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

Abstract

Ultrasound imaging is widely used for disease diagnosis in clinical practice due to its non-invasive and portable advantages. However, current deep learning-based models are predominantly designed for specific organs or tasks, leading to performance with limited cross domain generalization and insufficient exploitation of available ultrasound data such as ultrasound images and videos. To address this limitation, we propose UIV-US, a unified multi-organ and multi-task framework for ultrasound image and video understanding across B-mode and contrast-enhanced ultrasound (CEUS). UIV-US operates within a shared parameter space to jointly support segmentation and classification tasks across both images and videos, enabling consistent representation learning across modalities and temporal settings. We design a Structured Semantic Conditioning Prompt (SSCP) that comprises five semantic attributes grouped into structural and spatial priors for adaptive conditioning of shared features, forming the foundation for scalable learning from large-scale ultrasound datasets. To further bridge image and video reasoning, a Memory-Reinforced Temporal Attention (MRTA) is introduced to ensure temporal consistency for video segmentation, while a segmentation-aware Feature-Level Attention (FLA) strategy performs reliable temporal aggregation for video classification. Extensive experiments on a large-scale multi-organ benchmark comprising 27,804 segmentation images, 7,320 classification images, and over 1,400 ultrasound videos demonstrate the effectiveness of UIV-US. The proposed framework achieves 89.81% Dice Similarity Coefficient for image segmentation, 96.28% for video segmentation, and 89.26% for CEUS segmentation. For classification, it reaches 85.76% Area Under the ROC Curve on image, 80.32% on video and 72.16% on CEUS, demonstrating strong performance across tasks.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1246_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{HuaQia_Towards_MICCAI2026,
        author = { Huang, Qiang AND Jia, Dengqiang AND Chen, Xiangmei AND Lin, Zehui AND Wang, Zhizhou AND Cui, Ligang AND Wang, Zhiyuan AND Tan, Tao},
        title = { { Towards Unified Image–Video Modeling Across B-Mode and Contrast-Enhanced Ultrasound for Multi-task Analysis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16878},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes UIV-US, a unified multi-task foundation model for ultrasound analysis that bridges the gap between image and video modeling across B-mode and Contrast-Enhanced Ultrasound (CEUS). The authors introduce a Structured Semantic Conditioning Prompt (SSCP) to adaptively modulate features for diverse organs and tasks, alongside a Memory-Reinforced Temporal Attention (MRTA) and Feature-Level Attention (FLA) to ensure temporal consistency in video segmentation and classification, respectively.

  • 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 paper tackles a genuinely challenging problem – unifying image/video, segmentation/classification, and B-mode/CEUS modalities within a single model. The experimental scale is impressive, covering 27,804 segmentation images, 7,320 classification images, and over 1,400 videos across multiple organs. This level of comprehensive evaluation is rare and valuable.
    2. The design of SSCP is well-reasoned. By grouping five semantic attributes into structural and spatial priors and injecting them at different network depths, the model effectively handles heterogeneity without requiring task-specific branches or model duplication.
    3. Beyond standard performance metrics, the paper includes a detailed ablation study and a qualitative kNN purity analysis (Fig. 3). The kNN analysis is particularly interesting as it visually demonstrates the transition of features from domain-aware (encoder) to task-aligned (decoder).
  • 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.

    Major:

    1. The paper claims that “most existing approaches remain predominantly image-oriented or segmentation-focused, and explicit unification of image and video reasoning… remains limited. ” However, several prior works have already proposed unified or promptable frameworks for ultrasound that handle multiple tasks and modalities. USFM is cited but not further compare against it in the main experiments, nor explain how UIV-US differs or improves upon USFM. Other representative work like URFM, UltraFedFM, USF-MAE are not discussed.
    2. The paper states that the source code will be made available, but many details are omitted: What is the backbone architecture? Is it based on SAM, ViT, or a custom design? How are the five semantic attributes (Task, Target, Mode, Modality, Position) obtained for each sample? Are they manually annotated, derived from dataset metadata, or inferred? This is critical for reproducibility.
    3. MRTA take the segmentation results of previous frames as reference attention. If the results of previous frames contain significant errors, will error accumulation occur in subsequent segmentation processes, and are there any methods to eliminate this risk?
    4. While the model uses a “shared parameter space,” the introduction of memory banks (MRTA) with a length of up to K=10, and cross-attention mechanisms may introduce significant computational overhead during inference. The paper lacks a discussion or quantitative analysis of inference speed (FPS) or memory usage, which is crucial for the claimed portable advantages of ultrasound.

