Abstract

Neoadjuvant chemotherapy (NAC) response prediction is clinically important for treatment stratification in breast cancer. However, robust pre-treatment pathological complete response (pCR) prediction remains challenging due to insufficient cross-modal modeling, multicenter imaging heterogeneity, and weak evidence-grounded interpretability. We propose ClinRAG-GRAPH, a Clinically informed Retrieval-Augmented Generation Graph framework, for pre-treatment pCR prediction from DCE-MRI, structured clinical variables, and biopsy-derived pathological biomarkers. ClinRAG-GRAPH constructs an intra-patient clinical-prior graph and applies a prior-guided relation-aware graph convolutional network for structured multimodal representation learning. To improve cross-center robustness, we introduce a dual-branch domain-adversarial learning strategy to suppress protocol-related MRI bias while preserving pCR-relevant features. To enhance interpretability, we further incorporate large language model (LLM)-driven subgraph RAG module that retrieves clinically analogous historical cases and integrates retrieved evidence for pCR inference. We assemble a large-scale multicenter NAC breast cancer cohort for extensive validation, drawing from two public sources and three in-house centers.Results show that ClinRAG-GRAPH achieves AUCs of 0.815 on the internal test set and 0.774/0.712 on two external test sets, demonstrating robust pre-treatment pCR prediction across centers. The code is available at the anonymized https://github.com/miccai26-1181/ClinRAG-GRAPH.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1181_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/miccai26-1181/ClinRAG-GRAPH

Link to the Dataset(s)

N/A

BibTex

@InProceedings{DuaYao_ClinRAGGRAPH_MICCAI2026,
        author = { Duan, Yaofei AND Huang, Yuhao AND Zhang, Tianyu AND Gao, Yuan AND Han, Luyi AND Wang, Xin AND Xie, Xinyu AND Liang, Xinglong AND Lu, Chunyao AND He, Muzhen AND Pang, Patrick AND Sun, Yue AND Mao, Ning AND Tan, Tao AND Mann, Ritse},
        title = { { ClinRAG-GRAPH: Clinical-Prior Retrieval-Augmented Graph Model with Domain Adversarial Learning for Breast pCR Prediction } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16882},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper describes a RAG LLM arpproach that combines clinical, radiological, and pathological information for breast cancer NAC response prediction. The paper uses an adversarial leaning approach, a GCN, and a subsequent LLM to make a combined prediction. The methodology is evaluated on an internal test set and 2 external test sets.

  • 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.
    • Novel use of LLM combined with a GCN. Clear description of graph construction.
    • Interdisciplinary data, which is the SOTA for pCR in breast cancer
    • Multiple test sets, including 2 external.
    • Detailed theoretical information.
    • SOTA models used for comparison.
  • 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.
    • Not enough information provided regarding the encoders
    • Not enough information about the preprocessing of the datatypes.
    • Appears as though the LLM is used to provide an inference based on the patients proximity to nearest records in the training data. Not clear if a different LLM or another method of finding similar records would be as good.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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

    Figure 3 shows clustering structure in the external datasets. Are those sub-clusters correlated with specific labels or features? Is this representative of selection biases in those datasets that are not present in the internal dataset?

  • 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.

    (6) Strong Accept — must be accepted due to excellence

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Extensive performance evaluation of a modern approach. Clustering analysis, ablation study and SHAP analysis provide important insights. It is important that the data encoding is better described.

  • 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.

    Thanks for addressing the comments I have, my recommendation is positive and unchanged



