List of Papers Browse by Subject Areas Author List
Abstract
Automatic generation of radiology reports seeks to reduce clinician workload while improving documentation consistency. Existing methods that adopt encoder-decoder or retrieval-augmented pipelines achieve progress in fluency but remain vulnerable to visual-linguistic biases, structural misalignment, and lack of explicit multi-hop clinical reasoning. We present NeuroSymb-MRG, a unified framework that integrates NeuroSymbolic abductive reasoning with active uncertainty minimization to produce structured, interpretable reports. The system maps image features to probabilistic clinical concepts, composes differentiable logic-based reasoning chains, decodes those chains into templated clauses, and refines the textual output via retrieval and constrained language-model editing. An active sampling loop driven by rule-level uncertainty and diversity guides targeted training optimization and promptbook refinement. Experiments on standard benchmarks demonstrate consistent improvements in structural coherence and interpretability and standard language metrics compared to representative baselines.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6461_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{FuRon_NeuroSymbMRG_MICCAI2026,
author = { Fu, Rong AND Lyu, Yiqing AND Meng, Chunlei AND Qi, Muge AND Jin, Yabin AND Zhao, Qi AND Bao, Li AND Gao, Juntao AND Shi, Fuqian AND Dey, Nilanjan AND Luo, Wei AND Zhang, Wenxin AND Fong, Simon},
title = { { NeuroSymb-MRG: Differentiable Abductive Reasoning with Active Uncertainty Minimization for Radiology Report Generation } },
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
This paper proposes the NeuroSymb-MRG framework to address key challenges in radiology report generation, including visual-linguistic biases, factual inconsistency, and the lack of explicit multi-hop reasoning. The method maps visual features into probabilistic clinical concepts, then generates structured intermediate representations through differentiable neuro-symbolic reasoning chains. It further employs multi-agent orchestration to coordinate visual evidence, knowledge constraints, generation modules, and conflict verification, and produces structured reports through rule decoding, retrieval-augmented draft generation, and constrained language model refinement. In addition, the paper introduces an active sampling and clinical feedback mechanism based on rule-level uncertainty. Experiments on MIMIC-CXR and IU X-ray show higher language-generation metrics than a range of baselines.
- 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.This paper tackles several common limitations of existing methods, including visual-linguistic biases, factual inconsistency, and the lack of explicit multi-hop reasoning, and proposes a unified framework centered on neuro-symbolic reasoning. The problem is meaningful and well motivated in the context of clinical report generation. 2.The most notable strength of the paper is its attempt to move beyond the conventional end-to-end paradigm for radiology report generation by introducing an explicit neuro-symbolic reasoning stage. Through the pipeline of “clinical concept prediction -> rule composition -> clause generation”,the paper incorporates a structured reasoning process into report generation. 3.Another task-specific highlight is the active uncertainty minimization mechanism based on rule-level uncertainty. This design combines uncertainty estimation with clinician-in-the-loop refinement, which is well aligned with the practical needs of medical generation tasks. 4.The paper also provides relatively comprehensive component-level ablation studies. In particular, the authors analyze the roles of core modules such as the differentiable logic layer, rule-guided decoder, active uncertainty minimization, retrieval, and verifier.
