List of Papers Browse by Subject Areas Author List
Abstract
Interpretability is essential for trustworthy medical image diagnosis. However, existing concept-driven interpretable methods have key limitations: Concept Bottleneck Models (CBMs) require scoring all predefined concepts during both inference and manual intervention, imposing a substantial burden on clinicians, while rationale-based generative approaches often select concepts by class discriminability, which can drift from diagnostic ontologies. To address these issues, we propose Neuro-Symbolic Rule Distillation (NeRD), a framework that produces efficient, ontology-grounded reasoning chains that are sufficient yet non-redundant, without manually crafting diagnostic rules. Experiments on three datasets demonstrate strong diagnostic performance and interpretability, and blinded expert evaluation confirms the clinical plausibility of NeRD rationales. Our method further enables a first intervention study for Multimodal Chain-of-Thought-based diagnosis, achieving efficient and effective concept-level intervention. Our code is available at: https://github.com/HongxiY/NeRD.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2430_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HongxiY/NeRD
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanHon_NeRD_MICCAI2026,
author = { Yang, Hongxi AND Jiang, Yiwen AND Yan, Siyuan AND Chow, Jamie AND Li, Eunis AND Poon, Charlotte AND Fong, Stephanie AND Zhao, Xiangyu AND Mehta, Deval AND George, Yasmeen AND Ge, Zongyuan},
title = { { NeRD: Neuro-Symbolic Rule Distillation for Efficient Ontology-Grounded Chain-of-Thought in Medical Image Diagnosis } },
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 NeRD, a neuro-symbolic framework that distills logical rules into compact, ontology-grounded MCoT reasoning chains for medical image diagnosis. The method avoids exhaustive concept scoring (as in CBMs) and class-discriminability drift (as in WISE). The main contributions include automatic rule induction via LogicCBM, a three-step distillation process (activation, simplification, grounding) to generate non-redundant case-specific rationales, a blinded expert evaluation of clinical plausibility, and a simulated concept-intervention study.
- 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 method is novel and well motivated, bridging rule‑based interpretability and generative MCoT reasoning.
Generated MCoTs are compact (≈6 concepts per case) and shown to be clinically plausible by expert assessment.
Experiments cover diagnostic accuracy, concept fidelity, expert evaluation, and intervention efficiency. Code is provided.
- 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 abstract and introduction claim “first expert‑in‑the‑loop study”. However, Experiment 3 uses ground‑truth annotations to automatically replace concepts – no real clinician is involved. This is an oracle simulation, not a human‑in‑the‑loop study. Can the authors clarify or correct this claim?
In Experiment 1, the 8 cases per evaluator are sampled from the training set. Because rule induction already fits the training distribution, this may bias the results. Would the findings hold on the test set?
The paper does not compare with some recent concept bottleneck works, such as Evidential Concept Embedding Models (MICCAI 2024) and Concept Rule Learner (MICCAI 2025) by Gao et al. Would the authors add a discussion or experimental comparison?
- 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
The method currently uses ground‑truth concepts for rule grounding. An end‑to‑end evaluation with predicted concepts would strengthen the practical relevance. Also, the “ontology‑grounded” claim could be better supported by incorporating ontological priors earlier in the rule induction process. Minor issues: the zero‑shot prompt is not provided; there is a grammar error in Sec. 2.2 (“an selected rule” should be “a selected rule”).
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a novel method, but two major flaws prevent acceptance in its current form: (1) the claimed “expert‑in‑the‑loop” study is not what the authors actually did (no real clinician); (2) expert evaluation uses training cases, which may bias the results.
