Abstract

Deep learning models can effectively use Rapid Evaporative Ionization Mass Spectrometry (REIMS) data for surgical margin assessment. However, their clinical adoption remains challenging due to limited generalization to operating room conditions. This difficulty arises because models are typically trained on labeled spectra collected from resected tissue samples, while they must operate on noisy, unlabeled data acquired directly during surgery. In addition, the black-box nature of deep learning models makes it difficult to understand and systematically improve their behavior. Concept-based learning offers a promising way to address these challenges by mapping raw measurements to human-understandable concepts. However, supervised concept-based approaches rely on concept annotations, which are difficult to obtain in complex mass spectrometry workflows. We propose Agent-Guided Concept Discovery, a framework that learns meaningful concepts directly from data without requiring predefined concept labels. During training, a reasoning agent refines semantic descriptions of the learned concepts and adaptively adjusts their weight based on diagnostic relevance. These concepts are further grounded using a biochemical knowledge graph to ensure consistency with known metabolic relationships. Across Skin and Breast Cancer datasets, our model improves balanced accuracy and sensitivity over the baseline. In a representative intraoperative case, it shows fewer false positives, indicating better generalization to surgical conditions.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5424_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/nooshinmaghsoodi/cem-qwen-reims

Link to the Dataset(s)

Our dataset is private, and we are not allowed to publish it.

BibTex

@InProceedings{MagNoo_AgentGuided_MICCAI2026,
        author = { Maghsoodi, Nooshin AND Jamzad, Amoon AND Policelli, Robert AND Farahmand, Mohammad AND Srikanthan, Dilakshan AND Kaufmann, Martin AND Ren, Kevin Y. M. AND Merchant, Shaila AND Varma, Sonal AND Walker, Ross AND McKay, Doug AND Rudan, John AND Fichtinger, Gabor AND Mousavi, Parvin},
        title = { { Agent-Guided Relational Concept Discovery: Toward Interpretable Surgical Margin Assessment } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16893},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a an approach to perform surgical margin assessment (benign/cancer) based on REIMS data. The method combines a foundation model for mass spectra data with concept-based learning that maps measurements to human-understandable concepts, whose relevance scores are incorporated as auxiliary supervision during training of the margin assessment. The resulting model is both adapted to the REIMS domain and more interpretable than the backbone foundation model.

  • 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 integration of a reasoning agent, grounded in medical knowledge, to add interpretability is timely and novel to the domain of REIMS data.
  • 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.

    Methodological

    • Eq. (2), Definition of p_{ik}: The text describes p_{ik} as representing the relative presence of concept k, yet it is defined as σ(s_{ik}), where σ conventionally denotes the sigmoid function. The sigmoid maps activations to (0, 1) but does not yield a relative presence across concepts. The authors should either correct the equation or reconcile the textual description with the actual formulation.
    • Missing regularization term in the loss function: Following Eq. (2), the authors introduce a regularization term based on the cosine similarity of the concept weights w_k. However, this term does not appear in the final loss function defined in Eq. (7). It is therefore unclear at what stage of training this regularization is applied, or whether it is applied at all. This discrepancy must be resolved.
    • Concept memory mechanism: The paper does not adequately describe how the memory of concept descriptions is managed. It is unclear whether all concepts ever generated by the agent are retained, or only the most relevant are stored. Furthermore, the storage mechanism itself is not specified: it is not clear whether concepts are maintained in an external memory module or simply within the LLM context window. This requires a precise technical description.
    • Alignment loss and relationship to [14]: The authors state that the alignment in Step 4 is “inspired by [14],” yet upon consulting that reference, the alignment loss appears identical to the one proposed there. If the method is directly adopted from [14], this must be stated explicitly. If meaningful modifications were made, the authors are obliged to clearly articulate what distinguishes their formulation from that of [14]. This also adds to the judgement of novelty of the paper.
    • Underdefined hyperparameter λ: The weighting term λ in Eq. (7) is never defined, nor is any guidance provided on how to set it. In the absence of released source code, this constitutes a significant reproducibility concern that must be addressed.

    Experimental

    • Missing baseline, REIMS foundation model [7]: The experimental comparison is limited to a Transformer (whose architecture specifications are not provided) and the DreaMS model [5]. Given that a foundation model for REIMS data exists and is cited as [7], the absence of a comparison against it is a conspicuous gap that the authors must either address or explicitly justify.
  • 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

    In addition to the major issues outlined above there are some minor improvements that would strengthen the paper:

    • Contribution framing: In the introduction, the first listed contribution is described as being achieved through the other two contributions. A contribution that is merely the product of other contributions does not stand as an independent one. The authors should reformulate this, either by removing it as a separate item or by reframing it as the overarching goal of the work.
    • Unexplained notation: The term “m/z regions” is introduced without explanation. While this notation is standard within the REIMS community, a brief clarifying sentence would improve accessibility for a broader readership. Additionally, it should be noted explicitly that z here refers to the charge state and is distinct from the embedding variable z defined in Eq. (1), to avoid notational confusion.
    • Performance gap across cancer types: The results suggest a larger performance gap between the proposed method and baselines in breast cancer than in skin cancer. The authors should discuss this discrepancy and, where possible, offer mechanistic or data-driven explanations for the differential behavior across tissue types.
  • 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 recommend a weak reject for the paper. Since the novelty is limited, due to the outlined major issue related to the alignment loss. In addition, the experimental baselines are insufficient to fully understand the added value.

