List of Papers Browse by Subject Areas Author List
Abstract
Breast cancer is highly prevalent worldwide. While breast ultrasound (BUS) serves as a primary screening modality, its reliable diagnostic interpretation remains inherently challenging.
Most existing BUS classification methods prioritize direct image-to-label mapping and overall accuracy, neglecting the alignment with clinical diagnostic rules and decision traceability.
Such limitations hinder their integration into rule-governed clinical workflows and diminish the trustworthiness of automated diagnostic decisions. Although agent-based frameworks offer a modular paradigm for organizing complex decision processes, their potential for rule-aligned and fully traceable BUS diagnosis remains largely unexplored.
To address these challenges, we introduce RACA, a Rule-Aligned Collaborative Agent framework for evidence-grounded BUS classification. Our RACA introduces a novel, three-stage framework driven by explicit clinical rules. First, a perception module consolidates geometric and semantic evidence into structured diagnostic evidence. Then, a risk modeling agent aggregates the evidence under explicit clinical constraints to derive a rule-aligned risk estimate and an initial prediction. Finally, a critic-guided calibration agent refines predictions through memory-based retrieval. Extensive experiments on three BUS benchmarks demonstrate the superiority of RACA over the representative classification and medical agent-based frameworks. Comprehensive interpretability analyses and case demonstrations validates that its traceable decision pathway aligns with clinical diagnosis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2740_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Swecamellia/RACA
Link to the Dataset(s)
N/A
BibTex
@InProceedings{CheLin_RACA_MICCAI2026,
author = { Chen, Lingyu AND Wang, Yue AND Li, Cheng AND Jin, Lin AND Zhang, Daoqiang AND Liao, Hongen AND Chen, Fang AND Meng, Qingjie},
title = { { RACA: Rule-Aligned Collaborative Agents for Evidence-Grounded Breast Ultrasound Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a Rule-Aligned Collaborative Agent (RACA) framework for breast ultrasound (BUS) classification. The core idea is to transform conventional deep learning-based classification into a clinically grounded, rule-based, and traceable decision-making process. The main contributions are as follows:
#1.The authors design a rule-aligned collaborative AI agent framework that reformulates BUS classification as a rule-driven, three-stage decision process. #2.They introduce an Evidence-Grounded Perception (EGP) module, which converts traditional implicit feature learning into an explicit diagnostic attribute space (e. g. , shape, margin, echogenicity, and posterior features). Furthermore, a Rule-Aligned Risk Modeling mechanism is developed to map diagnostic attributes into a continuous risk score, incorporating consistency and uncertainty modulation. In addition, a critic-guided calibration module with memory retrieval is proposed, which leverages historical cases via KNN-based risk estimation to emulate experience-driven decision refinement. #3.Extensive experiments on multiple BUS datasets demonstrate the effectiveness of the proposed framework in terms of both classification accuracy and interpretability.
- 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.Clinically meaningful problem formulation. The paper explicitly aligns the diagnostic process with BI-RADS-style clinical rules, reformulating BUS classification as a rule-governed decision pipeline rather than a pure image-to-label mapping. #2.Well-structured multi-agent system design. The proposed framework decomposes the task into multiple collaborative agents (e. g. , perception, risk modeling, and calibration), enabling modular reasoning and structured information flow. #3.Strong emphasis on interpretability and traceability. The introduction of attribute-level diagnostic evidence, explicit risk aggregation, and step-by-step decision traces provides a transparent and auditable reasoning process. #4.Comprehensive experimental evaluation. The method is validated on three BUS datasets (BUSI, BUV, and BUSBRA), with clear ablation studies demonstrating the contribution of individual components.
- 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 in the overall Agent architecture. The proposed multi-Agent framework built upon AutoGPT follows a typical three-stage design: perception Agent → reasoning Agent → calibration/reflection Agent. This paradigm is highly similar to several recent established medical Agent works, including MedAgents (ACL 2024), MDAgents (NeurIPS 2024), and AgentHospital (Arxiv 2024). The paper only adapts this pipeline to breast ultrasound classification, without notable innovations in Agent collaboration, dynamic planning, communication, or decision-making mechanisms. Therefore, the contribution is more application-oriented than methodological. #2.The memory-based calibration module is poorly motivated and designed. The case-retrieval calibration lacks clear justification. The memory construction, feature representation for retrieval, and generalization risks are not well analyzed. Furthermore, the calibration strategies (dual-threshold veto, top-k rescue, gating) are heuristic post-processing steps rather than autonomous Agent-driven decisions, which deviates from the core idea of intelligent Agent behavior. #3.Agent autonomy is overclaimed; the framework resembles a fixed modular pipeline. Despite being labeled as collaborative Agents, the system follows a predefined, fixed execution flow without dynamic planning, conditional branching, error recovery, or high-level self-reflection. The Agents lack real autonomy and interact only through fixed data passing. Thus, the framework behaves more like a rigid modular pipeline rather than an intelligent, adaptive Agent system, which is inconsistent with the title and claimed advantages. #4.The overall framework diagram is overcrowded, overly complex, and visually cluttered. The excessive use of abstract notations (Astr, Asem, Afusion, etc. ), mixed arrows, and dense text make it difficult to follow the workflow at a glance. The three-stage structure claimed in the paper is not clearly visualized in the figure. The authors are suggested to simplify the diagram, use clear labels, separate stages visually, and reduce visual clutter to improve readability.
