Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Real-time video segmentation of the prostate in Transrectal Ultrasound (TRUS) is essential for image-guided interventions. While conventional 2D methods suffer from inter-frame inconsistencies by disregarding temporal context, 3D architectures incur prohibitive latency. To resolve this dilemma, we present a Temporally Consistent Learning Framework that distills temporal coherence into a 2D network during training, preserving single-frame inference efficiency. Our design is driven by a key clinical observation: the prostate exhibits geometric stability, whereas the surrounding acoustic environment fluctuates due to physiological motion and transducer pressure. Because conventional temporal constraints propagate erroneous gradients from these unstable regions, we introduce a Confidence-Weighted Temporal Consistency objective derived from optical flow warping residuals, selectively attenuating contributions from unreliable regions. Complementing this pixel-wise constraint, a Dual-scale Prototype Alignment Module enforces semantic coherence through contrastive optimization of local boundary and global semantic features. Furthermore, to eliminate the need for dense per-frame video annotations, we employ geometric equivariance-based pseudo-labeling with knowledge distillation from a pretrained teacher. Extensive experiments on SUN-SEG and our newly introduced TRUS-V benchmark (2,679 frames) demonstrate state-of-the-art accuracy and temporal consistency at real-time speed. Code and dataset are available at https://github.com/DYDevelop/DTC-TRUS.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0460_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/DYDevelop/DTC-TRUS

Link to the Dataset(s)

https://khdp.net/database/data-search-detail/TRUS-V

BibTex

@InProceedings{KimDon_Distilling_MICCAI2026,
        author = { Kim, Dong Yeong AND Lee, JunGyu AND Choi, Jaewon AND Seo, June Young AND Kim, Myeongseop AND Choi, Jinwook AND Kim, Taek Min AND Kim, Young-Gon},
        title = { { Distilling Temporal Coherence Into 2D Networks for Transrectal Ultrasound Prostate Video Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16883},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper
    1. Propose a temporal coherence distillation framework that injects temporal knowledge into 2D networks, balancing segmentation stability and real-time inference.
    2. Design confidence-weighted temporal consistency and dual-scale prototype alignment to suppress flicker and resist ultrasound speckle noise.
    3. Release TRUS-V, a multi-view TRUS prostate video segmentation benchmark with 2,679 annotated frames.
    4. Adopt geometric equivariance pseudo-labeling and teacher-student distillation to reduce video annotation costs.
  • 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 work proposes a self-supervised temporal consistency framework by using knowledge distillation and contrastive learning. It distills temporal knowledge into a 2D network during training, enabling pure single-frame inference without runtime overhead. This elegantly resolves the trade-off between 2D flickering and 3D high latency.
    2. The authors introduce a multi-view transrectal ultrasound prostate video benchmark with 2679 annotated frames, which can be an important contribution in public datasets for this clinical task. The authors claim code and dataset will be open-sourced to support reproducibility.
    3. Experiments on SUN-SEG and TRUS-V validate the method. The approach achieves state-of-the-art results with an ultra-fast inference speed of 127.97 FPS, showing strong clinical potential.
  • 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.Although the method achieves competitive results on SUN-SEG, the Dice on SUN-SEG-Easy is still lower than DALA. It would strengthen the empirical study if more recent and relevant baselines on SUN-SEG were included. Besides, there have been some other methods you didn’t compare with have even higher dice on SUN-SEG like LGRNet. Maybe you should compare more new methods on SUN-SEG, such as SAM2-based and SAM3-based to verify your excellent performance. 2.The paper should clarify the exact training protocol on SUN-SEG. Since the method is described as a self-supervised Student-Teacher framework trained on unlabeled videos, it is currently unclear whether the SUN-SEG annotations are used only for evaluation, or also for teacher pretraining / supervision. 3.The ablation study is only conduct on SUN-SEG (Easy) dataset. I think the ablation study on your own dataset is also important as you propose your method in the specific scene of prostate ultrasound video segmentation. In addition, the current ablation does not clearly explain the “partial usage” setting of the prototype module; a brief clarification, or a simple comparison between single-scale and dual-scale usage, would make the design choice more convincing. 4.TRUS-V is a useful contribution, but its current scale is still limited to 10 patients from a relatively narrow acquisition setting. It would be helpful to discuss this limitation more explicitly and clarify whether future extensions with more patients or acquisition diversity are planned. 5.Although the method does not use optical flow at inference, its temporal supervision is still derived from flow-based alignment during training. Therefore, the learned 2D model may implicitly encode scene-specific temporal regularities of the current TRUS setting. Since TRUS-V mainly contains paired axial/sagittal views from the same acquisition scenario, it would be helpful to clarify whether the learned temporal prior generalizes across views or acquisition settings, rather than being tied to a specific motion pattern.

