsupervised framing that is central to the paper’s positioning. I was also not fully convinced by the experimental margin over the strongest prior unsupervised baseline. On EchoNet-Dynamic, the pooled localization MAE is only modestly better than LMP, and the ES error is substantially worse than LMP, even though the headline localization number is slightly stronger. This makes the practical improvement less compelling than the overall presentation suggests. In addition, some parts of the technical description, such as the encoder design, are not described clearly enough, which further reduced my confidence.
Overall, I found the paper promising, but in its current form I see it as slightly below the acceptance threshold.
- 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.
The authors have constructively addressed my concerns about the ‘unsupervised’ claim by reframing the paper around self-supervised representation learning and adjusting the title accordingly. While the ES localization error on EchoNet-Dynamic remains higher than LMP, the overall contribution particularly the spline-based smooth latent dynamics and zero-shot pediatric transfer has merit.
Review #2
- Please describe the contribution of the paper
The manuscript introduces a spline autoencoder that works in latent space by enforcing consistency. The proposed method avoid explicit periodicity constraints during training.
- 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.Clear motivation with identification of real gap 2.Results outperform prior unsupervised baselines. 3.Zero-shot pediatric performance is a real positive.
- 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.My main issue is that claim of unsupervised. The latent learning is unsupervised. But the actual ED versus ES semantic assignment uses a post-training calibration axis computed from training-set ED and ES embeddings. That is supervision. The authors admit this in Section 2.3, yet the title and the repeated messaging still push “unsupervised cardiac phase discovery” .
2.Manuscript says both datasets have only one ground-truth ED/ES pair per video, and reports MAE to the temporally closest prediction. That means authors are not evaluating full-cycle phase tracking, beat consistency, or whether multiple predicted extrema are sensible. For me it is evaluating nearest-hit localization against a sparse annotation scheme. Maybe I understood the paper wrong. But would require a clarification on that.
3.The proposed method based on many engineering choices such as K=15 and etc. If authors can elaborate how those are made and justify the choices that would be great. for me it is like a pipeline with clever ingredients that works correctly.
4.Failure case analysis is weak at this moment. I dont get any insights into when, why and how it breaks.
- 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.
- 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?
Even though core idea is interesting, the misleading “unsupervised” claim and the under-justified pipeline lead to a weak reject.
- 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.
Authors are addressing relevant problems in the MIC domain. There are some issues raised by all the reviewers regarding the title. I think the authors have to change it in the camera-ready version. Authors have addressed my main concerns. hence I am going with acceptance
Review #3
- Please describe the contribution of the paper
The paper proposes a Spline AutoEncoder framework for the unsupervised identification of end-diastole and end-systole in echocardiography videos. By fitting a spline to the latent space, authors enforcing temporal smoothness through a second-order finite-difference penalty. A novel dynamical phase discovery algorithm is introduced, which uses spectral embedding of normalized latent velocities to recover a circular phase variable. Event localization is then performed by identifying the peak and valley of the projected latent trajectory onto a 2D plane using Partial Least Squares and an a SVD decomposition. The method was evaluated using and internal and external dataset (EchoNet-Dynamic and EchoNet-Pediatric). Authors perform a small ablation analysis.
- 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 method shows some degree of novelty by enforcing temporal smoothness through fitting a spline to the latent trajectory. Additional, authors propose to use the normalized velocities in the tangent space for spectral embedding to handle cyclic dynamics in high-dimensional latent spaces while remaining more stable than position-based methods. — The zero-shot evaluation on pediatric data (both A4CH and PSAX views) demonstrates that the model captures fundamental cardiac motion geometry rather than just dataset-specific appearance. — Competitive Unsupervised Performance — The spline-based formulation combined with the cross-reconstruction provides a mathematically grounded way to enforce temporal consistency without requiring periodic assumptions.
- 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.
— As noted by the authors, the recovered phase is invariant to global rotation. The reliance on a post-training calibration using a global average vector to distinguish ED from ES introduces a small supervised component (or at least a requirement for a labeled reference set) to achieve final labeling which goes against the main goal of the paper. It is not well explained how the authors achieve this lightweight post-training.
— While localization is stable, the labeling accuracy drops significantly (from 93.8% to 83.2%) when moving from A4CH to PSAX views in the pediatric dataset. This suggests the “global ED-ES axis” is not entirely view-invariant.
— While λ spline is ablated, there is limited discussion on the impact of the number of control points (n_ctrl =16) or the spline degree (p=3). It is unclear how these choices affect the capture of high-frequency motion details or arrhythmias.
— Results should include 95% confidence intervals (for example using bootstrap), especially for the pediatric zero-shot results in the errors (Table 1 and 2).
- 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?
The method proposed shows competitive Unsupervised Performance and some novelty
- 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
Constructive Comments:
— Include 95% CI. — A brief analysis of the inference time per video would be beneficial, as solving the normal equations for splines is efficient but spectral embedding can be computationally intensive for very long sequences.
