List of Papers Browse by Subject Areas Author List
Abstract
CT-based transesophageal echocardiography (TEE) simulation provides a vital environment for procedural training and autonomous robotic navigation. While recent advances approach realistic simulation via conditional generative models, they typically require strictly paired cross-modal data or costly expert annotations. Acquiring such supervision is practically infeasible due to the intrinsic mismatch between static 3D CT and dynamic operator-dependent 2D TEE, compounded by disparate imaging physics and non-rigid tissue deformations. To address this, we present the first generative TEE simulation engine capable of synthesizing high-fidelity TEE images directly from local CT slices without paired supervision. Our framework operates in a latent space constrained by an adversarial VQ-VAE, pre-trained on a large-scale TEE database to capture the modality-consistent prior. To tackle the central challenge of unpaired translation, we introduce Optimal Transport Coupled Flow Matching (OTC-Flow) as our core technical design, which establishes reliable pseudo-correspondences between CT-conditioned features and target TEE latents, enabling robust flow matching regression. Furthermore, as an indispensible complement, we integrate a pseudo-supervised anatomical consistency loop that leverages robust CT geometry to enforce strict spatial alignment in a completely label-free manner. Extensive experiments demonstrate that our framework significantly outperforms representative unpaired translation models in both fidelity and anatomical consistency. Code, preprocessing scripts, and split details are available at https://github.com/yunhaoli24/OTC-Flow.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0555_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/yunhaoli24/OTC-Flow
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiYun_OTCoupled_MICCAI2026,
author = { Li, Yunhao AND Feng, Yidan AND Chiu, Ho-On Alston Conrad AND Lam, Simon Cheung-Chi AND Pang, Yan AND Wang, Qiong AND Wong, Chun-Ka AND Qin, Jing},
title = { { OT-Coupled Flow: Towards Realistic TEE Simulation via Unpaired Cross-Modality Translation with Anatomical Consistency } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16893},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper develops and validates a synthetic TEE generation approach from CT using optimal transport coupled flow matching (OTC-Flow) for correspondence in latent space. The method operates without paired supervision, which is often infeasible due to the differences between static 3D CT and dynamic 2D TEE. The work is relevant to the conference as a methodological improvement for synthetic TEE generation with a clear clinical application in operator training and autonomous robotic naivgation development.
- 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 clinical background and gap are well-defined, specifically addressing the limitations of ray-casting and existing unpaired methods that require paired masks.
2.The introduction of flow matching for pseudo-correspondence is a novel architecture that is well-founded and detailed clearly in the methods section.
3.The framework is validated against existing state-of-the-art models with favorable performance showing clear improvement.
4.An ablation study is included to identify the specific architectural contributions.
5.The figures of generated images are compelling and the synthetic images do appear to be of high quality. I am impressed by the fidelity and realism of the synthetic TEE as it is a difficult modality to simulate
- 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 results lack standard deviations on the reported metrics, which makes it difficult to assess the consistency of the performance.
2.There is no discussion of runtime or computational overhead, which is of interest for a simulation engine. No discussion of potential deployment, is this meant to be a real-time engine or a offline tool for pre-generating a synthetic dataset from the more data-rich CT domain?
3.While the perceptual quality is good, the paper does not show how these images affect downstream tasks, such as whether a system trained on these images generalizes better to real TEE than existing methods. A more thorough discussion of limitations and expected results for future experiments particularly in terms of generalization would strengthen the conclusions.
- 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 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
No additional comments
- 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?
This is a strong technical work with a good contribution. The methodology is interesting and shows a clear improvement over existing unpaired translation models. While there are some missing details regarding runtime and downstream generalizability, the overall quality of the synthesis and the well-defined clinical motivation justify an acceptance.
- 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.
My initial recommendation was accept. I agree with the comments from the other reviewers, and feel that the rebuttal sufficiently addresses the concerns raised. Requested clarifications were made, and the acknowledgement of lack of downstream task evaluation clear, which should be addressed as future work.
