List of Papers Browse by Subject Areas Author List
Abstract
Ultrasound image classification is essential for computer-aided diagnosis. However, current methods often neglect clinical priors, leading to poor generalization in challenging scenarios and a lack of interpretability that limits clinical adoption. To address these issues, we aim to develop a medical-prior module that can be seamlessly integrated into existing pipelines to enhance both diagnostic performance and interpretability. In this paper, we propose an attribute-guided dual-branch framework for ultrasound classification that introduces domain-agnostic medical attribute priors, improving generalization while offering interpretable evidence. Specifically, a baseline branch follows conventional architectures and predicts image categories via a fully connected classifier. An attribute-guided branch injects domain-agnostic attributes as priors and produces human-interpretable decision cues. Finally, an adaptive decision module fuses the two branches in a data-dependent manner to yield the final prediction. Experiments across diverse ultrasound classification tasks demonstrate that our approach can be integrated into multiple backbones and state-of-the-art methods with low overhead, consistently improving accuracy and interpretability. Code is available at: https://github.com/zhaobo253-crypto/AttrGuide.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3025_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zhaobo253-crypto/AttrGuide
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaBo_Boosting_MICCAI2026,
author = { Zhao, Bo AND Li, Yapeng AND Liu, Juhua AND Du, Bo},
title = { { Boosting Ultrasound Image Classification via Attribute-Guided Dual-Branch Framework } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes an attribute-guided dual-branch framework for ultrasound image classification, aiming to improve both diagnostic performance and interpretability. The method consists of a baseline classification branch and an attribute-guided branch that injects domain-agnostic medical priors. An adaptive fusion module combines the two branches to produce the final prediction. The framework is designed to be plug-and-play and is evaluated on both public (BUSI) and private fetal ultrasound datasets, showing consistent performance improvements across different backbones.
- 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 problem in ultrasound imaging, where interpretability and robustness are critical. 2.The framework is clearly described and easy to integrate into existing models. 3.The idea of incorporating attribute-based priors is well-motivated.
- 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 is limited, as the approach largely builds upon existing ideas such as attribute-guided learning, concept bottleneck models, and feature fusion. 2.The performance improvements are relatively modest in some cases, particularly for certain backbones, which limits the overall impact. 3.The attribute design relies on manually defined expert knowledge, which may limit scalability and generalizability to other tasks or domains. 4.The evaluation is conducted on relatively limited datasets, with no strong validation under domain shifts or across diverse clinical settings. 5.While interpretability is claimed, the analysis remains somewhat limited, lacking deeper validation
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Suggestion for improvement: 1.Clearly differentiate the proposed framework from prior work, and provide stronger ablation studies or theoretical justification for the design choices. 2.Explore automatic or data-driven approaches for learning attributes, and evaluate the robustness of the model under different or noisy attribute sets. Additionally, discuss how the framework generalizes to tasks without expert-defined attributes. 3.More thoroughly demonstrate the plug-and-play capability across diverse architectures and tasks. 4.Include qualitative case studies showing how attribute predictions influence final decisions and conduct user studies or expert evaluation to assess clinical usefulness. 5.Provide visualization (e. g. , attention maps or attribute-region alignment) to support interpretability claims.
- 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 paper presents a well-motivated framework with practical value, but the contribution is incremental and the performance gains are modest. The work lacks clear differentiation from prior methods and stronger validation, particularly regarding robustness, scalability of attribute design, and interpretability claims. If the authors can address the above concerns, especially improving interpretability, generalizability, and more convincingly demonstrating the plug-and-play capability, the work could become a solid contribution.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This manuscript presents an attribute-based dual-branch framework for ultrasound classification. The proposed method consists of a standard classification branch and an attribute-guided branch: the former follows a conventional architecture with a fully connected classifier for category prediction, while the latter introduces so-called domain-agnostic medical attribute priors to provide human-interpretable decision cues. An adaptive decision module is then used to fuse the outputs of the two branches in a data-dependent manner for final prediction.
- 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 attempts to incorporate attribute-level priors into ultrasound classification in order to improve both predictive performance and interpretability. This general direction is meaningful, and the idea of providing semantically meaningful cues for decision-making is potentially interesting.