- 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 core claims of the paper are improved factual consistency and clinically grounded generation. However, the main experiments still rely primarily on surface-level text overlap metrics such as BLEU, ROUGE-L, and METEOR, while lacking more direct evidence from factual consistency metrics, clinical correctness evaluation, or expert human assessment. As a result, the current experimental results are still insufficient to fully support the paper’s central claims, especially stronger claims such as reducing contradictions and improving clinical grounding. 2.The paper integrates many components into a single system, including the visual backbone, concept predictor, differentiable logic, rule-guided decoding, retrieval augmentation, LLM refinement, verifier, knowledge agent, and active uncertainty minimization. While this makes the framework comprehensive, the methodological storyline remains somewhat unfocused, and it is difficult to quickly identify what the paper’s single most important methodological contribution is. At its current stage, the work reads more like a carefully designed pipeline than a sharply focused method paper. The novelty boundary and contribution hierarchy would benefit from further clarification. 3.The specific mechanism of multi-agent orchestration is still not sufficiently clear. Although the paper introduces multiple specialized agents and presents a conflict resolution mechanism based on confidence aggregation, it does not adequately explain the input/output of each agent, their invocation order, or how they actually interact throughout the reasoning and report generation process. Based on the current methodological description, this part appears closer to coordination and packaging of multiple functional modules than to a strict agent-based learning framework. It is therefore difficult to judge whether multi-agent orchestration constitutes a core algorithmic contribution or mainly serves as a high-level system organization strategy. 4.The connection between active uncertainty minimization and promptbook refinement still requires further clarification. The abstract, contribution summary, and framework figure state that rule-level uncertainty can drive clinician-in-the-loop adjudication and promptbook refinement. However, the paper mainly elaborates on the definitions of uncertainty estimation and active sampling, while providing only limited explanation of the representation of the promptbook, how it is updated, and how it concretely operates in Module 2.Therefore, this part currently reads more like an appealing high-level design idea, while its operational details and actual contribution still need to be made more specific.
- 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
This paper presents a unified neuro-symbolic framework for radiology report generation, aiming to alleviate the problems of factual inconsistency and insufficient reasoning in traditional end-to-end methods through explicit concept prediction, rule composition, and uncertainty-aware refinement. The overall direction is meaningful, the method is fairly well designed, and the ablation study is relatively thorough. In my view, there are three main aspects that need to be strengthened. First, the current experiments still do not provide sufficiently direct support for factual consistency, clinical grounding, and contradiction mitigation, because the main evaluation still relies largely on surface-level metrics such as BLEU, ROUGE-L, and METEOR. Second, the method integrates many modules, which makes the contribution hierarchy not yet sufficiently clear; I would encourage the authors to more explicitly highlight the single most central methodological contribution in the revision. Third, the specific operational mechanism of multi-agent orchestration and promptbook refinement still needs to be elaborated further, especially the input-output relationships among modules, their invocation order, and their concrete impact on the final generated reports. If these issues can be addressed more clearly in the rebuttal or in a revised version, the paper would be more convincing.
- 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?
I give a Weak Accept because the paper proposes a relatively complete neuro-symbolic framework to address factual inconsistency and the lack of explicit reasoning in radiology report generation, and it attempts to move beyond the traditional end-to-end generation paradigm through a structured pipeline of “clinical concept prediction -> rule composition -> clause generation”.In addition, the active uncertainty minimization mechanism driven by rule-level uncertainty, together with the relatively comprehensive ablation study, are also strengths of the paper. At the same time, however, the paper’s central claims—especially improved factual consistency, reduced contradictions, and clinically grounded generation—still lack sufficiently direct experimental support, since the main evaluation remains focused on text-overlap metrics such as BLEU, ROUGE-L, and METEOR.Moreover, the method contains many components, which makes the overall contribution hierarchy and novelty boundary somewhat unclear.The specific mechanisms of multi-agent orchestration and promptbook refinement also need to be explained more clearly.Therefore, I believe this work has certain potential and is slightly above the acceptance threshold, but it still depends considerably on the rebuttal to clarify these key issues.
- 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 #2
- Please describe the contribution of the paper
The paper describes a novel approach to apply differential abductive reasoning in the task of radiology report generation. A framework, NeuroSymb-MRG, is proposed to realise abductive reasoning and uncertainty minimisation.
- 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 main strengths include:
- the application of differential abductive reasoning in radiology report generation.
- a sophisticated framework to combine the above with uncertainty minimisation
- a comprehensive set of experiments which show consistent improvements from the proposed model
- 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 main weakness is the use of evaluation metrics. For radiology report evaluations, it is well known that Rouge/METEOR scores are limited as they do not deal with semantics and contexts (eg negations) well. It would be important to see the use of other more clinically relevant scores like RATE score, GREEN or similar.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission 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.
