Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

Automatic ultrasound video object segmentation (VOS) is crucial for thyroid nodule diagnosis and treatment, which still remains challenging due to small nodules and blurred, noisy boundaries induced by tissue deformation and probe motion. Although recent diffusion VOS methods have shown superior performance by gradually reconstructing tumor structures from noise features, they typically treat heterogeneous lesion cues equally throughout the segmentation process, neglecting the fact that their contribution varies across different denoising stages. Motivated by the coarse-to-fine annotation logic of clinicians, we propose a Discriminative Synergistic Temporal Diffusion (DSTD) framework that can explicitly assess the step-wise contribution of motion and texture guidance during each denoising timestep. Specifically, a Retentive Motion Perception module is designed to capture temporally consistent motion patterns while suppressing irrelevant tissue noise, and a Discriminative Texture Perception module evaluates the confidence of different texture representations to enhance ambiguous boundaries. Furthermore, a Synergistic Controller is introduced to adaptively balance these complementary guidance cues in a timestep-dependent manner. Moreover, we construct a large-scale ultrasound thyroid nodule dataset with 374 pixel-level annotated videos. Extensive experiments on 3 medical video datasets demonstrate that DSTD consistently outperforms state-of-the-art methods, involving small nodules and blurred boundaries. We shall release our dataset, code, and results upon publication.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1074_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiJia_Discriminative_MICCAI2026,
        author = { Li, Jialu AND Wang, Hongqiu AND Hu, Junpu AND Xiao, Ruiqiang AND Hu, Ying AND Lv, Faqin AND Wang, Qiong AND Zhu, Lei},
        title = { { Discriminative Synergistic Temporal Diffusion for Ultrasound Video Thyroid Nodule 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

This paper proposes a Discriminative Synergistic Temporal Diffusion (DSTD) framework for ultrasound video thyroid nodule segmentation, which dynamically integrates motion and texture guidance cues during the diffusion denoising process. In addition, the authors construct a large-scale ultrasound thyroid nodule dataset containing 374 pixel-level annotated videos, and evaluate the proposed method on multiple medical video segmentation benchmarks.
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 method introduces a Retentive Motion Perception (RMP) module to capture temporally consistent lesion motion patterns, a Discriminative Texture Perception (DTP) module to enhance boundary representation using wavelet-based texture features, and a Synergistic Controller (SC) to adaptively balance these cues at different diffusion timesteps.
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.While the paper introduces a new dataset consisting of 374 annotated ultrasound videos, more details regarding the annotation protocol, inter-observer agreement, and dataset diversity would improve the clarity and reproducibility of the work. 2.The paper focuses mainly on architectural design and quantitative results, but provides limited analysis of why the proposed method works. Additional visualization or interpretability analysis would help better understand the contribution of the motion and texture guidance modules. 3.The experimental analysis could be further strengthened. In particular, more comparisons with recent diffusion-based or transformer-based medical video segmentation methods would make the evaluation more convincing. 4.More visulizations can be added to show the learned representations of the propsoed model.
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?

I recommend Weak Accept because the paper addresses an important clinical task and proposes a reasonable diffusion-based framework that integrates motion and texture cues, and the method achieves improvements over several recent methods. The introduction of a dataset with 374 annotated ultrasound videos is also a valuable contribution. However, the paper would benefit from clearer dataset details, deeper analysis of the proposed modules, and broader comparisons with recent methods.
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