- 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.Limited novelty and insufficient clarification of the role of the medical attribute semantic space. The overall novelty appears limited. The method mainly relies on CLIP to construct a medical semantic space and uses a dual-branch framework for prediction. However, the manuscript does not clearly explain what specific role this medical attribute semantic space plays in the proposed method. The authors use customized prompts to obtain language embeddings and compare them with image features, but CLIP is originally trained on natural image-text pairs. The paper does not sufficiently justify why such a semantic space can be expected to work effectively for a medical ultrasound task without further adaptation or evidence. 2.The notion of “domain-agnostic medical attribute priors” is not clearly defined. The manuscript repeatedly emphasizes domain-agnostic medical attribute priors, including in the abstract, but this concept itself is not clearly explained. If these attributes are truly domain-agnostic, it is unclear why they would be effective for a specific ultrasound classification task. On the other hand, if they are indeed informative for the task, then their relationship to the specific imaging modality, disease categories, or clinical context should be more clearly explained. At present, this key concept remains vague, and the methodological motivation is therefore not sufficiently convincing. 3.The description of the attribute branch and Figure 2 is unclear. The paper states that the attribute branch is derived from CLIP, but Figure 2 does not clearly show the corresponding modules, making the connection between the figure and the text difficult to follow. In addition, several important details are missing: What exactly are the ultrasound-specific prompts? How are these prompts designed? How is the patch-level attribute similarity map generated? How are the text attributes matched with image patch features? Since prompt design and attribute similarity computation are central to the proposed method, these components should be described more explicitly and concretely. 4.The category definition of the private dataset is incomplete. The manuscript mentions that the private dataset contains seven classes, but it does not specify what these seven classes are. For a classification task, the class definition is a basic and necessary part of the dataset description, and it should be clearly provided. 5.The experimental design is not sufficient to support the effectiveness of the attribute module. In Table 1, the authors seem to aim to demonstrate the benefit of the proposed attribute module. However, if the goal is to show that this module is effective, the more appropriate design would be to report results with and without the attribute module across all evaluated backbones, rather than only presenting the BU-Mamba results. The current experiments are not sufficient to show that the proposed module consistently improves performance across different backbone architectures. 6.The claimed interpretability is not yet convincing. Figure 3 attempts to demonstrate interpretability, but the current evidence is not sufficiently convincing. If the interpretability mainly comes from predicting attributes corresponding to manually designed prompts, then this seems closer to matching predefined semantic labels rather than genuinely explaining the model’s decision-making process. In other words, since the attributes themselves are predefined by the authors, it remains unclear whether predicting these attributes should be considered true interpretability. The authors need to better justify why these attribute predictions constitute meaningful explanatory evidence rather than merely reflecting the predefined prompt design.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
In summary, while the topic is relevant and the general idea is potentially interesting, the current manuscript has substantial weaknesses in methodological clarity, experimental support, and the justification of its interpretability claims. The authors are encouraged to more clearly define the proposed attribute priors, justify the use of CLIP in the medical ultrasound setting, provide more complete methodological details, strengthen the ablation experiments, and offer a more rigorous explanation of what is meant by interpretability in this work.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for their detailed response. The rebuttal addresses most of my major concerns by providing additional evidence on backbone generalizability, task generalizability, attribute-table robustness, self-learned attributes, and the use of MedCLIP. These additions substantially strengthen the experimental support for the proposed attribute-guided framework. I also appreciate the clarification of the methodology, including how LLM-generated attributes are encoded by CLIP, mapped into the visual feature space, and used to generate attribute heatmaps and final predictions. The added definition of the seven fetal plane classes and the explanation of “domain-independent” attributes also improve the clarity of the manuscript. Some concerns remain regarding the precise interpretation of attribute-based interpretability and the terminology of “domain-agnostic/domain-independent” priors, but I believe these issues can be addressed through careful revision. Overall, the rebuttal improves the clarity and support of the work.
Review #3
- Please describe the contribution of the paper
The main contribution of this paper is a lightweight, plug-and-play attribute-guided dual-branch framework (AttrGuide) that can be seamlessly integrated into any existing ultrasound classifier. It introduces a parallel attribute-guided branch that injects domain-agnostic clinical priors (via frozen CLIP text embeddings and a fixed class-attribute matrix) and an adaptive fusion module that dynamically reconciles baseline logits with attribute-based logits. This design simultaneously improves classification accuracy, generalization, and human-interpretable decision cues while adding <5% training-time and parameter overhead.
- 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.
1Practical plug-and-play design: Requires only minimal modifications to existing backbones (ResNet, ViT, Mamba) and delivers consistent gains without re-training the whole model. 2Clinical relevance: Explicitly uses domain-agnostic medical attributes derived from public guidelines, providing human-readable evidence (attribute prediction accuracy 83–87%) that clinicians can verify. 3Efficiency: Extremely low overhead (e.g., +0.16–1.93 min per fold, +0.5–1.3 M parameters) while improving strong baselines (e.g., BU-Mamba from 87.86% → 88.72% on BUSI; ViT-B +4.3% Acc on BUSI). 4Thorough validation: Extensive experiments across two datasets (public BUSI + private fetal 7-class), multiple backbones, multi-task settings, and detailed ablations showing the necessity of both branches and learnable fusion.
- 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.
1The improvement on the strongest baseline (BU-Mamba) is relatively modest (+0.86% Acc). 2The fetal dataset is private, limiting direct reproducibility of that experiment. 3Novelty is incremental rather than groundbreaking: the approach combines ideas from Concept Bottleneck Models and CLIP-based attribute guidance, but does not introduce a fundamentally new paradigm compared to recent MICCAI works on prompt-based or attribute-guided medical VLMs (e.g., refs 7–9, 17–18, 20–24).
