List of Papers Browse by Subject Areas Author List
Abstract
Automated radiology report generation for chest X-ray (CXR) is essential for reducing radiologists’ workload and enhancing diagnostic efficiency in pulmonary screening. While vision-language models (VLMs) have shown promise, most existing methods rely on holistic free-text supervision, which often provides highly concise, incomplete clinical signals and leads to factual hallucinations. To address this issue, we propose TGST, a Template-Guided Structured Training framework that reformulates report generation as a fine-grained, clinically-aligned visual question answering (VQA) task. Specifically, we first construct a tree-structured clinical template for pulmonary findings, decomposing complex reports into anatomical entities and their specific attributes. We then introduce a two-stage pipeline that maps original reports to the template and leverages a medical VLM to infer missing visual details, creating dense supervision signals. By supervising the model on these localized clinical queries, TGST enforces precise visual-textual alignment and penalizes reliance on linguistic priors. Extensive experiments on the IU-Xray dataset demonstrate that TGST significantly outperforms state-of-the-art methods. Notably, our framework achieves substantial gains in factual correctness (GREEN score) and node-level observation accuracy across multiple VLM backbones, highlighting its potential for reliable and systematic automated pulmonary screening.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1715_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{XiaZiq_Reformulating_MICCAI2026,
author = { Xiao, Ziqin AND Liao, Zehui AND Lu, Zilin AND Hu, Shishuai AND Leng, Shaoyi AND Zhang, Jingfeng AND Xia, Yong},
title = { { Reformulating Chest X-Ray Report Generation as Fine-Grained Visual Question Answering with Structured Clinical Report Templates } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a template-guided structured training framework that redefines report generation as a fine-grained, clinically aligned visual question answering task. The proposed method does not rely on global free-text supervision and reduces factual hallucinations. The results also highlight its reliability and application potential.
- 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 problem addressed in this paper is widespread: VLMs tend to generate hallucinations when producing reports. 2.The two-stage strategy employed in this paper is supported by strong theoretical grounding. 3.The approach of this paper is interesting, novel, and reasonable. 4.The experimental results validate 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.This paper does not discuss the limitations of the proposed hierarchical clinical template. Is this template easily transferable to diagnosis of other organs or anatomical regions? 2.The experimental results show significant improvements. Are these results reproducible? Which component of the design contributes most to the performance gain? 3.Does the proposed method have a preference for the choice of baseline models? The performance improvements vary substantially across different models. Does this imply that the effectiveness of the method is highly dependent on the choice of baseline model?
- 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 approach of this paper is novel and reasonable. However, the paper does not elaborate sufficiently on the methodological details. Moreover, the experimental results indicate that the method’s performance depends heavily on the choice of baseline model, which may pose challenges for result reproducibility.
- Reviewer confidence
Somewhat confident (2)
- [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
2.Strengths
1.the authors propose a template-guided structured training framework that reformulates report generation into a fine-grained VQA task. By decomposing reports into anatomical entities and attributes, the approach provides more precise and clinically meaningful supervision, reducing reliance on incomplete free-text signals.
2.the authors design a two-stage pipeline to generate dense and localized supervision signals by mapping reports to structured templates and inferring missing details. This strategy improves factual correctness and visual-text alignment, leading to significant performance gains over existing methods.
- 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.
2.Strengths
1.the authors propose a template-guided structured training framework that reformulates report generation into a fine-grained VQA task. By decomposing reports into anatomical entities and attributes, the approach provides more precise and clinically meaningful supervision, reducing reliance on incomplete free-text signals.
2.the authors design a two-stage pipeline to generate dense and localized supervision signals by mapping reports to structured templates and inferring missing details. This strategy improves factual correctness and visual-text alignment, leading to significant performance gains over existing methods.
- 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.
Major Weaknesses:
1.The method leverages a strong VLM to fill in missing parts of the template, providing additional supervisory signals. However, it is unclear how the correctness of the generated content is ensured. In particular, how does the approach guarantee that the completed content is grounded in the actual visual evidence, rather than introducing hallucinated information? Furthermore, the paper lacks intuitive examples demonstrating this process.
2.The basis for constructing the structured clinical template shown in Figure 1 is not clearly specified. It would be helpful to clarify whether the template is designed based on established clinical guidelines, expert annotation, or data-driven methods. The authors are also encouraged to indicate whether this template will be publicly released. In addition, its generalizability remains unclear. Specifically, can the proposed template be directly applied to other datasets, such as MIMIC-CXR, or does it require dataset-specific modifications?
3.It is unclear why experiments are not conducted on a more widely recognized and authoritative dataset such as MIMIC-CXR.
- 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 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?
see the weaknesses
- 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
This paper proposes TGST, a Template-Guided Structured Training framework for chest X-ray report generation. The method reformulates report generation as a fine-grained visual question answering (VQA) task using a tree-structured clinical template and a two-stage pipeline for structured supervision construction. The authors aim to improve factual correctness and reduce hallucination in medical report generation. Experiments are conducted on the IU-Xray dataset using multiple VLM backbones, showing improvements over baseline methods in both free-form report generation and structured template completion.