Review #2
- Please describe the contribution of the paper
This paper proposes a CT-to-TEE synthesis framework that uses flow matching in a modality-consistent latent space learned by an adversarial VQ-VAE pre-trained on unlabeled TEE data. The core contribution is Optimal Transport Coupled Flow Matching (OTC-Flow), which enables synthesis from unpaired data by computing entropic optimal transport (solved via Sinkhorn-Knopp) over an L2 cost matrix between CT condition embeddings and a dynamically updated TEE latent bank, then using discrete argmax assignment as pseudo-correspondence supervision for the flow matching velocity field. A pseudo-supervised anatomical consistency loop is introduced on top of OTC-Flow to provide further structural constraint. The authors evaluate the framework using three generative metrics and three perceptual and anatomical metrics, and report that OTC-Flow achieves top performance compared to existing SOTA models, with ablation studies isolating the contribution of each component.
- 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.This paper tackles a genuinely difficult unpaired synthesis problem with strong clinical motivation.
2.The architecture cleanly separates representation learning (VQ-VAE, pre-trained offline and frozen), correspondence finding (OTC, per mini-batch), transport learning (flow matching), and decoding (VQ decoder, inference only). Each stage has a clear responsibility, and this modularity means components can in principle be replaced or improved independently.
3.The choice of argmax assignment over barycentric projection is well-motivated: it preserves the discrete support of the VQ latent space and avoids feeding off-manifold blended vectors into the decoder. This is a thoughtful design decision.
4.Sinkhorn-Knopp is the appropriate efficient solver for entropic OT at this scale.
5.OTC-Flow achieves the best score on all six reported metrics across both generative and perceptual/anatomical categories, with a notably large margin on FRD (27.33 vs. 38.90 for the next best method). Consistent wins across diverse metric families are harder to attribute to evaluation choices than improvements on a single metric.
6.The ablation study isolating OTC-Flow and segmentation supervision versus the full pipeline supports the contribution of each component.
- 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.View diversity in the VQ-VAE training data is not characterized. The 81,510 frames may be dominated by a few standard TEE views — frame count alone does not establish view diversity. This has two compounding effects: (1) the VQ-VAE codebook may overfit to common views and poorly encode rare ones, and (2) because the latent bank is initialized from a frozen encoder and only updated by mini-batch enqueueing of training latents, any view bias present at initialization persists throughout OTC-Flow training. Per-view performance is also not reported, so it is difficult to assess whether the framework generalizes across the standard TEE view set.
2.The semantic validity of the L2 cost matrix is not isolated from the structural effect of OT. The reported gains over random and nearest-neighbor baselines are encouraging but do not isolate the source of improvement. OT marginal constraints enforce assignment diversity independently of the cost matrix’s semantic quality, so the improvement could reflect this structural regularization rather than meaningful anatomical similarity in the L2 embedding space. Given the CT–TEE domain gap, L2 distance between embeddings may primarily capture modality appearance rather than anatomical compatibility.
3.The downstream segmentation task used for Dice evaluation is not specified. Dice score is highly dependent on which structures are segmented and how the segmentation model was trained or labels were obtained. Without these details, the metric is difficult to interpret.
4.Sample counts are not reported in the comparison and ablation tables, which limits statistical interpretability of the results.
- 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
Questions for Authors:
1.What is the distribution across standard TEE views in the 81,510 training frames?
2.Could the authors provide qualitative examples of CT-TEE pairs assigned by OTC during training? This would help assess whether the cost matrix is producing anatomically coherent correspondences.
3.What segmentation method and target structures are used for the Dice evaluation?
4.Could the authors report the number of test samples used in the comparison and ablation tables?
- 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?
This paper addresses a clinically meaningful and technically difficult problem with a principled approach. OTC-Flow is a thoughtful integration of entropic optimal transport with flow matching in a modality-consistent VQ-VAE latent space. The architecture is cleanly modular, the evaluation combines generative and anatomical metrics appropriately, and the reported results show consistent best-in-class performance across all six metrics with a particularly large margin on FRD.
My concerns are about reporting completeness and mechanism isolation rather than soundness. View diversity of the VQ-VAE training data is not characterized, the gains over random and nearest-neighbor baselines do not isolate whether improvement comes from L2 cost matrix semantics or from OT marginal-constraint diversity, and the Dice evaluation and sample counts are underspecified. These are clarifications, not fundamental flaws, and most are addressable in the rebuttal or camera-ready without additional experiments.
The paper makes a real methodological contribution, the results consistently support its main claims, and the weaknesses do not undermine the central contribution. I recommend weak accept, with the expectation that the authors address the reporting gaps and ideally provide qualitative evidence for the OT correspondences in the rebuttal.