Review #2

  • Please describe the contribution of the paper

    This paper addresses the clinically significant task of pretreatment pathological complete response (pCR) prediction for breast cancer patients undergoing neoadjuvant chemotherapy (NAC). The authors propose a novel multimodal framework (ClinRAG-GRAPH), with the following primary contributions: 1.Clinically Informed Multimodal Graph Modeling: The authors introduce a prior-guided, relation-aware graph convolutional network (PRAttn-RGCN). Moving beyond conventional feature concatenation across MRI, clinical, and pathological data, the method constructs an intra-patient directed graph. By encoding clinical practice guidelines into hierarchical edge priors, it enables more clinically grounded cross-modal fusion. 2.Domain Adversarial Learning for Multicenter Generalization: To mitigate heterogeneity arising from scanner variability and imaging protocols in multicenter DCE-MRI data, the paper implements a dual-branch domain-adversarial learning strategy. This approach effectively disentangles domain-invariant representations from site-specific noise, substantially improving robustness on independent external cohorts. 3.LLM-RAG Mechanism for Enhanced Interpretability: The framework integrates an LLM-driven subgraph retrieval-augmented generation (RAG) module. During inference, rather than outputting a scalar probability alone, the model retrieves historical cases with analogous graph attention patterns and multimodal signatures. This evidence-grounded reasoning significantly improves the clinical interpretability and trustworthiness of the predictions. 4.Extensive Multicenter Validation: The method is evaluated on a large-scale, multicenter breast cancer cohort comprising two public datasets and three institutional datasets. Comprehensive internal comparisons and rigorous external testing robustly demonstrate the model’s performance and translational potential under realistic distribution shifts.

  • 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 strongest aspect of this work lies in its rigorous, clinically grounded experimental design and thorough evaluation framework. Specific strengths include: 1.Large-Scale Multicenter Cohort and Rigorous External Validation: Cross-institutional generalization remains a critical challenge in medical AI. The authors assembled a substantial cohort across five institutions (two public, three internal). Rather than relying exclusively on internal cross-validation, they reserved two fully independent external centers (Ycenter and Qcenter) for unbiased out-of-distribution testing. This validation strategy provides robust empirical evidence of the model’s resilience to real-world distributional shifts. 2.Clinically Aligned Evaluation Metrics: Beyond relying on AUC alone, the authors report balanced accuracy, sensitivity, and specificity. This multi-metric approach appropriately accounts for the class imbalance inherent in clinical datasets and directly reflects diagnostic utility, particularly in balancing false-positive rates with the reliable identification of true pCRs. 3.Comprehensive Ablation Analysis: Despite the methodological complexity—integrating graph-based modeling, adversarial domain adaptation, and retrieval-augmented generation—the ablation experiments (Table 3) are methodically designed. By isolating the clinical prior (CP), graph attention mechanism (PR), domain adversarial component (ADV), and the RAG framework, the authors demonstrate that each element yields measurable performance gains, particularly on external cohorts. This systematic validation confirms that the architectural choices are substantively motivated and functionally integrated rather than arbitrarily combined.

  • 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.

    While the overall methodology is sound, the presentation of the proposed framework (particularly Figure 1) and the lack of critical implementation details require significant improvement: 1.Poor presentation clarity and severe text-figure mismatch in Figure 1: (a) Figure 1 lacks self-explanatory elements, leaving several visual components unexplained: 1) Unexplained abbreviations: Abbreviations such as Emb. and Tab. Enc. appear in the figure, but their full terms are not provided in the figure itself or the caption. 2) Unexplained symbols: The tag next to the nodes and the “scissors” symbol in Part C are not explained regarding their specific operations or states in either the figure or the caption. (b) There is a noticeable disconnect between the figure and the Methodology section, lacking terminology consistency: 1) Figure 1B uses descriptive module names like Self-Transformation, Linear Transformation, and Message Weighting, whereas Section 2.2 describes the process entirely using rigorous mathematical symbols. The visual elements lack a direct mapping to the equations, making it difficult to follow. 2) The figure introduces concepts not explained in the text: For example, the tag is shown in Figure 1A, but Section 2.1 (which introduces the prior graph construction) never mentions the concept of a “mask” or its specific implementation mechanism. 2.Missing implementation details of the MRI feature extractor: In Section 2.3, the imaging node initialization is defined as f-img(·). However, in the implementation details (Section 3.1), the authors only mention the cropping and concatenation sizes of the DCE-MRI images (192, 256, 256), omitting the specific architecture of the Imaging Backbone . This critical omission significantly hinders the reproducibility of the paper. 3.Privacy and compliance risks regarding LLM clinical deployment: The paper utilizes an LLM for retrieval and explanation but fails to specify whether it relies on calling an external API or local deployment. In real-world clinical Medical Image Computing (MIC) scenarios, transmitting sensitive data—including patient graph embeddings and multimodal features—to external LLM APIs (e.g., DeepSeek) poses significant privacy and compliance risks. Conversely, locally deploying a large-parameter LLM requires massive computational resources. The authors should clarify their deployment strategy. Furthermore, to truly demonstrate the clinical viability of the LLM framework, it is highly recommended to conduct validation experiments using a locally deployable, small-parameter LLM to assess whether the performance drop remains within an acceptable range.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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 well-motivated, clinically relevant multimodal framework with rigorous multicenter validation and thorough ablation studies. The methodological novelty, particularly the integration of clinical priors with graph attention and domain-adversarial learning, is clearly demonstrated. However, the manuscript requires minor revisions to improve figure-text consistency, clarify key implementation details (especially the imaging backbone), and address LLM deployment privacy considerations. Addressing these points would strengthen both reproducibility and clinical readiness, warranting a weak 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.