- 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.Clinically motivated problem formulation. The paper addresses an important and practical problem in radiology report generation, focusing on improving factual consistency and structured reasoning, which is highly relevant for clinical applications. 2.Structured supervision design. The proposed tree-structured template and decomposition into node-level attributes provide a reasonable way to introduce structured inductive bias into report generation. 3.Comprehensive experiments. The authors evaluate the method on multiple VLM backbones and report improvements across both free-form report generation and structured evaluation metrics, including factual correctness (GREEN score). 4.Consistent empirical gains. The proposed TGST framework shows stable improvements across different models, suggesting some generalizability of the approach.
- 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.Limited methodological novelty. The core components of the framework largely consist of existing techniques, including LLM-based information extraction, structured template construction, VQA reformulation, and standard LoRA fine-tuning. The contribution is mainly a combination of known ideas rather than a fundamentally new learning framework or algorithm. 2.Pipeline-style design. The method is heavily reliant on a multi-stage engineering pipeline. While effective, it is not clear whether the performance gains come from conceptual innovation or from improved data processing and supervision design. 3.Potential self-training bias in Stage II. The second stage uses a medical VLM to infer missing attributes, which may introduce model-induced bias or propagate hallucinations into the training supervision. This issue is not sufficiently analyzed or mitigated. 4.Limited theoretical insight. The paper lacks a deeper discussion on why the proposed decomposition improves learning beyond intuition. A more formal explanation of the learning dynamics or supervision advantage would strengthen the contribution. 5.Dataset limitation. Experiments are conducted only on IU-Xray, which is relatively small and may limit the generalizability of the conclusions.
- 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?
While the paper addresses an important clinical problem and demonstrates consistent empirical improvements, the methodological novelty is relatively limited and primarily lies in the engineering of a structured training pipeline. The lack of deeper algorithmic contribution and limited analysis of potential failure modes reduce the overall impact.
- 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
Author Feedback
We sincerely thank the reviewers and ACs for their constructive comments and for recognizing the value of our method.
GQ1(R1,R2)-Template Design and Generalizability: Our hierarchical clinical template was designed with clinicians based on established radiology reporting principles and common pulmonary findings, rather than dataset-specific heuristics. It is expected to generalize to other CXR datasets (e.g., MIMIC-CXR), although minor vocabulary adaptation may be needed due to reporting differences. While the current version focuses on pulmonary findings, TGST can be extended to other regions by redefining the template nodes. We are extending it to whole-chest structured reporting on larger datasets for a future journal release.
GQ2(R2,R3)-Stage II Self-training Bias: We agree that VLM-based template completion may introduce biased or hallucinated pseudo-labels. However, unlike free-text pseudo-report generation, our process is constrained by the structured template and closed-set VQA formulation. Stage-I anchors the template to explicit findings from the original reports, while Stage-II only fills missing attributes within predefined candidate spaces. Logical pruning and null assignment further reduce unsupported completions. Although hallucinations cannot be fully eliminated, the strong empirical improvements suggest that the generated supervision remains sufficiently reliable in practice. We will further incorporate grounding verification and pseudo-label quality control in the journal extension.
GQ3(R2,R3)-Limited Validation: We acknowledge that evaluation only on IU-Xray may limit the assessment of large-scale generalizability. We chose IU-Xray as a controlled benchmark to validate TGST in a pulmonary-focused setting. Unlike conventional report generation methods, TGST additionally requires hierarchical template construction, fine-grained attribute definition, and structured VQA supervision, which become substantially more complex in whole-chest and large-scale settings. Nevertheless, the consistent improvements across multiple VLMs provide encouraging evidence for the effectiveness of the proposed framework. We are extending TGST to whole-chest structured reporting and larger-scale datasets for a journal extension.
GQ4(R1,R3)-Performance Gain and Supervision Advantage: The performance gains mainly come from reformulating report generation into structured fine-grained VQA, rather than isolated engineering components. Conventional long-text supervision provides sparse and entangled signals, while TGST decomposes report generation into constrained attribute-level objectives, yielding denser supervision and more localized visual-text alignment. We agree that additional ablations would further strengthen the analysis and will include more detailed studies in the journal extension. All experiments use public VLMs/datasets and standard LoRA fine-tuning, and the structured template and code will be released.
R1Q1-Baseline Model Dependency: The results do not suggest that TGST depends on a specific backbone, as consistent improvements are observed across all evaluated VLMs. The varying gains mainly reflect differences in medical-domain adaptation and performance ceilings. General-domain VLMs such as InternVL and Qwen benefit more from the structured clinical supervision, while medical-domain models such as HuatuoGPT already possess stronger medical priors, leading to smaller but still significant gains.
R3Q1-Clarification of Novelty: Our contribution is not a new algorithm, but a clinically structured training paradigm for chest X-ray report generation. Specifically, we reformulate free-text report generation into template-guided fine-grained VQA with hierarchical clinical decomposition. This restructures sparse long-text supervision into explicit attribute-level clinical reasoning signals, enabling more localized visual-text alignment and improved factual consistency.
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.
This paper proposes a template-guided structured training framework that reframes report generation as a fine-grained VQA task to reduce hallucinations. Reviewers shared a similar impression of the strengths, including clinically motivated formulation, structured supervision design, and consistent empirical gains across models. Though some weaknesses were also pinpointed by the reviewers, e.g., unclear template generalizability to other organs/datasets; potential bias from VLM self-training in Stage II; experiments only on a small IU-X-ray dataset, all three reviewers lean towards accepting the paper. Overall, reviewers and AC acknowledge its practical value in reducing factual errors, given the clinical relevance. It will be helpful if the author could address the comments in the revision.