— In Figure 3, the authors show a label assignment error. A more detailed discussion or a confusion matrix on which types of heart rates or pathologies (e.g., tachycardia) lead to these flips would be clinically valuable.
— It would be interesting to see an ablation comparing the PLS-based projection (Step 3 in Algorithm 1) against a standard PCA projection to further justify the “phase-conditioned” aspect.
— In Fig 1, clarify if the partitions are mutually exclusive or if they can overlap in time, as this affects the difficulty of the cross-reconstruction task
- 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 method proposed shows competitive Unsupervised Performance and some novelty
- 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.
The authors addressed most of the concerns I pointed out.
Author Feedback
We are grateful for the valuable feedback from all reviewers. We respond to the suggestions and concerns below:
(R1/R2/R3) Unsupervised claim and paper positioning We agree that wording around “unsupervised” should be framed more cautiously. While SplineAE training and localizing trajectory extrema are label-free, resolving the semantic ambiguity of ED/ES on detected extrema requires label information. We clarify that there is no fine-tuning or gradient-based optimization on labels. Instead, we compute the sign of the dot product between the extrema axis and the training-set global ED/ES axis, as described in Section 2.3. Our intended contribution is primarily representational, as described in Section 1: Our self-supervised SplineAE discovers low-rank loop-like latent dynamics from echocardiography. Moreover, instead of enforcing a fixed 2D latent bottleneck [15] or explicit periodicity constraints [6], we rely only on temporal smoothness and consistency. Label-free latent geometry discovery of cardiac phase is supported by topology analysis, qualitative observation of latent loops, cardiac event localization, and zero-shot pediatric transfer.
ED/ES detection is a downstream task to evaluate clinical relevance of learnt representations. To reflect this positioning, we will rephrase the title to: “Self-supervised Spline Autoencoder for Smooth Latent Dynamics in Echocardiography”. We will revise the wording of manuscript content accordingly, to clearly explain self-supervised training, label-agnostic phase extraction with event localization, and label-informed ED/ES semantic assignment.
(R1) Performance margin and ES error In addition to relaxed constraints during representation learning, our label-free localization performance is competitive. Localization error is equal to pooled error with perfect ED/ES labelling. Meanwhile, ED/ES MAE degradation are caused by semantic label flips after otherwise accurate localization, where errors are as large as the ED-ES interval.
(R2/R3) Failure case We observe localization failure when a coherent latent loop does not emerge, often in low-quality videos with chambers moving out of view. Pediatric PSAX ED/ES MAE rises due to the 10% reduction in label assignment accuracy from A4C, as semantic directions in latent space may shift with view. However, the label-agnostic localization derived from trajectory geometry remains stable, as the MAE remains at 3.0 frames.
(R2/R3) Hyperparameter choices Cubic splines balance smoothness and expressivity, while nctrl=16 for 64-frame partitions, approximately 1 control point per 4 frames, captures curvature without fitting frame-level noise. K=15 empirically balances preserving cross-cycle semantic neighbours against over-connecting temporally adjacent samples, which depends partly on frame rate. We clarify that splines are only used in training; inference uses encoder embeddings directly in Algorithm 1. (R1) Reproducibility The full architecture and implementation will be included in the code repository release. Briefly, the encoder uses residual blocks [3×3 conv, GroupNorm, GELU, 3×3 conv], with 2 residual blocks before each downsampling stage, and 5 such stages before average pooling.
(R2) Nearest-hit localization The reviewer’s interpretation is correct. This protocol follows LMP [15] for EchoNet Dynamic under the same sparse-label limitation, so the comparison is like-to-like. Qualitative inspection of videos and Fig.3 suggest that our method aligns trajectory extrema to ED/ES events without dense over-prediction.
(R3) Other concerns We reported mean±SD, that could be reported as 95CI for clarity. Cross-partitions are mutually exclusive if the video is longer than 127 frames, but may randomly overlap if the video is too short. We also hypothesize that model errors or geometry irregularities may correlate with clinical pathologies. However, the EchoNet datasets lack pathology labels. We believe this is an important direction for future work.
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 tackles a phase detection problem with an interesting idea. However, it is critically undermined by an inaccurate central premise: the unsupervised claim is technically incorrect for the end-task. The quantitative advantage is marginal, and in a key metric, it is worse. The paper falls below the acceptance threshold due to a flawed central claim and insufficient performance advantage.
- 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 address well previous concerns and all reviewers agreed to accept 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.
The rebuttal addressed reviewers’ concerns and all three reviewers moved to accept post-rebuttal.
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 paper presents an interesting self-supervised representation-learning approach for cardiac phase analysis in echocardiography. The main concern was that the original “unsupervised” framing was overstated because the final ED/ES semantic assignment uses a training-set ED/ES alignment axis. The rebuttal addresses this issue constructively and after the rebuttal, all reviewers moved to acceptance. The evidence support accept. Congratulations!