Review #2

Please describe the contribution of the paper

This paper proposes a well-motivated approach that explicitly considers the heterogeneous effects of motion and texture cues over time to address complex medical video scenarios. Combined with a large-scale, pixel-level annotated thyroid nodule video dataset, this approach promises to make a valuable contribution to the field. Extensive experiments on multiple medical video datasets demonstrate that this method offers significant and sustained improvements over existing approaches, particularly in handling complex scenarios such as small lesions and poorly defined boundaries.
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.DSTD explicitly considers the contribution of heterogeneous cues over time. Compared to previous structures that directly and uniformly handle all guiding features, DSTD achieves higher segmentation accuracy across a variety of challenging clinical segmentation scenarios. 2.The proposed dataset containing 374 pixel-level annotated ultrasound videos is a valuable contribution and may contribute to future research. 3.This paper demonstrates sustained improvements over various existing techniques on multiple datasets, including complex cases with micronodules and irregular boundary shapes. 4.This coarse-to-fine diffusion structure enhances practical relevance and is well-suited for handling ultrasound-specific challenges such as micronodules and blurred boundaries.
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 dataset appears to be a high-quality, large-scale video dataset, the statistical analysis and diversity description of this ultrasound thyroid nodule video dataset in Section 1 of this paper are limited. 2.Some implementation details (e.g., hyperparameters or module settings) could be further clarified to improve reproducibility. 3.While the method is effective and well-motivated, this work would be more valuable if further analysis were conducted, such as analyzing the respective roles of different wavelet bases in the DTP module. 4.The collaborative controller sounds interesting, but could be further clarified, for example, by more explicitly comparing it with existing fusion methods based on gating or attention mechanisms.
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.

(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 paper provide a clear motivation, a rigorously designed methodology, and comprehensive performance testing on multiple medical datasets. The proposed stage-adaptive diffusion framework effectively simulates the effects of motion and texture cues over time, aligning closely with clinical practice and improving performance in complex scenarios. The proposed large-scale labeled dataset makes a valuable contribution to the research community. This method demonstrates sustained advantages over various existing approaches and is supported by extensive ablation experiments.

However, several limitations were noted: limited statistical/diversity analysis of the dataset, missing implementation details affecting reproducibility, insufficient analysis of wavelet bases in the DTP module, and lack of explicit comparison with existing gating/attention fusion methods for the collaborative controller.

Despite these shortcomings, the core technical novelty, clinical relevance, and the valuable public dataset outweigh the issues. The weaknesses are addressable in a revision and do not undermine the paper’s overall impact.
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 have solved all my questions. Given the method novlety and the contribution of high-quality dataset to related research, I recommend for acceptance.

Review #3

Please describe the contribution of the paper

It proposes a video segmentation framework, DSTD, that attempts to adaptively weight motion and texture cues across diffusion timesteps for ultrasound thyroid nodule segmentation.
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 a highly pertinent and notoriously difficult challenge in medical image analysis: segmenting small thyroid nodules with blurred boundaries in ultrasound videos. The problem statement is well-defined and practically significant. The authors demonstrate consistent and solid quantitative improvements (Dice, Jaccard, etc.) across multiple datasets, including their private dataset and two public medical video benchmarks (DPSTT and SUN-SEG). The results indicate a robust generalization capability. Compiling and pixel-level annotating 374 ultrasound videos is a highly labor-intensive process. Even pending clarification on its relationship to prior releases, this represents a valuable resource for the medical AI community.
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 paper introduces two major modules: Retentive Motion Perception (RMP) and Discriminative Texture Perception (DTP). However, in ultrasound video, consistent motion patterns and texture representations are often inextricably linked. The paper fails to prove that these are truly heterogeneous cues rather than two different mathematical ways of looking at the same pixel intensity changes. The SC is essentially an additional attention/weighting layer. The paper argues it mimics clinical logic, but provides no proof that the model actually focuses on motion first and texture later. It might simply be learning to ignore noisy features regardless of their types. 2.KANs are a recent trend, but the authors do not justify why a standard MLP or attention mechanism wouldn’t suffice. Without a specific ablation comparing them to standard components, it is not clear whether the insertion of this layer is necessary or a complexity bloat. 3.The paper uses Haar, DB3, and Dmey wavelets, but does not provide any explanation as to why these specific kernels were chosen or why three is the optimal number. If the model is truly “adaptive,” it should theoretically work with any bank of filters; without this proof, the choice feels like manual hyperparameter tuning disguised as discriminative perception. 4.The authors highlight the release of a 374-video thyroid ultrasound dataset as a primary contribution. However, a highly relevant existing dataset for this specific task, VTUS (100 videos), was recently introduced alongside the ViVim framework (Yang et al., TCSVT 2025). First, it is crucial to clarify the relationship between this newly proposed dataset and VTUS. Does the 374-video dataset include the 100 videos from VTUS? If so, the claim of releasing a ‘new’ 374-video dataset needs to be appropriately contextualized to accurately reflect the incremental contribution. Alternatively, if the two datasets are indeed entirely distinct with no overlap, then the authors have omitted a critical external validation. Since this submission focuses specifically on thyroid nodule segmentation, relying solely on general medical video datasets (breast BUV2022 and polyp SUN-SEG) for external validation—while ignoring the most relevant domain-specific public dataset—leaves a significant gap in demonstrating the model’s true generalizability. 5.The paper claims real-time performance for a diffusion-based framework. However, diffusion models are typically characterized by high iterative latency compared to linear-time models like Mamba-based models (e.g., Vivim). The authors must report the exact Frames Per Second (FPS) and Inference Memory usage to demonstrate a significant speed advantage or comparable efficiency to existing works likes Vivim. 6.The Baseline removes all memory-related enhancements. This sets an artificially low bar.8.Please rate the clarity and organization of the paper.
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?