    [1] URFM: a general Ultrasound Representation Foundation Model for advancing ultrasound image diagnosis [2] From pretraining to privacy: federated ultrasound foundation model with self-supervised learning [3] USF-MAE: Ultrasound Self-Supervised Foundation Model with Masked Autoencoding

  • 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

    Please refer to the weaknesses part.

  • 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 and clinically relevant problem: building a single foundation model that can handle the diverse data types encountered in ultrasound examinations. The scale of the experimental validation is substantial, and the results demonstrate consistent improvement over several task-specific baselines. The SSCP conditioning mechanism and the lightweight temporal modules are well-designed and supported by ablation studies. However, the paper has several significant weaknesses that prevent a higher score. The novelty against ultrasound foundation model, the computational cost, and the implementation details are not discussed. Despite these issues, the work represents a solid engineering effort and a useful contribution to the field of ultrasound image analysis. With a clearer explaination and revision, the paper could become a accept.

  • 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.

    UIV-US has made substantial technical contributions to the important research direction of unified multi-task ultrasound modeling and possesses distinct clinical value. The systematic design of its methods and comprehensive experimental validation meet the acceptance criteria for methodological papers at MICCAI. The authors have provided sufficient and satisfactory responses to all reasonable questions in the rebuttal.

    If possible, it is suggested that the authors make the following revisions in the final version: 1.Supplement comparisons with at least one or two reasonable baseline methods for CEUS and video classification tasks. 2.Add confidence intervals or standard deviations from multiple runs to the main experimental results. 3.Present detailed dataset-level results in the Supplementary Material.



Review #2

  • Please describe the contribution of the paper

    The main contribution of this paper is the proposal of UIV-US, a unified ultrasound framework that jointly handles image and video analysis, B-mode and CEUS modalities, and both segmentation and classification tasks within a shared-parameter architecture. To support this unification, the authors introduce a Structured Semantic Conditioning Prompt (SSCP) to inject task-, target-, modality-, and position-related priors into shared features, and further design MRTA for temporally consistent video segmentation and FLA for segmentation-aware temporal aggregation in video classification. The paper therefore contributes a more general and scalable modeling framework for multi-organ, multi-task ultrasound analysis, supported by extensive experiments on a large heterogeneous benchmark.