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a practical and well-executed plug-and-play module that meaningfully addresses the interpretability–performance trade-off in ultrasound classification. Experiments are solid and results are consistent. However, the absolute gains on top-tier baselines are modest and the core idea builds incrementally on existing attribute-guidance and CBM literature. A score of 4 reflects a solid methodological contribution suitable for MICCAI, but not a “must-accept” breakthrough.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers for their constructive feedback and for recognizing the significance of our work. Q1: Generalizability and task coverage (R1,R2,R3) A1: (1) We validated our method on four mainstream backbones: CNN (ResNet), Transformer (ViT), Mamba (BUMamba), and Diffusion (DiffMIC-v2). It improved their performance by 1.8%, 4.3%, 0.86%, and 3.4%, respectively, showing backbone generalizability. (2) We added a thyroid ultrasound dataset with a 4.3% gain. Consistent improvements across fetal, breast, and thyroid ultrasound tasks validate task generalizability. Q2: Ablation studies (R1,R2) A2: (1) Attribute-table robustness: The attribute table is LLM-generated without manual expert annotation. It is robust: under the same prompt, three generated tables yielded stable accuracies (85.4%, 85.6%, 84.9%). Furthermore, a progressive analysis shows that manual expert selection further elevates performance to 86.4%, demonstrating the potential for even higher gains when manual expert knowledge is integrated. (2) Automatic attribute learning: Self-learned embeddings achieve 82.8%, showing utility without semantic vocabularies; however, our explicit attributes with offline encoding are stronger. (3) Medical CLIP: Replacing standard CLIP with MedCLIP improves accuracy (85.4% to 86.3%), showing medical text priors benefit the model. Q3: Attribute priors and method comparison (R1,R2,R3) A3: (1) Attribute priors: Attribute priors encode discrete attribute tables into semantic anchors via offline CLIP. Attributes such as shape and margin serve as domain-invariant priors; our fetal dataset was collected from three hospitals, reflecting cross-center differences. (2) Comparison and versatility: Existing methods like medical VLMs, alignment methods, and CBMs often need scenario-specific designs, limiting plug-and-play versatility. In contrast, our method uses LLMs to auto-generate task-specific discriminative attributes. Via a dual-branch structure, we introduce attribute priors in a plug-and-play manner while preserving the original backbone. This significantly enhances generalization and universality across clinical tasks with 5% overhead (see Q1). Q4: Interpretability, qualitative cases, clinical use, and reproducibility (R1,R2,R3) A4: (1) Interpretability: Attribute prediction accuracy (83.28%) provides reliable decision clues. CAM attribute-region maps will be released with the code. (2) Qualitative analysis: On the breast test set, we report three-branch predictions and top-5 attributes per sample. Cases show our attribute/fusion branches correctly predict malignancy when the visual branch misjudges it as benign, proving fusion corrects visual-head errors. (3) Clinical utility: We plan to invite senior physicians to evaluate diagnostic accuracy and clinical acceptance. (4) Reproducibility: We will release code and model weights upon acceptance, enabling reproducibility. Q5: Terminology, CLIP attribute anchors, methodology, and dataset definition (R2) A5: (1) Terminology: “Domain-independent” refers to data-source independence; e.g., shape/margin remain center/device-invariant. (2) Methodology: a. ChatGPT generates task-discriminative attributes; b. CLIP extracts initial attribute embeddings offline; c. An FC layer maps embeddings to visual space; d. Cosine similarity between visual features and embeddings generates attribute heatmaps; e. Final predictions are derived via weighted averaging of heatmaps, which are supervised by category-specific attribute labels mapped from category labels. (3) CLIP adaptation: CLIP is applicable because basic features, such as edges and shapes, are universal. Beyond these, our supervised optimization in step e refines all initial semantic embeddings—covering universal and specialized medical attributes—to bridge domain gaps and calibrate them to clinical nuances. (4) Dataset: Seven fetal planes: three-vessel; aortic arch; fetal bladder; transcerebellar plane; thorax; ductal arch; fetal orbits.
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 paper addresses an important ultrasound classification problem with a practical plug-and-play design and promising results, but reviewers note that the novelty is incremental and that the attribute-based interpretability and experimental support are not yet fully convincing. The authors are encouraged to clarify the role of the attribute priors, strengthen ablations and justification, and better demonstrate generalizability in rebuttal.
- 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 presents a practical and well-motivated plug-and-play attribute-guided framework for ultrasound classification, with consistent gains across multiple backbones and datasets. The rebuttal also clarified several key concerns, especially the role of the attribute priors, the robustness of the attribute table, and the generalizability of the method beyond a single backbone. Although the method is incremental and the interpretability claims could still be stronger, I think the combination of improved performance, low adoption barrier, and broad applicability makes this a solid contribution. Overall, I am inclined to accept this paper.
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.
One reviewer rates as acceptance and two reviewers have not involved in the rebuttal phase. I read the paper, the rebuttal file, and reviewers’ comments after the rebuttal. think that the authors solve the main concerns though some important issues still need to be fixed. But due to the novelty of the method, I tend to 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.
All the reviewers highlighted the practical importance of the paper due to its plug-and-play nature and its ability to generalize across diverse backbones, despite the modest novelty of the methodology.