List of Papers Browse by Subject Areas Author List
Abstract
Echocardiography is widely used for assessing cardiac function, where clinically meaningful parameters such as left-ventricular ejection fraction (EF) play a central role in diagnosis and management. Generative models capable of synthesising realistic echocardiogram videos with explicit control over such parameters are valuable for data augmentation, counterfactual analysis, and specialist training. However, existing approaches typically rely on computationally expensive multi-step sampling and aggressive temporal normalisation, limiting efficiency and applicability to heterogeneous real-world data.
We introduce EchoLVFM, a one-step latent video flow-matching framework for controllable echocardiogram generation. Operating in the latent space, EchoLVFM synthesises temporally coherent videos in a single inference step, achieving a \textbf{$\sim$50×} improvement in sampling efficiency compared to multi-step flow baselines while maintaining visual fidelity. The model supports global conditioning on clinical variables, demonstrated through precise control of EF, and enables reconstruction and counterfactual generation from partially observed sequences. A masked conditioning strategy further removes fixed-length constraints, allowing shorter sequences to be retained rather than discarded.
We evaluate EchoLVFM on the CAMUS dataset under challenging single-frame conditioning. Quantitative and qualitative results demonstrate competitive video quality, strong EF adherence, and 57.9\% discrimination accuracy by expert clinicians which is close to chance. These findings indicate that efficient, one-step flow matching can enable practical, controllable echocardiogram video synthesis without sacrificing fidelity. Code available at:\href{https://github.com/EngEmmanuel/EchoLVFM}{\texttt{ \faGithub\,EchoLVFM}}
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3595_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/EngEmmanuel/EchoLVFM
https://huggingface.co/spaces/EngEmmanuel/EchoLVFM
Link to the Dataset(s)
CAMUS Dataset: https://humanheart-project.creatis.insa-lyon.fr/database/#collection/6373703d73e9f0047faa1bc8
BibTex
@InProceedings{OlaEmm_EchoLVFM_MICCAI2026,
author = { Oladokun, Emmanuel AND Thomas, Sarina AND Šprem, Jurica AND Grau, Vicente},
title = { { EchoLVFM: One-Step Video Generation via Latent Flow Matching for Echocardiogram Synthesis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces EchoLVFM, which generates an echocardiogram video in a single forward pass given one reference frame and a target ejection fraction. The core idea is a masking scheme that lets training and inference work directly on clips of different lengths instead of padding or resampling to a fixed length, plus a reconstruction regulariser to stabilise training. On CAMUS it runs claims to run 50x faster than a 25-step linear flow-matching baseline at comparable FID and FVD, and SOTA quantitative and trumps radiologist 42% of the time.
- 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.
Investigation of 1-step sampling from flow matching models is still scarce, especially in medical imaging and ultrasound video generation. The contribution targets an under-studied area.
The qualitative study is welcome and complements the quantitative scores well.
Rejection sampling with an independent EF estimator is a well-chosen evaluation. It decouples sample quality from EF controllability, which conditional synthesis work often conflates. The protocol can move numbers against the method, and in places does, which is what a meaningful test should be able to do.
- 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.
Evaluation is on CAMUS only, with no seed variance reported. Several Table 1 deltas are small, on the order of 1-2 FID points, and point estimates leave no way to tell whether they sit inside training noise. An external dataset such as EchoNet-Dynamic is expected at this level of claim and is not provided.
The 50x speed-up compares EchoLVFM to the authors’ own re-implementation of a linear baseline, not to a published model. Comparisons to HeartBeat and ControlEchoSynth are transcribed across different splits and resolutions. The headline comparison is only clean internally, and the external one is not directly interpretable.
The Gen EF protocol samples target EFs that differ by at least 5% from the source, drawn over [0,100]. With single-frame conditioning, many of these targets are physically incompatible with the anatomy fixed by the condition. The reported R² is not separable from the harshness of the protocol, and the controllability claim is under-tested against clinically plausible targets.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper targets an under-studied area, the rejection-sampling evaluation is well-designed, and the clinician study sets the right norm for this application domain. The one gap that could tip it the other way is single-dataset evaluation without seed variance. If possible without running further experiments, adding variance would be welcome. For future work, evaluating on a larger cardiac ultrasound dataset such as EchoNet-Dynamic would make the submission stronger.
