Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

Accurate detection of end-systole (ES) and end-diastole (ED) frames is fundamental to echocardiographic assessment. Existing methods are typically developed in a view-specific manner, depend on auxiliary annotations or intensive visual modeling, which limits their generalizability. In multi-view modeling, keyframe detection is driven by shared cardiac motion, yet large appearance differences and motion patterns make unified modeling challenging. To address these issues, we propose FrameONE, a unified end-to-end framework for multi-view echocardiographic keyframe detection. FrameONE introduces a Hierarchical Motion Modeling strategy: an intra-view multi-task learning reduces appearance bias and promotes motion-focused representations within each view; an inter-view general motion learning module further separates view-agnostic dynamics from view-specific patterns, enabling shared yet flexible motion representation learning across views. Extensive experiments on 25,872 videos spanning four standard views demonstrate that FrameONE achieves state-of-the-art keyframe detection accuracy with strong cross-view generalization. Code is available at \url{https://github.com/szuboy/FrameONE}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3530_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/szuboy/FrameONE

Link to the Dataset(s)

EchoNet-Dynamic: https://aimi.stanford.edu/datasets/echonet-dynamic-cardiac-ultrasound Echo-pediatric: https://echonet.github.io/pediatric/ Echo-LVH: https://echonet.github.io/lvh/

BibTex

@InProceedings{CheRus_FrameONE_MICCAI2026,
        author = { Chen, Rusi AND Huang, Yuhao AND Zhang, Hongyuan AND Tian, Chao AND Ji, Shunan AND Zhang, Yuhan AND Ni, Dong},
        title = { { FrameONE: Hierarchical Motion Modeling for Universal Multi-view Echocardiographic Keyframe Detection } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16896},
        month = {September},
        page = {pending}
}

Reviews

Review #1

Please describe the contribution of the paper

This paper proposes an intra-view, multi-task learning approach to reduce appearance bias and promote motion-focused representations within each view. To detect end-systole and end-diastole frames, a general motion learning module is used to further separate view-agnostic dynamics from view-specific patterns. This enables shared yet flexible motion representation learning across views. In addition to phase detection the ultrasound view is identified.
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 multi-view and also extended in a multi-cycle setting.
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 algorithm was only tested on a private dataset.
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 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
- The use of FP and FN metrics in multi-cycle settings is not clear.
- The definition of the ground truth for the regression process defined in Equation (2) is unclear.
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?

Although the formulation is novel, it is difficult to evaluate the proposed method because the experimental results are based only on a private dataset.
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 proposed approach is novel and interesting, and the authors have agreed to clarify the presentation. Therefore, acceptance is recommended.

Review #2

Please describe the contribution of the paper

The paper presents a deep learning model for the estimation of End-Diastole (ED) and End-Systole (ES) events from echocardiographic image sequences. The model takes as input a sequence of echocardiographic frames and jointly predicts the view type and the keyframe locations in a multi-task learning framework.

The proposed approach relies on two main components: (i) an intra-view multi-task learning strategy aimed at reducing appearance bias and promoting motion-focused representations within each view, and (ii) an inter-view general motion learning module designed to disentangle view-agnostic dynamics from view-specific patterns. The method is evaluated on four echocardiographic datasets (one per view), resulting in a large-scale dataset of 25,872 videos.
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 strengths of the paper are as follows:
- Large-scale dataset. The use of a dataset comprising more than 25,000 videos is a strong asset, as it enables the training of robust models and supports the reliability of the reported results.
- Well-motivated architecture. The proposed design, combining multi-view learning with the disentanglement of shared temporal dynamics and view-specific residuals, is original and appears to be a promising strategy for this task.
- Strong experimental validation. The comparison with recent state-of-the-art methods, showing consistent improvements in favor of the proposed approach on a large and diverse dataset, further supports the effectiveness 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.
The main weaknesses of the paper are as follows:
- Limited clarity and consistency in the methodological description. Figure 2(b) presents key technical elements that are important for understanding the method. However, there are inconsistencies between the notation used in the figure (e.g., $\alpha(t)$, $\beta(t)$, $B_{\text{global}}$, $B_{\text{view}}$) and that used in the main text (e.g. $U_g$, $U_l^v$, $W_g$, $W_l$). In addition, some equations (e.g., $\alpha(t) = \text{Softmax}(W_a - \Delta F(t))$) are not clearly aligned with the definitions provided in the manuscript, which makes the method difficult to follow.
- Insufficient description of key components. The decomposition of the motion signal $\Delta F(t)$ into a shared rhythm component and a view-dependent residual is central to the proposed approach. While the conceptual idea is well explained, important implementation details remain unclear. In particular, it is not specified which matrices are learned (e.g., $U_g$, $U_l^v$, $W_g$, $W_l$), how they are initialized, and how they interact within the model.
- Limited validation of the disentanglement mechanism. The paper does not provide explicit evidence that the learned representations effectively capture shared and view-specific motion components. For instance, it would be useful to analyze whether the matrices $U_g$ and $U_l^v$ correspond to meaningful global and view-dependent dynamics at convergence.
- Lack of discussion on model capacity and design choices. Figure 4 suggests that the optimal dimensionality for $U_g$ and $U_l^v$ is relatively low (e.g., $256 \times 2$ and $256 \times 4$). It is unclear whether such low-dimensional representations are sufficient to capture the complexity of cardiac motion. This aspect deserves further discussion in the Experiments and Results 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 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.