    The authors’ rebuttal has mostly addressed my concerns. They clarified the specific architecture of the MRI encoder, explained that the LLM used in the paper is API-based, and stated that during inference the LLM only receives de-identified embeddings and attention features rather than raw patient data, which helps mitigate privacy concerns. They also committed to adding explanations for the abbreviations and symbols in Figure 1.Overall, this paper addresses a clinically important problem, presents a method with a certain degree of novelty, and provides relatively comprehensive experimental validation. Taking into account the clarifications provided in the rebuttal, I am inclined to maintain my final decision of Weak Accept.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a graph-based framework for breast pCR prediction. First, it introduces a clinical-prior-guided relational attentional graph convolutional network to encode multimodal features. Second, it incorporates an adversarial learning strategy to effectively leverage multicenter data during training. Third, the paper integrates a large language model into the inference process to further enhance prediction performance 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.The proposed prior-guided graph convolutional network and adversarial learning strategy effectively improve the model performance.

    2.The experiments are conducted on multi-center data, providing strong evidence for effectiveness.

  • 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 proposed GCN largely follows standard design paradigms; therefore, its technical contribution mainly credits to clinical priors. However, the formulation and utilization of these priors require further clarification and experimental validation. For instance, why are the strong priors fixed? Are soft priors learnable? How does the initialization of these priors affect performance? What if use the R-GCN without priors, in addition to Table3? Additionally, it would be helpful to report the statistical properties (e.g., mean and standard deviation) of the learned priors after training.

    2.The generalization capability of the proposed method is not fully convincing. Although it significantly outperforms competing methods on the internal dataset and the external Ycenter, its performance drops substantially on the external Qcenter, where it only achieves marginal improvements, while other methods remain relatively stable across datasets. Moreover, the proposed adversarial learning strategy is intended to enhance robustness to multicenter data; however, Table 3 shows a performance decrease on the Qcenter when adversarial learning is applied. This observation requires further explanation.

    3.The use of the LLM requires additional clarification. Specifically, what data-dependent inputs are provided to the LLM (beyond the generic prompt illustrated in Figure 5)? Furthermore, if the LLM-based inference produces results that are inconsistent with those of the GCN, how are the final predictions determined?

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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

    1.What does τ represent in Eq. (2)?

    2.Table 2 should adopt a consistent formatting style (e.g., consistent use of commas and hyphens).

  • 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 proposed method demonstrates significant improvements on the internal dataset and the external Ycenter. However, further clarification of certain methodological details and a more thorough analysis of its generalization ability are still needed.

  • 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.

    The rebuttal nicely addresses my main concerns and I would recommend accept the work to MICCAI.



Author Feedback

We thank all reviewers (R) for their feedback. We have clarified the concerns and will improve the writing accordingly.