- Reviewer confidence
Somewhat confident (2)
- [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’ rebuttal adequately addresses my main concerns. In particular, the ablation numbers (random FRD 41.50 → NN 33.65 → OTC 27.33) partially isolate the contributions of L2 cost matrix preference and OT marginal-constraint structure, which was my primary technical concern. The Dice evaluation targets (MMWHS structures) and protocol are now clarified, and the explanation of the TEE training data sourcing (complete intraoperative videos from two centers, long-dwell repetitions removed) reasonably mitigates the view diversity concern. Minor gaps remain (segmentation model not specified, no qualitative OT pair examples) but these do not affect the core claims.
Review #3
- Please describe the contribution of the paper
This paper proposes unpaired CT-to-TEE image synthesis using a TEE-specific VQ-VAE latent prior combined with OT-Coupled Flow Matching (OTC-Flow) and a CT-derived anatomical consistency loop. Experiments compare against unpaired translation baselines on MMWHS CT and a 377-patient, two-hospital TEE cohort. Main contributions include:
- A generative TEE simulation model from CT slices, without paired data, pretrained on a fairly large dataset (80K+ TEE images).
- An optimal-transport-based pseudo-correspondence mechanism to align CT-conditioned latent features with the learned TEE latent distribution, replacing adversarial or diffusion-based alignment losses in prior art.
- Evaluated on a 377-patient, two-hospital TEE cohort with both generative quality metrics (FID, KID, FRD, LPIPS, NIQE).
- 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 targets a clinically meaningful problem: realistic TEE simulation is relevant for training, procedural planning, and robotic guidance, and the manuscript clearly motivates why paired CT-TEE supervision is difficult to obtain.
- The main technical idea is coherent and reasonably novel for this setting. Combining a TEE latent prior, OT-based pseudo-correspondence, and conditional flow matching is a thoughtful way to address unpaired CT-to-TEE translation.
- The use of a relatively large real TEE corpus for latent pretraining and domain matching is a substantive experimental asset, especially compared with many prior ultrasound synthesis works that rely on smaller curated datasets.
- 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.
Overall the paper is well organized. Here are my comments:
1.The central claim is patient-specific realistic TEE simulation, but the evaluation is still largely indirect. Distributional metrics such as FID/KID/FRD/LPIPS/NIQE do not by themselves establish that the generated image faithfully reflects the anatomy of the source CT, and the paper does not provide any evaluation on real annotated TEE dataset. For instance, can AI model trained on this synthetic dataset perform view classification on EchoNet TEE-view-classifier dataset?
2.The experimental protocol is underspecified in several places that matter for acceptance. In particular, the manuscript does not clearly describe patient-level train/validation/test splits, how leakage is prevented between VQ-VAE pretraining and evaluation, how many CT volumes are effectively used from the source domain, or how the Dice-based anatomical evaluation is constructed.
- 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.
(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 studied problem is clinically relevant and the overall direction is promising, particularly for procedural training and simulation. However, the validation of the proposed method is limited and primarily focuses on low-level image quality metrics (e.g., FID, LPIPS) and pseudo-label-based anatomical consistency. There is no evaluation on real annotated TEE data, expert assessment, or downstream clinical tasks. As a result, it remains unclear whether the generated images are clinically meaningful or usable in practice, beyond superficial visual realism.
- 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 thank the authors for the responses. I have no more questions.
Author Feedback
We thank all reviewers and focus on the main concerns bearing on the assessment.
R1.We appreciate the support. With same split, OTC-Flow ranked best on all six metrics (FRD 27.33 vs. 38.90). OT/Sinkhorn and latent bank are training-only; inference uses the CT encoder, latent transport, and frozen VQ-VAE decoder for planning/training/simulation.
R2-1.Thanks for pointing this out. The real TEE frames are complete intraoperative videos from two centers, not repeats of a few static standard views. They include key valve/chamber/intervention planes plus adjacent pose transitions. We retained de-identified key frames and removed long-dwell repetitions. Since clinical TEE uses continuous probe manipulation, frame-wise view labels require expert adjudication and can be ambiguous near transitions. Continuous conditions and CT-projected anatomy better fit this task.