(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?

My main concerns are 1) the limited clarity and consistency in the methodological description; 2) the insufficient description of key components in the methodological section and 3) the limited validation of the disentanglement mechanism. These points need to be clarified to strengthen the paper.
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’ responses to my comments, particularly regarding the improvement of consistency between the mathematical expressions in the figures and the text, as well as the clearer description of the method implementation, are satisfactory and improve both the quality and the readability of the paper. I therefore support the revised version of the manuscript incorporating these modifications.

Review #3

Please describe the contribution of the paper

The paper presents a keyframe detection method that generalizes to echocardiography sequences from different views. The method learns a shared motion representation, enabling ES and ED frame detection that bypasses appearance differences between views. The method is validated on 4 echocardiography datasets and compared to 5 state-of-the-art methods. The results show competitive performance across various datasets.
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 tackles the challenging problem of keyframe extraction and motion modeling in multiview settings. The experimental evaluation spans multiple datasets and includes comparison with several methods from the literature, which strengthens the credibility of the results. The consistent performance across different views suggests that the proposed method is promising for multiview keyframe detection.
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 problem of multiview ES/ED detection is relevant; however, the motivation could be strengthened by more clearly explaining the practical limitations of view‑specific models and the advantages over a simple pipeline combining view classification with view‑specific inference. Additionally, results show that view‑specific models consistently achieve the best or second‑best performance, e.g, OFM or the proposed method trained on a single view, which warrants further discussion and justification of the added value of the proposed multiview approach.

Other remarks:
- Section 2.2: Ug and Ul are introduced, but there is no explanation of how they are computed/obtained. This point should be clarified since the global and local decomposition is a key contribution of the paper.
- Sparisty on z (equation 6) is not justified/discussed.
- Compared methods are not introduced, which makes it difficult to interpret the results and the impact of the components on the proposed method.
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?

The methodology is clearly explained, and the experimental results are promising. However, the paper does not sufficiently substantiate or discuss the motivation and clinical relevance of the proposed approach.
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 #4

Please describe the contribution of the paper

The authors introduce a unified model for predicting key frames, namely end-diastole (ED) and end-systole (ES), from echocardiography videos acquired from different views, including A4C, A2C, PSAX, and PLAX. The proposed model integrates a hierarchical motion modeling strategy with both intra-view multi-task learning and inter-view general motion learning, aiming to achieve better view-agnostic representation learning for ED/ES frame regression. The paper compares the proposed approach with several baseline methods for cardiac phase detection and claims improved performance on most metrics across multiple views.
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 strength of this paper is that it attempts to model both global motion and view-specific motion within a unified framework. In particular, the authors use multi-task learning to capture view-specific motion representations and inter-view general motion learning to model view-agnostic dynamics across different echocardiography views. This design improves the downstream task of ED/ES regression when combining all four views and appears to achieve competitive performance compared with other baseline methods.
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.There are several aspects of the paper that are not clearly presented in the current version. For example, in Section 2.2, the relationship between ht and Δft^ is unclear and is not illustrated in Figure 2. It is also not evident how the baseline methods are trained. Are they trained separately for each view, while only the proposed FrameOne model is trained jointly on all four views? In addition, for the multi-cycle evaluation, it is unclear whether the authors report false positive counts, false negative counts, or false positive/negative rates. The units of the evaluation metrics should also be explicitly stated in the tables to avoid confusion.

2.In Table 1, it is unclear whether the best-performing method is significantly better than the second-best method. Most of the differences in MAE, assuming the unit is frames, are smaller than 0.5 frames. I am not convinced that such an improvement is meaningful, especially considering that the proposed method requires more training data from multiple views as well as ED/ES annotations. I would expect the authors to better justify the practical significance of this improvement. For example, does improved ED/ES detection meaningfully affect downstream tasks such as ejection fraction estimation or strain analysis?
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 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

1.How are regression labels assigned to frames that are neither ED nor ES? Are the labels linearly interpolated, or is another assignment strategy used? 2.In equation (4), how is U(l,v) activated when the input video comes from one of the four views? Is the view label required beforehand so that the corresponding U(l,v) can be selected? If so, does the method require prior knowledge of the view during inference?
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, I find this paper interesting because it proposes a unified model for cardiac phase detection across multiple echocardiography views. The architectural design is well motivated from both intra-view and inter-view perspectives, and the overall methodology is reasonably presented. However, I would expect the authors to clarify the issues raised above, particularly regarding methodological presentation, evaluation details, and the practical significance of the reported improvements.
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.

I think the authors have addressed most of my concerns related to method and evaluation details. I will recommend accept.

Author Feedback

We thank the reviewers for recognizing the novelty (R1), well-motivated architecture (R2,R3,R4), and promising experimental results (R2,R3,R4). Clarifications have been provided to address the comments.