  • 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?

    The results are based on a comprehensive consideration of the author’s methodological innovation, dataset contribution, and writing quality.

  • 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 rebuttal well addressed my concerns. Hence, I maintain my positive score on this work.



Review #2

  • Please describe the contribution of the paper

    The paper presents a framework for real-time Transrectal Ultrasound (TRUS) prostate video segmentation. The framework proposes a distillation framework that embeds temporal coherence into a 2D network during training. This resolves the trade-off between the flickering artifacts of 2D models and the high latency of 3D networks.

  • 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 motivation to maintain temporal stability within efficient 2D networks serves as a good research motivation. 2.Despite space limitations, the authors have conducted a relatively comprehensive experimental validation. 3.The introduction of a novel prostate segmentation dataset is a highly valuable contribution that effectively addresses the critical scarcity of this data type in the current field.

  • 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 introduction of the video segmentation paradigm in this paper is not comprehensive in the background description. Apart from 3D network design, many works [1,2] have adopted a network design that combines a 2D feature extractor with a temporal feature interaction module, and also have achieved good running speeds. However, this paradigm is not clearly mentioned in the Sec. Abstract & Introduction of the paper. 2.The experimental section omits comparisons with some recently advanced VPS work, such as SALI [1,2]. Even if the proposed framework does not outperform these advanced models in terms of accuracy metrics, the authors should still include them for an objective evaluation. It would be beneficial to analyze other potential advantages of the proposed framework compared to these advanced models, such as inference speed, memory efficiency, etc. 3.As shown in the ablation study (Table 3), employing a single-scale strategy for $\mathcal{L}_{proto}$ actually leads to a slight performance drop (Dice: 0.721 vs. 0.722), whereas the dual-scale strategy yields significant improvements (Dice: 0.735 vs. 0.701). It is necessary to provide a more detailed analyses to explain this phenomenon. Why does the single-scale fail, and how do the dual-scales produce such a substantial performance gain? 4.The manuscript is unclear about the actual data usage during training. In the Sec. Introduction, the authors wrote “To circumvent the prohibitive cost of dense video annotation”. However, the experimental datasets that are fully annotated (e.g., SUN-SEG). The authors should more clearly clarify: Did the training process utilize all the dense video annotations provided in the datasets? If so, how does the current experimental setup demonstrate the claim of “circumvent the prohibitive cost of dense video annotation”?

    [1] HilbertMamba: Local-Global Reciprocal Network for Uterine Fibroid Segmentation in Ultrasound Videos. MICCAI-2024.[2] SALI: Short-Term Alignment and Long-Term Interaction Network for Colonoscopy Video Polyp Segmentation, MICCAI-2024.

  • 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?

    Please see the 7-th point.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    1.The author did not respond to my first comment. The introduction section did not provide a comprehensive overview of the existing video segmentation methods, and it neglected to introduce some of the existing approaches that utilize 2D feature extraction networks combined with temporal feature fusion modules.

    2.The author claimed that this work could effectively meet the demand for dense video annotation. However, the experiments conducted did not verify this claim. Moreover, there was no comparison with some existing SOTA video segmentation methods, such as LGRNet and SALI. Given that the manuscript cannot include any additional experimental results, I am inclined to reject this article.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a temporally consistent learning framework for real-time video segmentation in transrectal ultrasound (TRUS). The main contribution lies in distilling temporal consistency into a standard 2D segmentation network during training, thereby avoiding the computational overhead associated with 3D or video-based models at inference time. Specifically, the method introduces: (1) a confidence-weighted temporal consistency loss based on optical flow to enforce pixel-level alignment across frames; (2) a dual-scale prototype alignment module to enhance feature-level semantic consistency; and (3) a self-supervised learning strategy that combines geometric equivariance and knowledge distillation, enabling training without dense frame-wise annotations. The framework aims to achieve temporally stable predictions while maintaining real-time inference efficiency. The approach is validated on the public SUN-SEG dataset and a newly introduced TRUS-V benchmark consisting of 2679 frames, and the code and dataset will be made publicly available.

