List of Papers Browse by Subject Areas Author List
Abstract
Automated Late Gadolinium Enhancement (LGE) MRI report generation is clinically important for supporting structured cardiac assessment and improving diagnostic efficiency. However, existing methods predominantly rely on CLIP-style global vision–language alignment, which captures coarse image-level semantics but fails to model region-level scar enhancement patterns and cross-slice structural continuity essential for reliable LGE interpretation.
To address this limitation, we propose Global–Local Structure-Aware Alignment for LGE Report Generation (GLSA-RG). The framework introduces a unified Structural Graph Representation to encode intra-slice regional interactions, cross-slice continuity, and structured textual dependencies. Building upon these representations, a novel Structure-Aware Graph Alignment mechanism performs fine-grained multi-instance contrastive alignment, complementing global volume–report supervision. This global–local alignment strategy enables explicit structure-consistent cross-modal correspondence while preserving overall semantic coherence.
Extensive experiments on two clinical datasets demonstrate that GLSA-RG better captures subtle myocardial enhancement patterns and maintains cross-slice structural consistency, resulting in clinically more reliable LGE MRI reports.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3705_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/luxiaozhou89-netizen/GLSA-RG
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhoChe_GlobalLocal_MICCAI2026,
author = { Zhou, Chenyang AND Tian, Yun AND Lu, Wenqi AND Hothi, Sandeep S. AND Zheng, Jinyu AND Zhao, Shifeng AND Duan, Jinming},
title = { { Global-Local Structure-Aware Alignment for Automated LGE MRI Report Generation } },
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 the clinically important but relatively underexplored task of automated LGE MRI report generation. It proposes GLSA-RG, a framework that explicitly models intra-slice regional interactions, cross-slice structural continuity, and structured textual dependencies through a unified graph representation, and further combines fine-grained structure-aware graph alignment with global volume-report alignment to improve report consistency and clinical reliability. The paper also introduces a two-center paired LGE MRI-report dataset and validates the method on both internal and external test sets.
- 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 paper presents a task-specific and clinically relevant framework that explicitly models local enhancement patterns, cross-slice continuity, and structured textual dependencies for LGE report generation. (2)The experimental study is reasonably comprehensive, including two-center data, external validation, main comparisons, and multiple ablations, with consistent improvements over the reported baselines.
- 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)While well designed, the method is still largely built on existing ideas in 3D medical vision-language modeling, graph-based structure modeling, and contrastive alignment, so the novelty is somewhat incremental. (2)The evaluation relies heavily on automatic text metrics and GPT-5-based scoring, with limited human clinical assessment, error analysis, and ethics-related documentation.
- 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?
This paper presents a clinically relevant and reasonably well-motivated framework for LGE MRI report generation, with consistent gains on both internal and external datasets. The main limitations are that the methodological novelty is somewhat incremental and the evaluation relies mostly on automatic metrics and LLM-based scoring rather than stronger human clinical validation. Overall, I believe the paper makes a useful contribution and is slightly above the acceptance threshold, although there is still room to strengthen the evidence and presentation.
- 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 main contribution of this paper is the proposal of a global–local structure-aware alignment framework (GLSA-RG) for automated LGE MRI report generation. Instead of relying only on coarse global image–text alignment, the method explicitly models fine-grained anatomical and pathological structure through unified visual and textual graph representations, including intra-slice regional relations, cross-slice continuity, and standardized myocardial segment dependencies. It further introduces structure-aware cross-modal alignment to better connect local imaging patterns with report content, aiming to generate more clinically consistent and reliable LGE reports. In addition, the paper presents experiments showing improved performance over prior methods and contributes a multi-center paired LGE MRI–report dataset.
- 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 task motivation is well justified. Automated LGE MRI report generation is clinically meaningful, and the paper clearly identifies the unique challenges of this task, especially the need to capture subtle local enhancement patterns and cross-slice structural continuity. 2) The analysis of existing methods is comprehensive and convincing. The paper provides a clear discussion of why conventional global vision–language alignment is insufficient for LGE report generation and explains the limitations of prior 3D medical report generation frameworks in a logical way. 3) The technical design and experimental validation are relatively thorough. The proposed framework is well structured, with explicit modeling of intra-slice, cross-slice, and textual structural information, and the paper includes both comparative experiments and ablation studies to support the effectiveness of each component. 4) The paper contributes a valuable multi-center dataset. Given the scarcity of paired LGE MRI-report data, the construction of a multi-center dataset is a meaningful practical contribution that can support future research in this area.