  • 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 has several major strengths. 1.Meaningful unified formulation. The work addresses an important limitation in current ultrasound analysis, namely that most existing methods are organ-specific, task-specific, or developed separately for images and videos. In contrast, this paper proposes a single shared-parameter framework that jointly supports segmentation and classification across image, video, B-mode, and CEUS settings. This unified design is interesting because it moves beyond narrowly specialized pipelines and is closer to how heterogeneous ultrasound data are encountered in practice. 2.Well-motivated method design. The proposed components are aligned with the problem setting. SSCP provides structured conditioning of shared features, while MRTA and FLA separately address temporal consistency for video segmentation and temporal aggregation for video classification. This makes the framework technically coherent rather than simply combining multiple tasks in one model. 3.Broad experimental coverage. The evaluation spans multiple organs, modalities, tasks, and temporal settings, including both public and private datasets. This makes the study more comprehensive than many prior ultrasound works that focus on only one organ or one task. 4.Empirical results with useful ablations. The method shows competitive performance across several settings, and the ablation study helps verify that SSCP, MRTA, and FLA each contribute to the final performance. This strengthens the credibility of the proposed design.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The paper has several important weaknesses. 1.The method description and figures are not sufficiently clear. Figure 1 is difficult to read and some design choices are not well explained. In particular, the “Yes/No” branches around the mask decoder in the segmentation pathway are confusing, and it is unclear why the corresponding outputs are routed into the video branch in the illustrated way. In addition, the variable P in Eqs. (3) and (4) is not clearly defined, and the implementation details of SSCP remain too abstract. The statement that “Task, Target, Mode, and Modality are injected via learnable feature modulation, while Position is modeled through spatial bias” is conceptually reasonable, but the paper does not explain clearly how these priors are encoded and inserted at different layers. This makes the method harder to reproduce and assess. 2.Some key parts of the segmentation-aware classification pipeline are under-specified. The classification branch is described as using segmentation-aware representations and mask-guided ROI features, but the paper does not clearly explain how the masks are obtained for all classification settings, especially when classification datasets may not naturally provide segmentation supervision. This point is important for understanding both the practicality of the method and the fairness of the comparison with baselines that may not use such intermediate structural cues. 3.The empirical comparison is not fully fair or fully complete. In Table 1, the proposed method is trained in a supervised/fine-tuned setting on the benchmark, while several baselines are reported under zero-shot or prompt-based settings, which makes the comparison less clean. At the same time, some key settings are not matched with sufficiently strong competing baselines in the main table, especially for CEUS and video classification. Therefore, the claim of strong superiority across all settings is not yet equally convincing everywhere. 4.The quantitative presentation hides important dataset-level details. The paper mainly reports mean DSC and mean AUC/Accuracy aggregated across datasets. Given the substantial heterogeneity across organs, modalities, and tasks, this presentation makes it difficult to judge whether the gains are consistent across datasets or driven by a subset of easier benchmarks. Similarly, Table 2 would be easier to interpret if segmentation and classification ablations were presented more separately and explicitly. 5.The qualitative and representation analyses are not sufficiently convincing in their current form. Figure 2 only shows qualitative results of the proposed method, without visual comparison to representative competing approaches, so it is difficult to visually assess the claimed superiority. Likewise, Figure 3 is not sufficiently explained: the definition of kNN purity is not provided clearly, and the claim that decoder features exhibit better cross-domain alignment is not easy to verify directly from the figure. Overall, these analyses would benefit from clearer definitions, stronger visual evidence, and more careful interpretation. 6.Some ablation findings deserve deeper discussion. The MRTA ablation shows that increasing memory length improves performance, with the best result at K=10.However, the difference between K=6 and K=10 is still noticeable, especially for CEUS segmentation, and the paper does not discuss sufficiently why this happens or what the computational trade-off is. A more thorough sensitivity analysis would strengthen the temporal modeling claim. 7.The reproducibility and external validity are limited by insufficient dataset characterization. The paper uses multiple private datasets for video classification and CEUS analysis, but provides limited detail about acquisition conditions, scanner/vendor diversity, patient composition, annotation protocol, and center-level variability. Since the paper emphasizes a general and clinically relevant unified framework, more information is needed to judge how robust and transferable the reported results really are. In addition, the external validation is only reported for segmentation, not for classification or CEUS classification, so the evidence for broad cross-domain generalization remains partial. 8.The paper does not fully establish that the benefit comes from true unified multi-task learning. Although the framework is presented as unified, the training is performed in two stages: segmentation is optimized first, and then classification is trained with the encoder and segmentation branch frozen. This staged design is practical, but it weakens the claim that the gains arise from end-to-end unified learning itself. More direct comparisons with separate task-specific models using the same backbone would make this claim stronger. 9.The main novelty appears more integrative than fundamentally new. The overall unification across image/video, B-mode/CEUS, and segmentation/classification is meaningful, but the key components are relatively close to existing directions already discussed in the paper, including prompt-conditioned modeling (e.g., SAM, SAM2, MedSAM, SAM-Med2D, SAMUS), and memory-based temporal modeling for video segmentation (e.g., MemSAM). Therefore, the contribution seems to lie mainly in combining and adapting existing ideas into a broader ultrasound framework, rather than introducing a clearly new methodological principle.

  • 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.

    (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?

    I appreciate the ambition of this work. The paper tackles an important problem in ultrasound analysis by attempting to unify image and video modeling, B-mode and CEUS modalities, and segmentation and classification tasks within a single framework, and it is supported by relatively broad experiments across multiple datasets and settings. The reported results are also generally strong, and the ablation studies suggest that the proposed components are useful. However, the main factors that lowered my score are related to clarity, novelty, and the strength of the empirical evidence. First, the method description is not sufficiently clear in several key places: Figure 1 is hard to interpret, some variables such as P are not defined clearly, and the implementation of the semantic conditioning mechanism remains too abstract for full assessment or reproduction. Second, the comparisons are not entirely convincing, because the proposed model is trained in a supervised setting while several baselines are reported in zero-shot or prompt-based settings, and some important scenarios such as CEUS and video classification lack equally strong direct baselines in the main table. Third, many results are reported only as dataset-averaged metrics, which makes it difficult to judge whether the gains are consistent across organs and tasks. In addition, the classification branch relies on segmentation-aware representations, but the paper does not fully clarify how this is handled across all classification settings. Finally, although the overall framework is interesting, the methodological novelty appears more integrative than fundamentally new, as several core ideas are closely related to existing prompt-based unified ultrasound frameworks and prior memory-based video segmentation designs already discussed by the authors. Taken together, I find the paper promising, but in its current form the claims are not yet supported strongly enough for a clear accept.