Q1.Lack description for image encoder. (R1,R2) Thank you for pointing this out. We omitted details of the MRI encoder in the manuscript. The MRI encoder is DenseNet-121, as specified in encoder.py of the anonymized code. We will add this clarification to manuscript.

Q2.LLM mechanism and decision rule. (R1, R3) We thank the reviewers for the opportunity to clarify. PRAttn-RGCN is the diagnostic backbone, while the LLM-guided retrieval branch only provides auxiliary evidence by generating a schema-constrained retrieval plan over fixed similarity routes. Thus, the framework is not DeepSeek-specific, and any schema-compliant LLM can be substituted. For each query, the LLM receives the prompt together with patient’s sub-graph embedding, MRI/clinical/pathological embeddings, and edge-level attention weights. Finally, the LLM cannot “disagree” with PRAttn-RGCN because it does not produce a prediction. Any difference arises only between the PRAttn-RGCN output and the neighbor-derived estimate, which are combined deterministically: p_final = b·p_R-GCN + (1-b)·p_neighbors), b=0.7. PRAttn-RGCN remains the dominant decision source. Retrieval can adjust borderline cases but cannot override a confident backbone prediction.

Q3.Clarification of clinical priors. (R3) Thank you for this insightful comment. Strong-prior edges are derived from breast cancer guidelines and expert consensus, representing clinically validated and stable medical facts. We keep them fixed to preserve such knowledge and reduce the risk of learning dataset-specific shortcuts under limited sample size or center bias. For less deterministic relations, soft priors are learnable but initialized with larger weights, while uncertain relations are weakly initialized and mainly learned from data. Priors are injected only as a first-layer log-bias to calibrate early R-GCN message passing, leaving deeper layers fully learnable. As shown in Table 3, adding clinical-prior improves AUC over vanilla R-GCN.

Q4.Qcenter generalization and adversarial learning. (R3) Thank you for raising this point. We agree that Qcenter is a more challenging external domain, where our method still achieves the best at AUC, bACC and SEN. The adversarial branch is applied only to MRI to reduce protocol-related imaging bias, not all sources of external distribution shift. Its impact on pCR prediction is therefore limited, especially because prior studies and our analysis suggest that pathological factors dominate pCR prediction. This is also supported by Fig. 4, where few highly ranked edges are image-related. We will clarify that adversarial branch improves robustness when MRI bias is a major domain-shift source.

Q5.Privacy risks regarding LLM. (R2) We used API-based LLM for three reasons. First, privacy risks are mitigated, as stated in Sec. 2.4, the LLM only receives de-identified embeddings and attention features, not raw patient data. Second, API-based use of properly de-identified medical data is already an established practice in healthcare AI workflows when privacy and governance requirements are met. Third, local deployment is limited by computational resources.

Q6.Fig. 3 clustering and selection bias. (R1) Thank you for this observation. Fig.3 aims to visualize binary class separation across internal and external cohorts, not to claim that external sub-clusters are free from selection bias. We will clarify this interpretation.

Q7.Writing details. (R2,R3) We will correct all typos related to Table 2; revise notation errors and ambiguities, such as clarifying that τ denotes the temperature coefficient and replacing it with t; and explain undefined abbreviations (Emb. as Embedding, Tab. as Table, and Enc. as Encoder). We will clarify as edge-prior mask and the scissors symbol denotes the gradient-reversal-based adversarial decoupling branch.




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 proposed framework integrates clinical-prior-guided graph networks with LLM-based RAG for interpretable breast pCR prediction. Its strengths lie in the rigorous validation on extensive multicenter cohorts and a sound architecture that balances domain-adversarial robustness with evidence-grounded reasoning. Through the review scores as 4 or higher, there are several concerns that should be addressed via a thorough rebuttal: 1) More details on formulation and utilization of clinical priors, 2) Explanation regarding performance drop in external cohort, 3) LLM concerns raised by the three reviewers

  • 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 adequately addressed the concerns raised. The reviewers are now unanimous in their decision.



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.

    Congratulations!



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 addressed the reviewers’ concerns regarding privacy and generalization. They also provided more explanatory details.



back to top