R2-2.The L2 cost is computed between CT embeddings and the TEE latent bank used by OTC-Flow, shaped by the frozen TEE VQ-VAE prior and segmentation-regularized CT conditioning. Thus, it provides lower-noise target preference rather than raw-image similarity. Tab. 2 isolates the effects: random pairing removes cost-aware selection, while nearest-neighbor keeps local L2 preference without global OT balancing. Their FRD 41.50/33.65 vs. 27.33 indicates that L2 contributes target preference and OT improves balanced assignment.
R2-3/4.Dice evaluates CT-conditioned anatomical consistency: CT labels are projected by the same fan grid into pseudo-TEE labels, and synthesized-TEE segmentations are compared with projected CT labels. All baselines used the same protocol and official MMWHS targets (LV, LABC, LVBC, RABC, RVBC, AO, PA). The w/o Seg ablation worsens FRD/LPIPS, supporting this path. We used an 8:2 train/validation split, 20 MMWHS CT volumes, and 60K fan slices.
R3-1.Thanks for pointing out this confusion and for recognizing the value of TEE simulation in planning/training/simulation. Preoperative CT is routinely available, whereas pose-matched real TEE is acquired intraoperatively, making frame-wise annotations and strict CT-TEE pairs difficult to obtain. An unpaired framework generating TEE representations from preoperative CT is therefore a necessary basis. OTC-Flow couples CT fan-slice construction and CT-projected pseudo labels with OT-coupled flow matching to build CT-TEE latent correspondence. We hope this work brings this clinically needed but under-explored problem to the community and provides a reusable basis for expert assessment, training data generation, and navigation development.
The suggested EchoNet-style classifier is useful for view recognition, but TEE-only data without corresponding CT and paired TEE annotations cannot directly evaluate whether synthesized images preserve patient-specific anatomy from source CT. Our evaluation combines real-TEE distribution/perceptual metrics (KID/FRD/LPIPS/NIQE), CT-projected Dice, and qualitative comparisons in Fig. 3.These support improved realism and CT-conditioned structural consistency over unpaired baselines, not complete clinical deployability.
R3-2.For protocol and leakage, all comparisons used the same 8:2 train/validation split, with 20 MMWHS CT volumes and 60K CT fan slices. The 81,510 unlabeled TEE frames are used only for stage-I VQ-VAE pretraining to learn a TEE latent prior, without CT data or CT-TEE pairing; stage-II CT generation is separate from this prior-learning step, so there is no leakage between VQ-VAE pretraining and CT evaluation.
For Dice, CT labels are projected through the same fan grid into pseudo-TEE labels, and synthesized-TEE segmentations are compared with them. This evaluates source-CT structure preservation without real annotated TEE, consistent with prior anatomical-constraint practice. The w/o Seg ablation worsens FRD/LPIPS, supporting this constraint. Consistent with the abstract, code will be released upon acceptance for simulation-pipeline and Dice verification.
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.
All 3 reviewers recognize the approach as technically sound and novel, with a strong clinical motivation in addressing the inherent difficulties on obtaining paired CT and TEE data. The state-of-the-art performance across multiple metrics, and the use of large-scale TEE datasets with an interesting correspondence mechanism are noted as key strengths. However, one consistent concern raised by reviewers is the potential benefit of the proposed approach on downstream technical task/application that would demonstrate the clinical utility of the improved performance. There are also concerns regarding missing standard deviations to help assess statistical significance, and lack of discussion regarding the computational runtime. I encourage authors to provide a rebuttal to address reviewers’ 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 addressed the key concerns raised by the reviewers, particularly on the utility of proposed method on downstream technical task/application as well as clarifying the experiment protocol. While additional concern regarding runtime wasn’t directly addressed, the expected release of the source code for reproducibility would help address this concern. In summary, AC agrees with the reviewers and recognizes both the technical novelty of using flow matching for correspondence in latent space as well as significance of the proposed work in TEE simulation research, and thus, recommend acceptance.
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.
Reviewers are in favor overall of this work. While I miss a discussion on potential hallucinations in the synthesized TEE images, I still see merit in this work due to the strong coupling between modalities via segmentation masks. Reviewers acknowledge importance in clinical planning. Reviewer concerns regarding experimental setup were addressed well in the rebuttal in my opinion. I see this work as a meaningful contribution to be presented and discussed at MICCAI.
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.
This paper was reviewed by three experts in the field. Based on the reviewers’ post-rebuttal feedback, the paper is recommended for acceptance to MICCAI 2026.The authors are encouraged to incorporate the necessary changes discussed during the rebuttal process into the final camera-ready version.