- 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
Review #2
- Please describe the contribution of the paper
The paper proposes NeRD, a neuro-symbolic framework for generating compact, ontology-grounded multimodal chain-of-thought rationales for medical image diagnosis. Instead of selecting concepts mainly by discriminability, the method first induces explicit logical diagnostic rules over clinical concepts, then distills activated rules into case-specific reasoning chains through rule selection, concept grounding, and finally structures these into MCoTs. The key contribution is to bridge rule-based concept interpretability with generative multimodal reasoning in a way that is intended to be more clinically aligned, non-redundant, and intervention-friendly.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper has several strengths. First, it tackles a clinically meaningful interpretability problem. Second, the proposed pipeline is technically interesting: it combines induced rule activation, logical simplification, and concept grounding to build more structured rationales rather than relying purely on class-discriminative concept selection. Third, a strong blinded human expert plausibility study, and a clinician-intervention experiment. Fourth, the empirical results are solid: NeRD is competitive diagnostically and improves over baselines on both datasets.
- 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.
My main concern is that the evaluation is not strong enough for the paper’s central claims about interpretability and clinical usefulness. The main quantitative evaluation uses only diagnostic accuracy/F1 and a fairly weak interpretability metric: “concept accuracy” is computed only over the concepts that the model explicitly mentions in its generated chain-of-thought, rather than evaluating completeness, or alignment against the full clinically relevant concept set. A second concern is that the experimental scope is narrow. This makes it difficult to judge how well the framework would transfer to broader medical imaging settings, especially tasks with richer label spaces, more complex ontologies, or longer-form outputs. A third concern is that the framework depends on a fairly strong problem setup (concept-annotated datasets, relies on an explicit concept ontology and concept group structure) that may be non-trivial to reproduce.
- 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.
(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?
Although the paper is well motivated and the proposed rule-distillation framework is interesting, I do not think the empirical validation is strong enough to support the paper’s broader claims about interpretability. Overall, I see this as a promising paper with good ideas, but in its current form I would place it slightly below the bar; it feels closer to an interesting workshop-style paper unless the authors can substantially strengthen the generalization case.
- 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
Review #3
- Please describe the contribution of the paper
The paper introduces NeRD, a neuro-symbolic framework for generating compact, ontology-grounded multimodal chain-of-thought (MCoT) rationales for medical image diagnosis. The key idea is to first induce logical diagnostic rules over clinical concepts using LogicCBM, then distill the activated rules for each case through rule selection, logical simplification, and concept grounding, and finally convert the result into a concise rationale used to fine-tune an MLLM. The intervention study shows that NeRD supports concept-level correction more efficiently than a CBM baseline.
- 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.I enjoy the nice figures, which help to understand the work better. 2.The method is reasonable to me, and a solid contribution to explainable AI in the medical domain. 3.The method is novel. It uses induced logical rules, simplifies them into a compact form, resolves disjunctions through concept grounding, and then organizes the resulting facts into supportive and refutational evidence.
- 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.It seems that the paper is polished by a LLM, and the writing style is hard to follow. Could you please restate your contribution again, and why it is different from the previous work WISE? 2.The empirical scope is still fairly narrow and the validation is limited to two dermatology datasets. I wonder why not focus on other domains except for skin lesion diagnosis?
- 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 method is novel and good, but i don’t like the writing style, which is a little bit hard to follow.
- 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.
I think the authors have made some progress in answering the reviewer’s questions; however, specifically to me, they didn’t provide details on how to make their paper stronger in the final version. Thus i recommand reject. However, i recognize the method of this paper, so i believe it’s on the borderline.