(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 proposed approach is novel and experiments confirm the efficiency. I believe presenting this work at MICCAI would be beneficial to the community and stimulate good discussions along the methodology and applications.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper tackles the well-known issue of hallucinations and lack of interpretability in automated radiology report generation. The authors propose NeuroSymb-MRG, a framework that moves away from pure black-box text generation. Instead, it extracts visual features, maps them to clinical concepts, and then passes them through a differentiable “neuro-symbolic” logic layer that applies soft logical rules (AND, OR, NOT). The output of this logic layer is then used to retrieve relevant text fragments and fill in report templates. To make training more efficient and reliable, they also introduce an active learning loop that uses Monte Carlo dropout to estimate uncertainty, focusing the model’s learning on the most ambiguous cases.
- 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.
Combining differentiable logic with retrieval-augmented generation (RAG) and active learning is a really fresh take on report generation. It directly addresses the “black box” nature of current LLM-based approaches, which is a huge barrier to clinical adoption. By forcing the model to route its decisions through explicit, soft logical rules before generating text, the system provides a clear audit trail for why a specific sentence was generated. The model achieves very competitive scores on both MIMIC-CXR and IU X-ray, outperforming several recent and strong baselines like MRG-LLM.
- 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 relies entirely on standard NLP metrics (BLEU, ROUGE, METEOR). In the medical domain, these metrics are notoriously misleading. A generated report that says “There is a pneumothorax” and a ground truth that says “There is no pneumothorax” will have a very high BLEU score, but a catastrophic clinical error. You really need to include clinical efficacy metrics like CheXbert (F1-score for clinical observations) or RadGraph to prove the logic layer is actually improving factual correctness. 2.In the introduction and the ablations (Table 3), you mention a “learned feedback simulator to accelerate clinician-in-the-loop refinement.” However, looking through Section 3, I can’t find a clear explanation of how this simulator actually works. How is the clinician’s feedback simulated during the training loop? What data is it trained on? This needs a dedicated paragraph. 3.How are the diagnostic rules (the soft trees mentioned in Section 3.5) initialized? Do clinicians have to manually define the skeleton of these rules for every possible disease beforehand, or does the network learn the tree structures entirely from scratch? If it’s manual, I worry about how scalable this is to hundreds of rare findings. 4.Running 5 Monte Carlo dropout passes for active sampling during training sounds quite computationally heavy. It would be helpful to include a brief discussion on the training time/memory overhead compared to a standard end-to-end baseline.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission 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.
(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 integration of neuro-symbolic logic into radiology report generation is highly relevant to the MICCAI community, as it directly addresses the critical need for interpretability and hallucination reduction in medical AI. The architecture is well-thought-out, and the ablation studies are incredibly thorough, proving the value of the proposed components. The main factors holding it back from a perfect score are the reliance on purely lexical metrics (BLEU/ROUGE) instead of clinical accuracy metrics, and some missing methodological details regarding the feedback simulator and rule initialization. However, the core idea is strong enough, and the empirical results are solid enough, that I believe it should be accepted.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 core problem addressed is the lack of interpretability and factual reliability in automated medical report generation. Current “black-box” models often hallucinate or fail to follow logical clinical reasoning. The proposed NeuroSymb-MRG framework attempts to solve this by forcing the model to follow structured, logical steps before producing text.
The reviewers agree that the shift from end-to-end learning to a neuro-symbolic approach (logic-based reasoning) is a significant and “fresh” contribution. It creates a clearer audit trail for clinical decisions.
While the framework includes an “uncertainty” mechanism meant for clinician interaction, the actual utility is questioned because the evaluation relies on word-matching rather than clinical accuracy.
The paper presents a strong, technically sound framework that is highly relevant to the MICCAI community. However, the disconnect between the “factual consistency” claims and the “text-matching” evaluation must be addressed.