This submission addresses a highly relevant clinical problem and proposes a logically sound framework (DSTD) that intuitively mirrors the clinical diagnostic workflow. The reported quantitative segmentation results are also promising. However, the current manuscript contains critical gaps in experimental validation and clarity that prevent me from recommending acceptance at this stage. First, the lack of inference efficiency metrics (FPS and memory consumption) is a significant omission, especially considering the known latency challenges of diffusion models compared to baseline architectures like ViVim. Second, the architectural introduction of KAN layers and specific wavelet kernels lacks proper ablation to prove their necessity over simpler, standard components. Finally, the relationship between the newly proposed dataset and the existing VTUS dataset remains unclarified, and the model lacks validation on the public VTUS benchmark. Therefore, my initial recommendation is a Weak Reject. However, I consider this a borderline paper and I am open to raising my score to an Accept if the authors can satisfactorily address these key issues in their rebuttal.
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.

N/A

Author Feedback

We appreciate the favorable comments on our work (R1, R2), and the high-quality dataset contribution (R1, R2, R3). Below, we clarify all the issues. R1 1[Dataset]Thanks, we have added more details. 5 clinicians annotated using ITK-SNAP, with each frame annotated 3 times and pixel-wise majority voting to reduce observer variation. Our data aged 25–78 years with goiter, follicular adenoma, Hashimoto’s thyroiditis, papillary carcinoma. 2[Visualization]We have added visualizations of RMP motion guidance, DTP texture guidance to show that RMP focuses on temporally consistent tumor candidate regions, DTP highlights wavelet-domain tumor boundary areas. 3[Comparison]We have added recent transformer/SAM/diffusion-based comparisons: HRVVS, MedSAM2, and Diff-VPS achieve 62.54/73.08/79.47/73.54, 63.69/74.38/81.14/74.71, and 61.68/71.93/82.77/70.49 in Jac/Dice/Pre/Rec, while our method still performs better. 4[Visualization]We have visualized the learned representations of RMP, DTP, and SC across timesteps, showing that RMP learns temporally consistent tumor candidate regions, DTP captures wavelet-domain boundary textures, and SC dynamically balances their contributions across timesteps. R2 1[Statistics]Thanks. We have enriched Sec.1 and updated Fig.1 with age/pathology distributions, showing that the dataset covers patients aged 25–78 years, with an average age of 49.8 years, and diverse thyroid diseases. 2[Details]Our model is trained using BCE and Dice loss with a weight of 1:1 on 1 RTX4090 and EPYC 7763 CPU. The diffusion timestep and K are empirically set to 5 and 3. 3[Wavelet]DTP supports different wavelet kernels and kernel numbers for multi-scale and multi-orientation texture modeling. Using 4 and 5 kernels achieves 69.04/79.21/84.72/79.88 and 69.75/79.55/86.19/78.75 in Jac/Dice/Pre/Rec, respectively, showing improved adaptability to varied thyroid nodule morphologies. 4[Module]We have replaced SC with normal attention, the Jac/Dice/Pre/Rec are degraded to 67.16/77.91/83.17/78.55, showing the effectiveness of our SC. R3 1[Heterogeneous]Thanks.1)RMP and DTP use different input frames and feature domains. RMP inputs adjacent memory frames to extract temporally consistent motion features, while DTP only selects top-K similar frames (Eq3,Sec2.2) to extract wavelet-domain texture features. Memory size and K are empirically set to 8 and 3.2)[CVPR-Qian-Boosting Diffusion Model; WACV-Lee-Beta sampling] demonstrated the coarse-to-fine behavior of diffusion itself, SC only dynamically assists its behavior. We added