Author Feedback
We sincerely thank all reviewers for recognizing the novelty and motivation of our work. We have addressed their valuable comments below: 1.R1Q1 Expert-in-the-loop Claim We agree that Exp.3 did not involve live clinical interaction and therefore will revise our claim to reduce ambiguity. However, we would like to emphasise that intervention is indeed driven by human annotated input, as Exp.3 has evaluated how efficiently expert verified concept corrections rectify model errors. NeRD therefore retains strong potential for efficient and effective human-guided concept-level intervention, and this refinement does not affect our core claims/conclusions. 2.R1Q2 Human Eval in Exp.1 Our evaluation has two goals: (1) verify the clinical faithfulness of constructed MCoTs used for training (Exp.1); (2) assess if MLLMs can learn from them (Exp.2). Since induced rules on the training set but not the test set are used for subsequent MLLM training, Exp.1 primarily reports training-set results. We additionally evaluated sampled test cases as a sanity check and observed consistent trends (e.g. Derm7pt: All-Concepts/WISE/NeRD/NeRD-RD 0.75/1.25/1.50/1.00), which will be added to the final version. 3.R1Q3 Comparison with Evi-CEM & CRL Due to page constraints, we have only included a subset of CBM baselines in the paper. Comparisons of NeRD with Evi-CEM (MICCAI24) and CRL (MICCAI25) on two datasets are as follows (ACC/Intp.,%): Derm7pt: Evi-CEM 78.99/78.31, CRL 77.97/73.99, NeRD 80.76/76.76; F17k: Evi-CEM 88.04/92.66, CRL 88.51/92.55, NeRD 88.66/94.10.NeRD consistently achieves superior performance across almost all evaluation dimensions, which demonstrates its effectiveness. Results will be added in the final version. 4.R2Q1 Interpretability Metric Consistent with standard concept-based evaluation(e.g. CBM&WISE), we report concept accuracy against human annotations. We further assess MCoT clinical plausibility through a blinded expert study (Exp.1) using only MCoT without images. For completeness of concepts, NeRD outperforms the full concept set, suggesting that NeRD retains diagnostically relevant information while avoiding noise that may dilute salient evidence and reduce diagnostic confidence. Whereas for non-redundancy, NeRD-RD (randomly removing 1–2 concepts) degrades performance, confirming our selected concepts are necessary for correct diagnosis. 5.R2Q2 & R3Q2 Experimental Scope We extended the evaluation to the WBCAtt dataset (NeurIPS23) for white blood cell classification, a new medical imaging domain with richer labels (5 classes) and complex ontologies (24 morphological attributes). To simulate a challenging scenario, the model is trained using 30% training samples (n=1030) and evaluated on the official test set (n=3099). Main results (ACC/Intp., %): JointCBM 96.10/86.33, SequentialCBM 95.39/86.26, Zero-shot 35.43/36.91, WISE 96.13/85.87, NeRD (Ours) 96.64/92.81.Our NeRD surpasses all competing methods on WBCAtt, demonstrating strong generalization across diverse modalities, rich label spaces, and complex ontologies. 6.R2Q3 Generalizability We would like to emphasise the broad applicability of NeRD with these following reasons: (1) NeRD can operate in annotation-free settings where concepts can be generated by LLMs and pseudo labeled by CLIP. (2) NeRD does not consume any external ontology/knowledge base as input. The diagnostic logic is induced automatically, which is one of our key contributions. (3) Concept group structure is not required. Derm7pt has concept groups whereas F17k does not, and our pipeline operates on flat, ungrouped concepts in both datasets. 7.R3Q1 Difference from WISE WISE (Fig. 1b) uses decision trees to pick class-discriminative concepts, which may favor shortcuts that drift from diagnostic ontologies; NeRD (Fig. 1c) induces explicit logic rules and distills the activated ones per case, yielding ontology-grounded, sufficient, non-redundant rationales. We will improve readability in the final version.
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.
All reviewers recognize the merits of the paper, while they also have consistent concerns regarding the empirical evaluations and generalizability of the proposed method.
- 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 rebuttal addressed most key concerns, e.g., those regarding generalizability and experimental evaluations.
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.
The paper presents a novel and well-motivated neuro-symbolic framework for compact, ontology-grounded MCoT rationales, and the reviews consistently recognize its methodological contribution, clinical relevance, reproducibility through released code, and promising empirical evidence, including expert plausibility and intervention-oriented evaluation. Since concerns remain about the breadth of validation, reliance on concept annotations/ontologies, and the need to temper some clinical-usefulness claims, the work is placed marginally above the MICCAI acceptance bar.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
Two reviewers reject (R2=3, R3=4→Reject). R3 downgraded from 4 to Reject after rebuttal because ‘authors didn’t provide details on how to make their paper stronger.’ The ‘expert-in-the-loop’ claim is misleading (actually oracle simulation).