  • 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 proposed framework integrates pixel-level temporal consistency, feature-level prototype alignment, and self-supervised learning. These components are intuitively complementary, addressing geometric alignment, semantic robustness, and the scarcity of supervision, respectively. 2.Introduction of a TRUS video dataset.

  • 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 TRUS-V dataset is relatively small (10 patients), which limits the strength of conclusions regarding generalization. In addition, the manuscript does not clearly describe how the 10 patients are split into training and testing sets. Furthermore, key properties of the video sequences (e.g., frame rate, average sequence length) are not provided, making it difficult to understand their spatiotemporal characteristics. (2) The proposed method leverages unlabeled video data and self-supervised learning, while many baseline methods are trained in a fully supervised manner. This discrepancy in training settings raises concerns about the fairness of the comparisons. (3) Important implementation details are missing, such as loss weights and training hyperparameters. (4) In Section 3.5, the explanation following Equation (5) refers to lambda_s, which does not explicitly appear in the equation (5). (5) The segmentation results of DALA in Fig. 4 appear as grayscale maps rather than binarized masks, whereas other methods are shown with binary outputs. The reason for this discrepancy should be clarified. (6) The method section is not sufficiently clear, and the overall workflow of the proposed approach is not explicitly explained.

  • 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

    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?

    This paper addresses an important problem and proposes a framework with practical value. However, the experimental design and evaluation do not sufficiently support the central claims regarding temporal consistency. Strengthening the evaluation with dedicated temporal metrics, clarifying the training setup, and improving the coherence of the methodological presentation would significantly enhance the quality of the work.

  • 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.

    This paper addresses an important problem and proposes a framework with practical value.



Author Feedback

To AC and Reviewers: We thank the AC and reviewers for their comments. We have addressed the concerns regarding annotation, baselines, and clinical rigor below. All revisions will be fully incorporated into the final manuscript.

Clarification on Labeling Protocol (R1-R3) To clarify our protocol, we emphasize that our framework does not require any frame-wise annotations. The training follows a two-stage process: (1) Static Pre-training: An initial 2D teacher model is trained using only 10-15% random sparse image-level labels. (2) Label-free Adaptation: No additional annotations are used during the subsequent video adaptation process. Instead, we leverage self-supervised equivariance and distillation to empower the 2D model with temporal robustness. This approach effectively elevates static 2D models into robust video architectures without the prohibitive cost of video-domain labeling.

Asymmetric Comparison with Baselines (R1, R2) In our preliminary evaluations, representative baselines (SALI, HilbertMamba, LGRNet, SAM 2/3) showed comparable performance. Although these baselines were omitted from the original manuscript due to strict page constraints, we emphasize that such direct comparisons are fundamentally unfair, as these methods rely on extra temporal labels and heavy interaction modules. While such models may offer marginal gains via extra temporal supervision, they operate at a much lower throughput of 10 to 40 FPS. In contrast, our method achieves 127.97 FPS, satisfying the high-speed requirements of TRUS-guided surgery. Acknowledging the reviewers’ feedback, we will incorporate these additional comparisons and an extended related work section in the final manuscript.

Pioneering the TRUS-V Temporal Benchmark (R1, R3) Prior to our work, a dedicated temporal benchmark for TRUS-guided surgery was non-existent. To address this critical gap, TRUS-V uses a strict patient-level split (9 train / 1 test), yielding 2,405/274 frames across 20 axial/sagittal videos. Although limited to 10 patients, every single frame was exhaustively manually refined by an experienced radiology professor. To ensure diversity, the dataset covers major anatomical variations across both axial and sagittal planes. This rigorous clinical validation establishes TRUS-V as a highly reliable initial benchmark for evaluating real-time temporal consistency.

Prototype Ablation (R1, R2) The single-scale alignment, denoted as △, is suboptimal because local prototypes are noise-sensitive, while global prototypes lack boundary precision. Our dual-scale design combines global semantic stability with local boundary refinement via area-adaptive weighting, improving both structural consistency and mask overlap, as reflected by Sα (+0.08) and Dice (+0.025) in Table 3.To ensure fair generalization evaluation, we conducted these ablations on the public SUN-SEG benchmark rather than our private TRUS dataset.