  • 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 addressed my concerns well, and I believe the paper can be accepted.



Review #3

  • Please describe the contribution of the paper

    This paper proposes UIV-US, a unified ultrasound framework that aims to support both segmentation and classification across images and videos, covering both B-mode ultrasound and contrast-enhanced ultrasound (CEUS), within a shared-parameter architecture. The method combines three main components: (1) a Structured Semantic Conditioning Prompt (SSCP) that injects task-, target-, mode-, modality-, and position-related priors into intermediate features; (2) a Memory-Reinforced Temporal Attention (MRTA) module for temporally consistent video segmentation; and (3) a Feature-Level Attention (FLA) module for temporal aggregation in video classification. The paper evaluates the approach on a multi-organ benchmark that includes image segmentation, image classification, video segmentation, video classification, and CEUS tasks, and also provides module ablations and external validation on additional segmentation datasets.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The paper addresses an important and relevant problem setting. Rather than focusing on a single organ, modality, or task, the paper aims to unify image/video analysis and segmentation/classification across both B-mode and CEUS ultrasound. This broader formulation is interesting and clinically relevant because ultrasound data are heterogeneous in modality and temporal structure. 2.The overall framework is well motivated and conceptually coherent. The use of a shared encoder, semantic conditioning through SSCP, and separate temporal reasoning modules for video segmentation (MRTA) and video classification (FLA) forms a clear and reasonably well-structured methodological story. 3.The experimental scope is broad. The paper reports results on multiple tasks and datasets, including image segmentation, image classification, video segmentation, video classification, and CEUS settings. In addition, the paper includes ablation studies for SSCP, MRTA, and FLA, and also reports external validation on BUSI, BUSIS, and DDTI. 4.The reported quantitative results are promising. According to the paper, the method achieves strong aggregate performance across several settings, including 89.81% DSC for image segmentation, 96.28% for video segmentation, and 89.26% for CEUS segmentation, along with competitive classification results. The ablations also suggest that each proposed component contributes positively to performance. 5.The paper attempts to go beyond pure performance reporting. The external validation and the encoder/decoder feature analysis are useful additions that aim to support the claims of cross-domain robustness and representation alignment.

  • 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 methodological novelty appears moderate relative to prior work. The paper combines several familiar ingredients: shared-backbone multi-task learning, prompt/conditioning mechanisms, memory-based temporal modeling for video segmentation, and temporal aggregation for video classification. While the unified formulation is interesting, the paper does not fully convince me that the methodological contribution is sufficiently distinct from existing prompt-based and unified ultrasound frameworks such as SAM/SAM2 [11,25], MedSAM [19], SAM-Med2D [5], SAMUS [15], UniUSNet [16], and PerceptGuide [17], or from recent ultrasound foundation/universal modeling works such as USFM [10], TinyUSFM [18], and the multi-organ ultrasound foundation model with task prompts and anatomical priors [4]. In its current form, the paper reads more as a careful integration of existing ideas than as a clearly novel formulation. 2.The “unified” claim is weakened by the training protocol. The paper emphasizes unified multi-task learning in a shared parameter space, but the actual optimization is performed in two stages: segmentation is first trained with the classification branch frozen, and classification is then trained with the encoder and segmentation branch frozen. This is not the same as end-to-end joint multi-task optimization. As a result, it remains unclear how much of the benefit comes from genuine joint learning versus sequential reuse of a shared backbone. A direct comparison with true joint training, or with separately trained task-specific models under matched capacity, would strengthen the central claim. 3.Several comparisons are incomplete, especially for CEUS, video classification, and recent unified ultrasound baselines. Table 1 reports encouraging aggregate results, but the comparison matrix is uneven. In particular, the CEUS setting appears to include only the proposed method, and the video classification comparison is also limited. Moreover, for the image setting, the paper does not compare against recent unified ultrasound models from the UUSIC challenge, as reported in Diagnostic Performance of Universal-Learning Ultrasound AI Across Multiple Organs and Tasks: the UUSIC25 Challenge. These methods are particularly relevant because they also address single-model joint classification and segmentation across multiple organs, which closely matches the central claim of this submission. At minimum, the paper should discuss why such highly relevant baselines are not included. 4.The main quantitative table reports only averaged results across datasets. Table 1 presents mean DSC and mean AUC/Accuracy across datasets, which makes it difficult to assess whether the method is consistently strong across organs/datasets or whether some weaker cases are hidden by averaging. Given the breadth of the benchmark, per-dataset results would be important for a fair assessment of robustness and consistency. 5.Some key details about the conditioning and inference protocol remain unclear. The method relies on five semantic attributes in SSCP, including target and position. However, it is not sufficiently clear how these attributes are obtained at inference time, and whether they require manual input or additional metadata. Similarly, the classification branch uses segmentation-aware representations, but the paper does not fully clarify how mask guidance is handled in settings where only classification labels are available. These details matter for both fairness of evaluation and practical deployability. 6.The “foundation model” framing seems somewhat overstated. The paper refers to the method as a “foundation model,” but the experimental setup appears closer to a supervised unified multi-task framework trained on a curated benchmark. The paper does not demonstrate properties often associated with foundation models, such as broad transfer, low-shot adaptation, or zero-shot generalization. I would encourage the authors to moderate this terminology or provide stronger evidence supporting it.