- 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 #2
- Please describe the contribution of the paper
The authors introduce EchoLVFM, a framework for generating echocardiogram videos in a controllable way using a one-step latent flow matching approach. The main contribution is adapting MeanFlow-style flow matching for conditional video generation, where the model can be guided by clinically meaningful factors such as ejection fraction and partial video inputs. They also propose a new loss function. In addition, the method uses a masking and padding strategy that allows it to handle videos of different lengths without forcing them into a fixed temporal format.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper is technically novel, applying one-step flow matching to echocardiography video generation. It incorporates ideas such as MeanFlow-style one-step inference, a masked temporal loss, and padding-aware conditioning. The efficiency gains are significant, reducing the generation process from 25 steps to just a single step.
- 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.
There is no evaluation of whether the generated videos preserve pathology-specific features. It also lacks comparisons with state-of-the-art conditional video diffusion models adapted to this task, not just within echocardiography. It also appears that the linear approach, along with the methods described in references [28] and [13], produces more visually realistic results, despite being less efficient—a tradeoff that may be more important in a clinical setting.
- 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 paper is well written and technically novel. It also presents an efficient one-step generation framework, with strong results in both sampling speed and video quality. However, the evaluation feels somewhat limited. It mainly focuses on global EF adherence and perceptual realism, but doesn’t really address whether the generated videos preserve clinically meaningful structure or pathology-specific features.
- 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 #3
- Please describe the contribution of the paper
The paper proposes EchoLVFM, a one-step latent video flow-matching framework for synthesizing echocardiogram videos. It introduces a masked conditioning strategy to handle variable-length sequences without aggressive temporal interpolation.
- 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.
This article combines the single-step generation method of MeanFlow [1] with EchoFlow [2] and applies it to the medical scenario. If the experiments are carried out comprehensively, it has certain publication value.
[1]. Geng Z, Deng M, Bai X, et al. Mean flows for one-step generative modeling[J]. arXiv preprint arXiv:2505.13447, 2025.[2]. Reynaud H, Gomez A, Leeson P, et al. Echoflow: a foundation model for cardiac ultrasound image and video generation[J]. arXiv preprint arXiv:2503.22357, 2025.
- 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 core contribution is the global conditioning on ejection fraction (EF). However, there is not a single figure demonstrating this capability. The authors must provide qualitative comparisons showing how the same initial state evolves under different target EFs (e.g., normal vs. impaired LV function) to visually prove that the physiological dynamics are accurately controlled, rather than relying solely on the surrogate metrics in Table 1.2.The paper includes only one qualitative figure (Fig. 2), which presents a seemingly ideal scenario. For a highly heterogeneous modality like echocardiography, it is imperative to showcase a broader range of generated samples across different sequence lengths, image qualities, and pathological states. 3.While the clinical evaluation by expert cardiologists (58% confusion rate) is quantitatively impressive, the complete lack of visual examples from this 120-video test set leaves reviewers and readers unable to independently verify the perceptual quality of these synthetic videos. 4.In the domain of medical image synthesis, quantitative metrics (FID, FVD) often fail to capture clinically critical artifacts. Relying almost exclusively on tables while hiding visual evidence severely undermines the credibility of the proposed EchoLVFM framework. 5.An abstract is usually a coherent paragraph of text, but the abstract of this article has been split into three parts. This highly formulaic structure is a typical output format when ChatGPT or Claude writes or extracts an abstract. The author most likely copied and pasted the text generated by a large language model directly into the LaTeX template, not even bothering to delete the extra line breaks. This not only indicates that the author relied heavily on AI-generated text, but also reveals that they were in a hurry to complete the article and had a perfunctory attitude towards typesetting.
- 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 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
Generative AI papers—especially in medical imaging where anatomical fidelity is paramount—must adhere to strict visualization standards. The current manuscript reads like a technical report that forgot to include its figures. Including comprehensive, publication-quality visual comparisons (e.g., demonstrating variations in EF control, edge cases, and temporal consistency maps) is mandatory for this venue. Relying almost exclusively on tabular metrics undermines the scientific validity of a synthesis paper.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The primary factor driving this rejection is the severe discrepancy between the paper’s methodological claims and its visual evidence. The authors claim to achieve controllable video generation based on ejection fraction (EF). However, there is not a single publication-quality figure demonstrating this control (e.g., comparing generated sequences at varying EF targets from the exact same initial frame). In generative AI applied to medical imaging, tabular proxy metrics (like FID, FVD, or proxy EF MAE) are entirely insufficient without rigorous visual validation. Furthermore, the overall presentation lacks necessary academic-style rigor; the methodology requires a clear, modular technical flowchart explaining the masking mechanisms, and the results section requires extensive visual proof of anatomical fidelity. Without these comprehensive visual proofs and robust failure-case analyses, the clinical and scientific validity of the proposed EchoLVFM framework remains fundamentally unverifiable.
- 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
We thank the reviewers and the meta-reviewer for their feedback, and the meta-reviewer for the positive recommendation.