- 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 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?
My recommendation is weak accept. The paper targets AI Agent for medical image diagnosis, a highly topical and promising direction in recent medical imaging research. It proposes a collaborative multi-agent framework for breast ultrasound classification, with rule alignment and evidence grounding, which is clinically meaningful and well-positioned to the community. However, the major strength and critical weakness both center on its AI Agent design. While the multi-agent structure improves modularity and interpretability, the framework largely behaves as a fixed modular pipeline wrapped in Agent terminology. It lacks genuine autonomy such as dynamic planning, conditional reasoning, or adaptive interaction. The claimed agent collaboration is limited to predefined data flow rather than intelligent decision-making. Nevertheless, the paper is clearly written, experimentally solid, and provides useful insights for rule-aware clinical diagnosis. I therefore support weak accept provided the authors better justify and strengthen the authentic agentic behaviors.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for addressing my concerns in the rebuttal. In particular, I would like to emphasize that Figure 1 still requires major revisions. Regarding the comparative experiments in Table 1, I suggest including the latest medical foundation model methods published in 2025 or 2026 for a more comprehensive evaluation. Overall, the study applies state-of-the-art methods to BUS classification and achieves promising results.
Review #2
- Please describe the contribution of the paper
This paper proposes a collaborative agent framework designed to anchor the Breast Ultrasound (BUS) classification task to an explicit and rule-compliant clinical diagnostic workflow. Extensive evaluations that the proposed method consistently outperforms various representative classification models and medical agent models.
- 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 experimental results are detailed and comprehensive, accompanied by a complete set of figures and tables. 2.The proposed method by anchoring decision paths to clinical guidelines facilitates alignment with clinical diagnostic workflows and enables traceable risk assessment.
- 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 overall input-output workflow shown in Figure 1 is unclear. The figure appears to have been generated by AI. 2.The overall layout and formatting are user-unfriendly, resulting in poor readability—for instance, the text in Table 3 is too small, and Figure 2 is visually cluttered. 3.The list of references is incomplete. 4.The size of the dataset has not been specified. 5.The methodology is described unclearly, and its novelty remains to be determined.
- Please rate the clarity and organization of this paper
Poor
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Innovativeness, readability, reproducibility
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
The author did not answer all the questions very well.
Review #3
- Please describe the contribution of the paper
This manuscript tackles a persistent issue in medical image analysis: the opacity of end-to-end classification models. By orchestrating collaborative agents to follow BI-RADS clinical guidelines, the authors achieve commendable diagnostic traceability. The three-stage evidence-grounded framework is well-motivated, and the interpretability figures are compelling. However, relying on a complex cascade of foundational models demands a deeper investigation into error propagation, memory dynamics, and algorithmic efficiency before clinical adoption.
- 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.
Integrating BI-RADS guidelines into a multi-agent framework beautifully bridges the persistent gap between black-box deep learning and strict clinical diagnostic workflows.
- 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 pipeline heavily relies on upstream models like YOLO and nnU-Net. If the initial ROI extraction completely fails or isolates severe acoustic artifacts, downstream agents might confidently process garbage. Please provide an explicit failure-case analysis demonstrating how the risk modeling agent handles catastrophically wrong bounding boxes.
2.Employing a vision-language model for semantic parsing is interesting but risky. The manuscript lacks detail on the specific VLM used and the exact prompting strategy. Medical VLMs often hallucinate features. Detail how these hallucinations are tightly constrained before reaching the auditable diagnostic-attribute fusion agent.
3.Equation 1 introduces critical constants regulating temperature, instability scaling, and consistency penalization. A method claiming stable rule-based alignment must demonstrate that its clinical interpretability does not heavily rely on dataset-specific tuning. Omitting a sensitivity analysis for these hyperparameters weakens the generalizability claim.
4.The critic-guided calibration relies on a memory retrieval mechanism. I question how the size, quality, and diversity of this memory bank affect the final calibration. Please evaluate how the calibration agent behaves if the memory bank is drastically reduced or contains highly imbalanced prior cases.
5.Running multiple agents, a VLM, and hybrid segmenters introduces massive computational overhead compared to typical monolithic networks. For a system aiming at real-world clinical workflow integration, you need to report the end-to-end inference latency per image and discuss the pipeline’s hardware deployment requirements.