  • 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.

    (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 paper addresses an interesting problem and presents a broad experimental study with promising aggregate results. I also appreciate the effort to unify multiple ultrasound settings within one framework and the inclusion of ablations and external validation. However, I am not yet fully convinced by the paper’s level of methodological novelty or by the strength of the evidence supporting its central claims.

    My main concerns are:

    1.The novelty appears moderate relative to prior prompt-based, unified, and foundation-style ultrasound methods; 2.The “unified multi-task learning” claim is weakened by the staged training protocol rather than true joint end-to-end optimization; 3.The comparisons are incomplete in some important settings, particularly CEUS and video classification; 4.The paper reports mostly averaged results rather than per-dataset performance; 5.Several practical details of the conditioning and inference process remain unclear.

    Overall, I think the work has potential and is not far from the acceptance threshold, but in its current form I find the evidence insufficient to clearly support acceptance.

  • 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.

    I appreciate the authors’ careful and constructive rebuttal. The response clarifies several implementation details, especially the source and encoding of SSCP attributes, the use of predicted soft masks rather than ground-truth masks for classification, and the intended interpretation of the term “unified”. I also appreciate that the authors agree to moderate the “foundation model” framing.

    However, after considering the rebuttal, my main concerns remain only partially resolved. My concern is that the submitted manuscript itself lacks some key evidence needed to support its central claims. In particular, the authors clarify that the work is unified mainly at the model/workflow level rather than through true end-to-end joint multi-task optimization. This clarification resolves the wording ambiguity, but it also means that the “unified learning” claim should be interpreted more conservatively.

    In addition, the submitted manuscript has incomplete empirical support in several important settings. CEUS and video classification lack strong matched baselines, the comparison with highly relevant recent unified ultrasound methods remains insufficient, and the main quantitative table reports averaged results rather than per-dataset performance. These limitations make it difficult to assess whether the method is consistently robust across organs, modalities, and tasks, rather than mainly strong on aggregate metrics.

    Overall, I still find the problem important, the framework coherent, and the reported results promising. Nevertheless, I remain unconvinced that the submitted paper provides sufficiently strong methodological novelty and sufficiently complete empirical evidence to support acceptance at this venue. Therefore, I maintain my recommendation to reject.



Author Feedback

We thank all reviewers for their helpful feedback.

Q1: Novelty relative to prior ultrasound models. (AC,R1,R2,R4) We stress our novelties in the Contributions: UIV-US provides a shared-parameter B-mode/CEUS image-video segmentation/classification framework; SSCP builds a semantic coordinate system over Task/Target/Mode/Modality/Position; and MRTA/FLA bridge image-video segmentation/classification. The novelty lies in the unified ultrasound workflow rather than an isolated module. Prior works mainly focus on segmentation, encoder pretraining, video segmentation, or image-level segmentation/classification, and we will expand the discussion to the omitted works(R1, R4). Results support the design, e.g., image segmentation over PerceptGuide (89.81% vs. 87.62% DSC), image classification AUC (85.76% vs. 79.33%), and video segmentation over MemSAM (96.28% vs. 95.23% DSC). Following R4, we use “unified ultrasound framework” rather than “foundation model”.

Q2: “Unified learning” and two-stage training. (AC,R2,R4) We acknowledge that “jointly optimized” in Sec. 2.1 may have caused ambiguity. “Unified” is intended at the model level: shared architecture, SSCP-conditioned representations, and one workflow across B-mode/CEUS, image/video, and segmentation/classification. The two-stage strategy in Sec. 3.1 handles heterogeneous supervision: segmentation learns localization from masks, while classification uses image/video-level labels. This is for stable optimization, not model separation; UIV-US remains one framework at inference. We will clarify the wording to avoid implying one-step joint optimization.