  • 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 the authors for the detailed explanations in the rebuttal.

    I still have some concerns about how the regularization in step 2 is not playing a role in Eq (7). As far as I understand the model it is trained end-to-end (except the embedding generation from the DreaMS), therefore the regularization should play a role?

    All other points are addressed with the revision, therefore I change my rating to a “weak accept”.



Review #2

  • Please describe the contribution of the paper

    Addresses the clinical problem of real time tumour margin detection using mass spectroscopy and then the technical problem of training a classifier to be robust enough for good generalisation, using two ex vivo and one in vivo datasets. It does this by implementing an unsupervised learnable concept bottle neck model. Learned ‘concepts’ are then mapped to interpretable text outputs via an LLM which is sequentially guided by clinical guidelines, database query and a knowledge graph. This results in a more ‘generalisable’ model which is also interpretable.

  • 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.

    Well motivated, both the clinical problem and the technical element and generally well written and easy to follow.

    The proposed method is novel as far as I’m aware and the strongest element of the paper, having real clinical potential, and possible extensions to adjacent techniques (e.g. vibrational spectroscopy).

    Ablation studies assess the contribution of various elements, supporting conclusions.

  • 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 training regime seems well done but without the actual code I can’t be sure. One concern is that, for instance, one dataset has 693 burns from 91 patients. Was the data split performed at the level of the burn or the patient? If the former then it is likely that multiple burns from the same patient are in both the training and test set, violating assumptions of independence in the sets, resulting in over-fitting.

    There is some ambiguity in the description of hyperparameters in the paper, for instance the choice of K = 8.The way presented suggests this was pre-selected and fixed and not part of the hyperparameter optimisation? In general the training regime details are a bit sparse and not currently reproducible, but that should be addressed when the code is released.

    Although the proposed model does perform better, it might be a stretch to claim it is indeed more generalisable without a true external test set.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    Enjoyed reading the paper, thanks. Just some additional comments:

    Some small typos in the paper (e.g. REIMS vs reims).

    Equation 3 - I either don’t understand or there’s a mistake. I understand it as the scale parameter, alpha_k, is put through a sigmoid function to constrain it to {0,1}, but the notation indicates an identity function where alpha_k = sigmoid(alpha_k).

    “For each concept k, we measure the change in prediction when its contribution is removed from the classifier input.” How is the concept ‘removed’ – presumably the architecture is fixed so the concept vector can’t just be removed. Do you set alpha to 0?

    Figure 2 was not intuitive for me to understand and could benefit from more guidance for reader.

  • 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?

    Excellent use of guiding LLM outputs during the training process such that embedded representations become interpretable - not simply by throwing an LLM at it and hoping for the best, but a principled use of various structures (like a knowledge graph and database query).

  • 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.

    Points satisfactorily addressed.



Review #3

  • Please describe the contribution of the paper

    The paper proposes a novel framework that learns interpretable concepts from REIMS spectra without requiring manual concept annotations, by incorporating a reasoning agent into the training loop. A large language model–based agent iteratively refines concept semantics, assigns descriptions, and provides relevance feedback that directly influences model training via an alignment loss.

  • 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.Eliminating manual concept labeling is a strong practical advantage, especially in complex domains like mass spectrometry. 2.The integration of a reasoning agent into concept learning is innovative and goes beyond standard concept bottleneck models.

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

    1.The paper provides a high-level description of the agent-guided concept refinement process, but lacks sufficient implementation details (e.g., prompts, interaction protocol, update frequency), making the framework difficult to reproduce and evaluate rigorously. 2.While qualitative examples are provided, there is no quantitative metric assessing concept correctness, consistency, or faithfulness. 3.The paper relies on a relatively small LLM as a reasoning agent, but provides no evidence that its outputs are reliable enough for the claimed biochemical grounding and concept refinement. 4.The proposed “agent” largely combines retrieval, heuristic scoring, and LLM-based description, rather than functioning as a fully autonomous reasoning system. While its integration into training is interesting, it is unclear whether it offers advantages over simpler alternatives.

  • 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?

    The main factors influencing my score are related to concerns about the proposed agent framework. See the weakness section.

  • 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 rebuttal improves the methodological clarification, but my core concerns remain unresolved. In particular, the paper still lacks rigorous quantitative evaluation of concept correctness or faithfulness, and the reliability of the LLM-guided refinement process is not convincingly validated. Moreover, it remains unclear whether the proposed “agent” provides meaningful advantages over simpler retrieval or heuristic-based alternatives, as no ablation or comparative analysis is provided. Therefore, I remain unconvinced about the necessity and robustness of the proposed framework and maintain my recommendation for rejection.