- 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?
The conceptual shift from direct image-to-label mapping to structured evidence aggregation fundamentally improves the trustworthiness of automated medical image diagnosis.
- 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.
N/A
Author Feedback
We sincerely thank ACs, PCs, and reviewers for the constructive comments. We appreciate the recognition of our well-structured multi-agent design, strong clinical interpretability and traceability, and comprehensive validation. We provide point-by-point responses below. Code will be released upon acceptance. R2-Q1: Agent Novelty. Unlike general medical-agent frameworks (e.g., MedAgents) that rely on free-form collaboration and lack clinical-rule alignment and decision traceability, our method encodes BUS clinical rules as executable constraints by learning structured diagnostic evidence, estimating rule-aligned risk and performing critic-guided calibration, improving interpretability and traceability. R2-Q2: Memory Calibration. Our calibration is a rule-constrained retrieval and verification step, not heuristic post-processing. Specifically, the calibration agent generates candidate updates, and a rule-constrained selection policy chooses among them and the initial prediction, avoiding arbitrary label flips. Memory bank is built from validation cases, and retrieval keys are diagnostic-state embeddings (e.g., lesion shape) rather than raw images or labels. R2-Q3: Rule-Constrained Autonomy. Our agents perform rule-constrained coordination, not open-ended autonomy. Unlike fixed pipelines, each agent updates shared diagnostic states when reliability/consistency checks pass, e.g., the risk agent updates only when evidence confidence is sufficient, the critic revises only under rule-aligned retrieval agreement. This ensures clinically auditable decisions in high-risk BUS diagnosis. R2-Q4: Diagram Readability. In the final version, we will simplify Fig.1 by reducing visual clutter (e.g., arrows/text), grouping 3 stages clearly and enlarging labels for readability. R3-Q1: Presentation. Fig. 1 was manually designed in PowerPoint. See R2-Q4 for revision details. R3-Q2: Formatting. Larger font size and whitespace will be applied to Tab. 3 and Fig. 2 in the final version. R3-Q3: Reference Completeness. We will carefully re-examine the references to ensure completeness. R3-Q4: Dataset Size. Sec. 3 cited the public datasets. Final version will add specific dataset sizes. R3-Q5: Novelty. Please see R2-Q1 for novelty details: rule-aligned design, evidence-grounded methodology and traceable clinical workflow. R2 further noted our ‘well-structured multi-agent system design,’ which supports the contribution. R4-Q1: ROI. Our model mitigates poor-ROI impact by complementing ROI-based structural cues with image-level semantic cues. In a poor-ROI case, low ROI reliability down-weights unreliable structural cues (e.g., edge strength) and suppresses malignant trigger, while semantic cues (e.g. indistinct margin) still support malignancy. As a result, RACA still predicts malignant correctly, showing robustness to poor ROIs. R4-Q2: Hallucination. We use Qwen3-VL-4B-Instruct with a fixed JSON-only prompt (released with codes). The prompt restricts VLM outputs to four controlled diagnostic slots with confidence, evidence, and uncertainty fields. Before fusion, outputs pass parsing, vocabulary validation, and rule-based consistency checks. Unsupported claims are mapped to “unknown” or low-confidence, preventing hallucinations from entering diagnosis. R4-Q3: Parameter. Parameters in Eq. 1 were fixed across all three datasets without per-dataset tuning, yet RACA achieves consistent performance gains, indicating robustness rather than reliance on hyperparameter search. A full sensitivity analysis will be in future work. R4-Q4: Memory Robustness. Reducing the memory bank from 100→50% caused only a small ACC drop of ≤5%, while further reducing it from 50→0% led to a 5-10% ACC drop, indicating graceful degradation. Our memory banks are class-imbalanced (BUSI 64/48, BUV 33/10, BUSBRA 113/55 benign/malignant), supporting robustness to memory imbalance. R4-Q5: Latency. Runtime is 31.40 s/image on one RTX 4090 24GB GPU, with most latency from VLM (31.38 s/image).
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 presents a multi-agentic framework for strcutred evidence aggregation and traceable/interpretable breast ultrasound diagnosis, showing promising performance. Several concerns, espeically those regarding technical novelty (especially from the AI agent perspective), system sensitivity to key upstream models, and computational complexity, should be addressed.
- 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.
An interesting paper with the novelty of multi-agent design. Overall, the authors rebuttal could address most concerns regarding technical and evaluation details.
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.
R2 and R4 recommended Accept. Although R3 recommended Reject, the weaknesses mentioned appear to be relatively minor and could likely be addressed in the camera-ready version.
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 authors have clarified most of the concerns raised by reviewers. The manuscript proposes a quite interesting and novel agentic framework for medical image analysis, of which the merit is well acknowledged.