List of Papers Browse by Subject Areas Author List
Abstract
Automatic chest X-ray report generation can assist clinical reading, but vision–language models may hallucinate findings or miss subtle cues, and their localized evidence is often hard to inspect. We propose TraceCXR, an evidence-traceable framework that links generated clinical statements to concept-level node traces and localized visual evidence maps. TraceCXR constructs ProtoGraph, which makes clinical concepts evidence-addressable by associating them with disease-aware visual prototypes, so retrieved concepts can be traced back to candidate image evidence. Building on ProtoGraph, EviLoop Prompting retrieves patient-specific graph context from the image and grounds the retrieved cues to localized visual evidence; confidence-gated residual injection suppresses unreliable graph prompts. We further introduce Process-Supervised Node Planning to regularize clinically relevant node selection and concentrated evidence attribution during training. Experiments on IU X-ray, MIMIC-CXR, and CheXpert Plus improve report quality and clinical correctness, with stronger evidence-centric retrieval and competitive grounding and transfer performance. Code:
https://github.com/Blaise-H02/TraceCXR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3018_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Blaise-H02/TraceCXR
Link to the Dataset(s)
IU X-ray: https://openi.nlm.nih.gov/
MIMIC-CXR: https://physionet.org/content/mimic-cxr/
CheXpert Plus: https://aimi.stanford.edu/datasets/chexpert-plus
BibTex
@InProceedings{DonHan_TraceCXR_MICCAI2026,
author = { Dong, Hang AND Sun, Chao AND Yan, Hao AND Hu, Wei AND Du, Bo},
title = { { TraceCXR: Verifiable Chest X-ray Reports via Evidence-Addressable Graph Prompting } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses automatic radiology report generation by explicitly establishing the relationship between clinical evidence and textual statements.
- 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 strength of this paper lies in introducing clinical concepts as graph nodes and linking them with visual evidence, enabling the report generation process to perform reasoning that simultaneously verifies local information.
- 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 construction of the clinical graph is not described clearly enough. Specifically, how are the entities and relations extracted from reports further transformed into “concepts”? Are all entities and relations converted into concepts?
- Equation (3) defines the prototype bank, but its subsequent usage is not clearly explained. In addition, the relationship between the index obtained in Equation (6) and Equation (3) should be clarified.
- The role of Equation (8) needs to be further elaborated.
- In PSNP, P is defined as a pseudo target set, but its meaning and how it is computed remain unclear.
- Figure 3 shows that as the number of entities increases, all metrics first improve and then drop sharply around 400; the reason for this significant change should be explained.
- Based on the results in Table 2, using only the EviLoop module already achieves strong performance, while the addition of protograph and PSNP brings only marginal improvements. Therefore, the necessity and contribution of protograph and PSNP require further justification.
- 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.
(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 idea of explicitly validating visual evidence using concepts is novel; however, the paper exhibits several shortcomings in the presentation of its methodology.
- 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 proposes TraceCXR, which enhances visual traceability in report generation by building a ProtoGraph with visual concepts , retrieving from input image to graph nodes and from graph nodes to visual tokens through EviLoop Prompting. Process-Supervised Node Planning is further proposed as a regularizer
- 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 problems of unreliable and untraceable visual evidence in report generation is significant. 2.It is an interesting idea to build connection between visual concepts from predefined graph and visual patch tokens. 3.Comprehensive experiments and ablation studies demonstrate the effectiveness of the proposed method.
- 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.I am a little confused about the terms of “expose an auditable concept trace” and “lightweight RL planner to output an ordered trace.” in the Introduction section. It seems to me that the actual trace is the set of retrieved nodes, instead of an ordered path/graph with generated causal relations.
- Key hyperparameters in the retrieval process (number of selected nodes in visual graph retrieval top_K0, and expanded hops) are introduced but their values are not specified, and there is no sensitivity analysis for these hyperparameters as well. 3.(Minor) Missing references in Table 1.1
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The idea of connecting visual concepts from predefined graph and visual patch tokens to enhance reliability and traceability is interesting. However, the paper’s central claim of traceable report generation seems a bit overclaimed. The actual “trace” is a set of retrieved nodes with no causal relation.