(1) EF controllability protocol (R1, Meta). ‘Rec’ task conditions on the source EF (anatomy and EF coherent by construction) whilst ‘Gen’ uses a different EF. Under single-frame conditioning, the frame fixes cardiac anatomy and distant target EFs can be physically inconsistent with it. However, EF is also not uniquely determined by a single frame, as it depends on the relative volume change between ED and ES rather than absolute size at one time point. This is why pmf=50% Gen R-squared collapses to -1: with x_m carrying multiple frames at the original EF, requesting a different EF makes x_m an anatomical confounder. The relevant controllability measure is therefore the pmf=50% Rec result (R-squared=93%, MAE=3.1%, RMSE=3.8%), where anatomy and EF are coherent.
(2) 50x speedup and external baselines (R1, R2). The 50x claim is internal: EchoLVFM at 1 step versus an identical-architecture Linear FM baseline trained on identical data at 25 inference steps. No published variable-length linear-flow echocardiogram model exists as a directly comparable external baseline. HeartBeat and ControlEchoSynth values in Table 1 are quoted from the original publications at 256x256 and 64x64 respectively, with FVD computed at different clip lengths; we share splits with ControlEchoSynth, but HeartBeat splits were not released. R2’s observation that Linear is marginally more perceptually convincing on LPIPS and the clinician vote is exactly the efficiency-fidelity trade-off the paper reports: Linear needs much more compute for a small perceptual gain and clinics are not equipped with significant compute, which supports prioritising efficiency.
(3) Visual evidence, perceptual fidelity, and pathology preservation (R2, R3, Meta). The paper targets EF-conditioned rather than pathology-conditioned synthesis, which motivated our prioritisation of echocardiography-specific baselines. Within the 8-page limit we prioritised the blinded clinician study (n=120; two cardiologists each with 15+ years’ experience, calibrated against real examples) as the test of perceptual realism: experts of this seniority reliably flag artefacts inconsistent with cardiac anatomy or pathology, and the resulting 58% accuracy is near chance (50%). Fig. 2 includes an M-mode slice showing temporal consistency. The paper has a GitHub and HuggingFace page which contain many generated videos and fail cases. Pathology-aware evaluation is valuable future work, but neither CAMUS nor EchoNet-Dynamic provides pathology labels, limiting rigorous quantitative evaluation here.
(4) Seed variance and dataset scope (R1, Meta). Each Table 1 row already aggregates three independent noise draws per test video (300 samples), giving a partial variance signal. Multi-seed retraining and EchoNet-Dynamic evaluation are valuable next steps; these suggestions will be taken into account in future extensions of this work.
(5) Methodological novelty and presentation (R2, R3). EchoLVFM extends substantially beyond a MeanFlow+EchoFlow combination. EchoFlow already operates in a medical echocardiography setting and uses strictly linear flow with single-frame, fixed-length conditioning and generation. EchoLVFM contributes one-step MeanFlow-style generation, a masked temporal loss, a reconstruction regulariser, variable-length synthesis, arbitrary partial-video conditioning, and controllable generated sequence length, none of which EchoFlow or MeanFlow supports. R2’s call for non-echo-specific video diffusion baselines is reasonable; we know of none adapted to variable-length echocardiogram synthesis with EF conditioning. The manuscript is author-written; the abstract’s paragraph structure is a deliberate stylistic choice for readability, consistent with our prior work.
We thank the reviewers again for engagement that has helped us identify places where the manuscript can be clearer.
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 paper proposes EchoLVFM, a one‑step latent flow‑matching framework for controllable echocardiogram video synthesis. This aims to achieve single‑step video generation conditioned on clinically meaningful variables such as ejection fraction while supporting variable‑length sequences which is technically novel and relevant. One‑step flow matching remains under‑explored in medical video generation, and is a practical and well‑motivated design choice. The reported efficiency gains are substantial, and the evaluation is strong, combining quantitative metrics, rejection‑sampling analysis, and a blinded clinician study, which is appropriate for this domain.
The main limitations are in the scope and interpretation of the evaluation. Results are confined to the CAMUS dataset. The speed‑up can be better justified. The EF controllability protocol is intentionally challenging but includes physiologically implausible EFs under single‑frame conditioning, which seems almost too challenging to interpret. While qualitative examples and clinician results are encouraging, clearer visual demonstrations of different EF‑specific dynamics would strengthen the claims.
Reviewer #3 raises some legitimate concerns around visualization and interpretability; however, that review is weakened by speculation and tone, which I have not taken into account.
Overall, the paper addresses an under‑studied problem with a technically sound, efficient approach and a well‑designed evaluation for clinical relevance.