Author Feedback

We thank all reviewers for their constructive feedback. The paper was recognized for its clinical novelty, motivation, writing quality, possibilities to extend to adjacent techniques (R2) and the novelty for moving beyond standard concept bottleneck models by integrating a medically grounded reasoning agent that eliminates manual concept labeling (R1&R3). We address the primary concerns below:

1.Quantitative Concept Evaluation, and the role of LLM (R3& Meta-R). Quantitative concept evaluation for REIMS data is challenging, as this data type lacks biochemical ground-truth concept labels. Our work is grounded in a computer-assisted-intervention (CAI) setting, with sparse, intraoperative mass spectrometry data; as such, we assess the clinical relevance of concepts using well-established biological evidence in the literature: hormone-receptor-related activation patterns and known breast cancer pathways. Additionally, we demonstrate the quantitative value of the learned concepts indirectly in Experiment 3 by showing they improve intraoperative margins over no-concept baselines. To ensure reliability, our non-autonomous reasoning agent grounds LLM-generated descriptions in retrieved metabolic evidence, knowledge-graph links, observed concept activations, and prior concept behavior to guide concept updates. This is useful in CAI settings where fixed-rule alternatives are difficult to define.

2.Alignment loss (R1). Our main contribution to the alignment loss is the use of relevance feedback provided by the reasoning agent and incorporating them into training for label-free concept discovery. To compare with reference [14], we do not assume predefined concepts. Instead, we learn latent concepts and continuously evolve them via agent feedback incorporated directly into the loss function. We use the same alignment principle as [14] to give concepts labeled highly relevant by the agent greater influence on prediction.

3.Recommended baseline (R1 & Meta-R). We use DreaMS [5] as our foundation-model baseline, as it is pretrained on a broad mass spectrometry dataset and, to our knowledge, there is no publicly available large-scale REIMS-pretrained foundation model. While [7] is related, it starts from DreaMS and fine-tunes all parameters on a small private ex vivo REIMS dataset, making it a dataset-adapted DreaMS model rather than a REIMS foundation-model baseline. We also include a Transformer baseline to show performance when training directly on the ex vivo data.

4.Memory mechanism (R1). The memory is an external record. It stores concept information from each epoch, including m/z regions, relevance feedback, and classification outcome. In later epochs, overlapping concepts from previous epochs are retrieved for each candidate concept and provided to the LLM to improve consistency.

5.Methodological clarification (R1). Regarding p_{ik} in Eq. 2, the sigmoid maps each raw concept score independently to a value between 0 and 1, producing a bounded activation per concept rather than a relative presence across concepts. Regarding the cosine-similarity term in Step 2, it is applied only at concept initialization to encourage diverse starting directions and plays no role in the loss in Eq. 7; we will replace “relative presence” and “regularization term” with “bounded concept activation” and “initialization diversity term” respectively.

6.Clarification on data splitting and generalization (R2). We stratify the ex vivo dataset patient-wise, so the training and test data are from independent, non-overlapping patients with no data leakage. The built models are further evaluated on an intraoperative patient case to show generalization from ex vivo to intraoperative use.

7.Reproducibility (R1-3 & Meta-R). Lambda and the number of concepts were selected on validation data. We will clarify implementation and baseline details and prompts in the revision, and will release the code upon acceptance.




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.

    There are mixed reviews (A, 2WR). Reviewers are concerned that experiments/analysis are limited (quantitative metric, reliability, and insufficient baselines). Implementation details are not sufficient.

  • 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 interpretable learning from REIMS data for intraoperative decision support. Its main merit is the integration of agent-guided concept refinement into a concept-learning framework, reducing the need for manual concept labeling while aiming to produce clinically meaningful and biochemically grounded representations. This is a potentially valuable direction for the MICCAI community, especially for computer-assisted intervention and interpretable medical AI. The rebuttal resolves several important concerns. The authors clarified the patient-wise data split, reducing concerns about data leakage. The paper still lacks a quantitative evaluation of concept correctness, consistency, and faithfulness. Nevertheless, I find that the clinical novelty, interpretability-oriented formulation, and promising empirical results outweigh these remaining weaknesses. I recommend accept.



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.

    Following the rebuttal, two reviewers were satisfied, noting that the authors clarified the data split, baseline choice, memory mechanism, alignment loss, and several implementation details. However, one reviewer maintained a Reject recommendation, emphasizing that the paper still lacks rigorous quantitative evaluation of concept correctness and faithfulness, and that the reliability of the LLM-guided refinement process remains insufficiently validated. Overall, while some interpretability claims require stronger validation, the rebuttal addressed many methodological ambiguities and the proposed framework is clinically motivated and novel. I therefore recommend acceptance.



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 corresponding paper integrates concept-based learning with a reasoning agent for surgical margin assessment. During the rebuttal, some concerns regarding baselines and data splitting have been addressed; some concerns remain regarding compared methods (e.g., heuristic aggregation and concept validation). Given the overall positive assessment and reasonable reply regarding concept validation in the rebuttal, I lean towards a weak accept.



back to top