- 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 TraceCXR, a framework designed to improve the transparency and reliability of automated chest X-ray report generation. The primary contribution is a closed-loop “retrieval-grounding” mechanism consisting of three components: ProtoGraph, which links clinical concepts to visual prototypes; EviLoop, which retrieves and grounds these cues into an LLM via confidence-gated prompts; and a node planning module for regularizing the reasoning trace. Unlike standard black-box VLMs, TraceCXR provides an auditable evidence trail for each generated statement, aiming to reduce hallucinations and assist clinical auditing.
- 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 focus on “verifiability” is timely and addresses a major hurdle for the clinical deployment of large vision-language models.
The proposed EviLoop prompting successfully integrates knowledge graph retrieval with visual grounding, moving beyond simple image-to-text generation.
The experiments are quite comprehensive, covering multiple standard benchmarks (IU X-ray, MIMIC-CXR, CheXpert Plus) and demonstrating consistent improvements in both linguistic and clinical metrics.
The inclusion of evidence-centric evaluations (like Trace Coverage) is a good step toward quantifying model interpretability.
- 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.
Incremental Novelty: While the integration is well-executed, the individual components—such as prototype-based retrieval and KG-augmented generation—are relatively well-established in the field. The “EviLoop” concept feels like a refined RAG (Retrieval-Augmented Generation) framework tailored for radiology, rather than a fundamental breakthrough in vision-language architecture.
Proxy for Verifiability: The paper makes strong claims about “verifiability,” but this is primarily validated through a proxy metric (Trace Coverage). It is unclear whether these evidence trails are actually helpful or intuitive for radiologists in practice. A small-scale human-in-the-loop study comparing the “traceable” reports against standard outputs would have significantly strengthened this claim.
Ambiguity in Baselines: In Table 1, the baseline performance for certain models (e.g., MedGemma) appears unexpectedly low compared to published results, which might make the relative gains of TraceCXR look more pronounced. Some clarification on the specific configurations of these baselines would be beneficial.
Localization Performance: While the localized visual evidence maps are interesting, it is not entirely clear if they outperform standard attention-based or Grad-CAM maps provided by simpler models in terms of diagnostic accuracy.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Overall, this is a strong and well-motivated work that addresses an important gap in medical report generation. The structured “evidence loop” approach is a sensible way to improve model transparency.
My main reservation is the incremental nature of some components and the lack of human-in-the-loop verification to support the “verifiability” claim. Strengthening the paper with a small-scale clinician preference study or more detailed visualization comparisons against state-of-the-art grounded VLMs (like CXR-LLaVA) would significantly boost the paper’s impact. Additionally, provide more detail on the training stability when using the optional RL planner in PSNP.
- 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?
TraceCXR is a solid and well-executed work that addresses the critical issue of model interpretability in medical report generation. The transition from black-box generation to an auditable evidence loop is commendable and shows clear empirical gains across several key benchmarks. However, the technical novelty is somewhat incremental, and the primary claim of “verifiability” lacks direct validation from clinical users. Given its strong empirical performance and high relevance to the MICCAI community, I lean towards a Weak Accept.
- 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 paper aims to improve the transparency of radiology report generation by making every statement “verifiable” through an explicit evidence trail. The framework introduces ProtoGraph, which converts clinical entities into evidence-addressable nodes using disease-aware visual prototypes. A retrieval module called EviLoop Prompting identifies patient-specific graph context based on visual similarity, while a Process-Supervised Node Planning module regularizes the reasoning path. The result is a report accompanied by a concept-level node trace and localized visual evidence maps to assist clinician verification.
Reviewers have agreed that the paper has significant contributions to be accepted. Please consider the reviewers’ comments for further improvement.