Temporal Prior (R1) We utilize optical flow to acquire inter-frame temporal information without requiring any additional temporal supervision. Rather than inducing overfitting to specific probe trajectories, this mechanism serves as a training regularizer to distill artifact-invariant and anatomically stable representations. Our robust performance across both the TRUS-V and SUN-SEG datasets empirically demonstrates that our framework captures fundamental structural consistency rather than memorizing specific motion biases.

Implementation Details & Visualization (R3) We will revise Sec. 3 to detail the workflow, hyperparameters (AdamW, 1e-6, batch 8, 1 epoch), and explicit loss weights (1.0, 1.0, 0.25, 1.0) in Eq. (5). In Fig. 4, all rows display raw confidence maps rather than binarized masks. DALA’s grayscale appearance reflects uncertain predictions, whereas our model produces distinctly sharper activations. Crucially, all metrics were computed using an identical thresholding protocol to ensure strict fairness.



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.

    Across the reviews, there is a consistent appreciation for the practical motivation of the work. The idea of transferring temporal coherence into a lightweight 2D model is seen as a sensible way to balance stability and efficiency, and the dataset contribution is viewed as a useful addition, especially given the limited availability of TRUS video data. That said, the main concerns are less about the core idea and more about how convincingly it is validated. Reviewers repeatedly point out gaps in benchmarking particularly the absence of stronger or more recent video segmentation baselines and some ambiguity around the training protocol (e.g., the role of annotations and fairness of comparisons). The relatively small dataset and limited diversity also raise questions about generalizability. In addition, parts of the method and ablation analysis would benefit from clearer explanations and more thorough justification. My reading is largely in line with these points. The direction is interesting and practically relevant, but the current version leaves some uncertainty around how strong the empirical support really is. Addressing the evaluation gaps and clarifying the setup would make the contribution more convincing.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The paper addresses an important and clinically relevant problem: achieving temporally stable TRUS video segmentation while maintaining real-time inference efficiency. Reviewers generally agreed on the practical value of the proposed temporal distillation framework and the usefulness of the newly introduced TRUS-V dataset. The rebuttal successfully clarified several key concerns, including the annotation protocol, training procedure, dataset split, implementation details, and the rationale behind the dual-scale prototype alignment. These clarifications strengthen the validity of the proposed approach and make the experimental setup substantially clearer. Two reviewers maintained or strengthened their positive assessments following the rebuttal.

    One reviewer remained concerned about the completeness of the related work discussion and the absence of direct comparisons with several recent video segmentation methods. While these concerns are reasonable, they do not appear sufficient to outweigh the overall contribution, particularly given the method’s emphasis on achieving temporal consistency with lightweight 2D inference and its demonstrated efficiency advantages. Overall, I find the contribution technically sound, practically relevant, and supported by adequate empirical evidence. The paper presents a meaningful methodological contribution and introduces a valuable benchmark resource for the community



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.

    This paper addresses a practical clinical challenge in real-time TRUS prostate video segmentation by balancing the trade-off between the inter-frame inconsistencies of lightweight 2D networks and the high computational latency of 3D networks. I agree with the reviewers’ consensus regarding the method’s practical significance and the valuable contribution of the newly released, expert-annotated TRUS-V dataset. In the rebuttal, the authors resolved primary concerns regarding comparison baselines, evaluation dataset scale, labeling protocols, and general clarity issues. Regarding the absence of certain recent benchmarking baselines, I agree with the author’s rebuttal that the proposed method is emphasizing more on the high throughput (>120 fps) while maintaining a competitive performance against heavier video models. Overall, I think the practical soundness and open-source data contributions place this work in a strong position, and thus, I recommend accept.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    I recommend rejection. The paper addresses an important clinical problem, but the contribution appears limited. The method mainly combines existing ideas, including teacher-student distillation, equivariance-based pseudo-labeling, optical-flow temporal consistency, and prototype alignment, without clearly establishing sufficient novelty.

    The experimental comparison is also not convincing. Although some video baselines are included, the comparison with strong recent SOTA video segmentation methods is still limited. More importantly, the paper claims temporally consistent video segmentation, but the evaluation relies mostly on frame-wise segmentation metrics. Direct temporal stability metrics are needed to support this claim.

    Finally, the core assumption is unclear. If the model uses temporal information only during training but performs single-frame 2D inference, the paper should justify under what conditions such a model can reliably produce temporally consistent results. Without this clarification, the main claim is not fully convincing.

    Overall, due to limited novelty, insufficient validation, and unclear presentation, I suggest reject.



back to top