Q3: Acquisition and encoding of SSCP attributes. (AC,R1,R2,R4) At inference, SSCP attributes come from the requested task and known input labels. Task denotes output type; Target organ/lesion; Mode whole/local input; Modality B-mode/CEUS; and Position anatomical site. Thus, SSCP requires no additional test-time labels. If attributes are unavailable, corresponding prompts can be omitted; Table 2 shows removing SSCP degrades performance. Each attribute is one-hot encoded as P_a and mapped by a learnable prompt bank, e_a=P_aW_a. Task/Target/Mode/Modality generate channel-wise modulation h’=h(1+tanh(γ))+β; Position adds spatial bias h’=h+c_posB_xy. We will clarify this implementation and define SSCP prompt set P in Eqs. (3)-(4).

Q4: Fairness of comparisons and CEUS/video baselines. (AC,R2,R4) In Table 1, zero-shot SAM-family models assess direct transferability, fine-tuned methods assess task-adapted performance, SAMUS/MemSAM provide ultrasound image/video segmentation baselines, and UniUSNet/PerceptGuide represent unified image-level segmentation/classification models. These comparisons provide matched evidence for UIV-US in image/video segmentation and image classification. Since matched baselines for CEUS/video classification are relatively scarce, we acknowledge those results as evidence of effectiveness rather than superiority over specialized models.

Q5: Method clarification, temporal robustness, and evidence presentation. (R1,R2,R4) UIV-US uses a shared ViT-B(12) encoder. In Fig. 1, “Yes/No” indicates temporal routing: T=1 uses the image pathway, while T>1 activates MRTA/FLA. For classification, no gt mask is required; the branch uses the soft mask predicted in the same forward pass as an ROI cue and fuses it with global features. MRTA does not propagate previous masks as predictions; they are auxiliary memory, while each frame is segmented from current visual evidence, mitigating error accumulation. K=10 preserves more temporal context than K=6 but increases cross-attention cost; we will clarify this trade-off and report inference complexity. Table 1 reports averages because UIV-US targets cross-scenario performance and per-dataset reporting needs more space; dataset-level results will be released with code. We will add a “Task” column to Table 2, define kNN purity in Fig. 3, and avoid overclaiming generalization.



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 work tackles a well-motivated problem, and the proposed components (SSCP, MRTA, FLA) are coherently motivated and supported by ablation studies. However, several main concerns need to be addressed, but not limited to:

    1.Justification of novelty relative to prior methods, such as SAMUS, USFM, URFM, etc.

    2.The authors claim end-to-end unified learning within a shared parameter space, yet the model is trained in two sequential stages. The authors should clarify this inconsistency and justify the design choice.

    3.Sufficient details should be provided, particularly regarding how the five semantic attributes (Task, Target, Mode, Modality, Position) are encoded and acquired at both training and inference time.

    4.Unfair comparison settings (e.g., other methods are zero-shot ) and lack of baselines for CEUS settings. These make the experiments not 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 rebuttal addressed many of the main concerns and resulted in an overall positive post-rebuttal assessment. The response clarified important implementation details and the intended interpretation of the framework. While the “unified” claim should be presented more conservatively in the camera-ready version, I do not view this as sufficient reason for rejection. Overall, the paper addresses an important problem and provides sufficient merit for 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.

    This paper proposes UIV-US, a unified ultrasound framework that aims to support segmentation and classification across images and videos, B-mode and CEUS modalities, within a shared-parameter architecture. The work addresses an important and clinically relevant problem, and the scale of the experimental evaluation is substantial. Reviewers acknowledged the value of the unified formulation, the coherent design of SSCP, MRTA, and FLA, and the broad validation across multiple tasks, organs, and modalities. Following rebuttal, two reviewers explicitly changed their recommendations from Weak Reject/Accept to Accept after the authors clarified the novelty relative to prior ultrasound foundation models, the rationale behind the two-stage training strategy, the implementation and inference protocol of SSCP, and the fairness of the experimental comparisons. I find that the paper makes a meaningful contribution to unified ultrasound modeling and provides sufficient empirical evidence to support its claims. Overall, the strengths outweigh the remaining weaknesses, and 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.

    Accept

  • Please justify your recommendation.

    The work is more integrative than fundamentally novel, and the unified/foundation-model claims should be moderated. Still, the broad ultrasound evaluation and useful ablations provide sufficient practical value for acceptance.



back to top