Q1.Large-scale Public Dataset (R1) FrameONE is built on large dataset with 25,872 videos. Specifically, it includes three public datasets(98.3%), i.e., EchoNet-Dynamic (A4C), EchoNet-LVH (PLAX), and EchoNet-Pediatric (PSAX). Only one in-house A2C dataset is used due to the limited accessible resources.

Q2.Motivation of Multi-view Modeling (R3) 1) As noted by R3, multi-view modeling is challenging, cause simply mixing all views does not guarantee improvement. In Tab.2, direct multi-view training yields 3.093 MAE, even worse than FrameONE (Single). FrameONE addresses this via hierarchical motion disentanglement: IML reduces appearance bias, while IGM separates shared and view-specific motion, improving MAE from 3.093→2.700→2.362.2) Previous methods treat each view as an independent task, which prevents cross-view motion learning and generalization. In contrast, our motivation is that different views share the same underlying cardiac cycle and view-specific motion.

Q3.OFM vs. FrameONE (R3,R4) FrameONE achieves significantly lower MAE (2.907→2.386, p<0.001) and also shows improved stability and efficiency. 1) FrameONE reduces False Positive (3.438 → 2.176 frame). 2) Fewer parameters (43.93M vs. 14.35M) and higher speed (132.1 vs. 246.6 FPS).

Q4.Methodological Clarity (R2,R3,R4) A) Unified notation (R2): We unified B_global and B_view as U_g and U_l^v for the shared and view-specific bases, and alpha(t) and beta(t) as z_g(t) and z_l(t) for the corresponding motion coefficients. B) Clarification of key components (R2,R3): 1) Coefficient Estimation: W_g and W_l project motion feature difference Δf_t into a shared coefficient space z_g(t) and a view-specific coefficient space z_l(t), followed by softmax normalization (Eq.5). 2) Motion Reconstruction: z_g(t) and z_l(t) act as weights to combine the shared basis U_g and the view-specific basis U_l^v, producing the reconstructed motion feature Δf̂_t (Eq. 6). 3) Optimization: U_g and U_l^v use orthogonal initialization, while W_g and W_l use Xavier initialization. All four matrices are jointly optimized during end-to-end training. C) No view label is required during inference (R4): FrameONE predicts the view via a classification head, then selects the corresponding view-specific basis U_l^v.

Q5.Disentanglement Mechanism (R2) Tab.2 validates the effectiveness of the decomposition design. The global basis reduces mean MAE from 2.700 to 2.470 frames, while the local basis further improves performance by 0.613 frames, suggesting that the two representations capture different aspects of cardiac motion patterns.

Q6.Low-dimensional Representation (R2,R3) Low-dimensional representations are sufficient since ES/ED detection targets extrema along constrained cardiac motion trajectories rather than full anatomy [Ref.1]. In ablations (Fig.4), MAE is stable across basis dimensions, with the best performance at K_g=2, K_l=4.Ref 1: Rohé et al. Low-dimensional representation of cardiac motion using barycentric subspaces. MedIA, 2018. Q7.Implementation Details (R1,R3,R4) (1) FP/FN in multi-cycle (R1,R4): FP_Rate = N_FP / N_pred, FN_Rate = N_FN / N_GT. N_FP and N_FN denote unmatched predictions and missed GT, respectively. N_pred and N_GT are the total number of prediction and GT, respectively. (2) GT definition (R1,R4): ES and ED are assigned values 0 and 1, with intermediate frames represented via Gaussian interpolation with standart deviation 2.(3) Training Strategy (R3,R4): All baselines are trained independently for each view, whereas FrameONE is jointly trained on all four views. (4) h_t and Δf̂_t (R4): Δf̂_t is used for reconstruction loss, while h_t is a Transformer-based representation derived from z_t for ES/ED prediction. We will clarify the details in camera-ready version.

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 proposes a unified hierarchical motion modeling framework for ED/ES keyframe detection across multiple echocardiographic views. Reviewers appreciate the multi-view design, the large-scale dataset, and the improvements over recent baselines. However, they question the practical limitation, the motivation and clinical relevance. They also raise concerns about methodological clarity, missing public-dataset validation, weak evidence for the disentanglement. Given these mixed assessments, the paper is invited to rebuttal for the authors to address the concerns.
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 clarified the methodological details, notation inconsistencies, and demonstrated statistical significance of improvements, leading three reviewers to accept post-rebuttal. The recommendation is therefore Accept.

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 author’s response addresses the reviewers’ concerns.

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 a unified framework for multi-view ED/ES keyframe detection in echocardiography and demonstrates consistent performance across multiple cardiac views on a large-scale dataset. The rebuttal clarified several concerns regarding the methodological formulation, dataset composition, evaluation protocol, and motivation for multi-view modeling. Overall, the paper provides a useful contribution to multi-view echocardiographic motion analysis.

back to top

FrameONE: Hierarchical Motion Modeling for Universal Multi-view Echocardiographic Keyframe Detection

Author(s):