visualizations of RMP, DTP, and SC maps, showing that RMP focuses on temporally consistent candidate regions, DTP highlights tumor-related boundaries, and SC assigns higher motion weights at early timesteps but higher texture weights at later timesteps. This supports the complementary roles of RMP/DTP and the dynamic balancing effect of SC. 2[KAN]Replacing KAN with MLP and attention yields the degraded Jac/Dice/Rec of 67.31/77.68/75.83 and 68.02/78.13/77.87.We also clarify that KAN is not our key contribution, but an implementation choice in DTP. 3[Wavelet]We empirically set 3 kernels to balance performance and efficiency. DTP is compatible with different kernel numbers, where 4 and 5 kernels achieve Jac/Dice/Pre/Rec of 69.04/79.21/84.72/79.88 and 69.75/79.55/86.19/78.75, indicating diverse wavelet representations improve adaptability to varied thyroid nodule morphologies. 4[Dataset]Our dataset is entirely independent from VTUS. Our model achieves better Jac/Dice/Pre scores of 75.42/84.58/89.01 on VTUS and better Jac/Dice/Pre/Rec (Tab.1) on our dataset. 5[Efficiency]We clarify that “real-time” in Sec.1 refers to ultrasound imaging as a modality, not the inference speed of DSTD(15.73 FPS, 8.96GB infer memory). 6[Baseline]Apologize for mistakenly using “removing…with” instead of “replacing…with” in Sec.3.We clarified that Baseline (Tab.4) has the memory block, and it can be checked by the ‘Memory’ column in Tab.4.

Meta-Review

Meta-review #1

Your recommendation

Invite for Rebuttal
Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

The reviewers generally agree that this work addresses a clinically important and challenging problem in ultrasound video segmentation. The proposed diffusion-based framework is considered well motivated, and the reported quantitative improvements across multiple datasets are viewed positively. In addition, the introduction of a large pixel-level annotated ultrasound video dataset is regarded as a potentially valuable contribution to the community.

However, there are still several points that need to be clarified, including but not limited to:

1.The authors should provide clearer details on the newly introduced 374-video dataset, including annotation protocol, inter-observer agreement, diversity/statistics, and clarify its relationship (if any) to existing datasets such as VTUS.

2.Several architectural choices require stronger support, such as KAN layers and wavelet kernels. Additional ablation or clearer justification will help to make this convincing.

3.Given the diffusion-based design, inference efficiency should be reported more explicitly (e.g., FPS, memory usage), particularly in comparison with existing efficient video segmentation models.

4.Implementation details remain insufficiently specified and should be clarified to improve reproducibility.
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 overall positive and support acceptance. The rebuttal addressed the main concerns about the method design and clarified the contribution of the dataset, and the post-rebuttal reviewers who updated their opinions are supportive. One reviewer did not submit a post-rebuttal update, so this conclusion partly relies on the original recommendation, but the overall direction remains positive.

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.

I have read the manuscript, review comments, rebuttal letter. All three reviewers recommend acceptance (after rebuttal). The responses addressed the concerns about dataset details, ablation of architectural, and inference costs.

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 author’s response addresses the reviewers’ concerns.

back to top

Discriminative Synergistic Temporal Diffusion for Ultrasound Video Thyroid Nodule Segmentation

Author(s):