- 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 dataset description is somewhat limited. While the paper states that the dataset contains 2,038 cases from two clinical centers and uses one center for development and the other for external validation, it provides relatively limited detail about disease composition, report style variability, acquisition differences, and other characteristics that are important for assessing the difficulty and generalizability of the task. 2) The experimental analysis could be deeper. The paper includes ablation studies on the proposed graph modules, hyperparameters, and module order, but it lacks more detailed qualitative or error analysis showing exactly where the method improves clinical reporting, such as lesion localization, segment-level consistency, or cross-slice reasoning. Please consider adding some case studies. 3) The set of baselines could be further strengthened. Although the paper compares against several recent 3D medical vision-language models, it would be helpful to include more diverse or stronger baselines, especially methods with explicit structural modeling or more closely related report generation frameworks, to better position the contribution of the proposed method. 4) Could the authors discuss whether the proposed framework can be generalized to other volumetric MRI report generation tasks, such as brain MRI? In particular, while the global-local structure-aware alignment seems potentially general, some components appear strongly tailored to cardiac LGE, such as the short-axis continuity modeling and the AHA 17-segment textual graph. It would be helpful if the authors could clarify which parts of the framework are modality-agnostic and which parts would need to be redesigned for other organs.
- 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 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
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?
14.I gave this paper a weak reject because, although the proposed method is technically reasonable and the overall idea is promising, the experimental evidence is still not strong enough to fully support the claims. The paper presents encouraging results and some ablation studies, but the experimental section remains relatively limited for a paper making strong claims about clinically reliable LGE MRI report generation. In particular, I would like to see more thorough disclosure of experimental details, deeper case studies and error analysis to better understand where the method truly helps, and a broader validation of generalizability, for example through discussion of transferability or additional experiments on other public MRI datasets such as brain MRI. If the authors could provide richer experimental analysis, more transparent implementation and data details, and stronger evidence that the framework is not limited to this specific dataset and setting, I would be more inclined to support acceptance.
- 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.
After reading the response, I still have concern regarding the effectiveness of the proposed method. The author should include some more experimental analysis in the final version.
Review #3
- Please describe the contribution of the paper
This paper presents GLSA-RG, a Global–Local Structure-Aware Alignment framework for automated LGE MRI report generation. The core is a unified Structural Graph Representation that encodes intra slice regional interactions, cross slice continuity, and AHA 17 segment textual structural dependencies. A Structure-Aware Graph Alignment module performs fine grained multi instance contrastive alignment alongside global supervision. The authors also contribute a multi center LGE MRI dataset with paired reports. Experiments on two clinical datasets show the model outperforms state of the art approaches on standard NLG metrics and clinical accuracy, reliably capturing subtle enhancement patterns and structural consistency for clinically valid reports.
- 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 the following strengths. First, it clearly identifies the key limitation of existing CLIP-style global vision–language alignment for LGE MRI report generation: insufficient modeling of region-level scar patterns and cross-slice structural continuity, making the work clinically well-motivated. It proposes a novel Global–Local Structure-Aware Alignment framework (GLSA-RG) with unified structural graph representations, including intra-slice, cross-slice, and AHA 17-segment textual graphs, enabling fine-grained structure-aware visual-text alignment beyond coarse global matching. The authors build a multi-center LGE MRI dataset with paired reports to alleviate data scarcity. Extensive experiments and ablations show the model outperforms state-of-the-art methods on both internal and external validation, with improved clinical accuracy and structural consistency. The global–local alignment paradigm advances 3D medical vision-language modeling for structured cardiac report generation.
- 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 paper has the following strengths. First, it clearly identifies the key limitation of existing CLIP-style global vision–language alignment for LGE MRI report generation: insufficient modeling of region-level scar patterns and cross-slice structural continuity, making the work clinically well-motivated. It proposes a novel Global–Local Structure-Aware Alignment framework (GLSA-RG) with unified structural graph representations, including intra-slice, cross-slice, and AHA 17-segment textual graphs, enabling fine-grained structure-aware visual-text alignment beyond coarse global matching. The authors build a multi-center LGE MRI dataset with paired reports to alleviate data scarcity. Extensive experiments and ablations show the model outperforms state-of-the-art methods on both internal and external validation, with improved clinical accuracy and structural consistency. The global–local alignment paradigm advances 3D medical vision-language modeling for structured cardiac report generation.
- 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?
No
- 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?
This paper offers a strong technical contribution with real potential for clinical translation in automated LGE MRI report generation. The proposed GLSA-RG framework, with its global-local structure-aware alignment and unified structural graph representation, addresses a critical, often-overlooked limitation of existing CLIP-style vision-language methods: the failure to model fine-grained regional scar patterns, cross-slice continuity, and AHA 17-segment anatomical dependencies. This paradigm is generalizable and could benefit other 3D medical vision-language and structured reporting tasks, and the authors’ multi-center paired dataset is a valuable contribution to the field. However, there are several key issues that require clarification, including insufficient methodological details for reproducibility, incomplete clinical validation with missing segment-level performance metrics, lack of interpretability visualizations, and unaddressed real-world clinical applicability.
- 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 thank the reviewers for their valuable feedback and address their concerns below.
AC, R1W1: Novelty (1) Task novelty: We address automated LGE MRI report generation, a clinically important, challenging, and underexplored task, as acknowledged by R1, R3, and R4.(2) Framework novelty: Although some components draw on existing ideas, our main contribution is a unified, task-specific framework for cardiac MR and structured report generation that does not rely on segmentation. R4 recognizes it as a “novel global-local structure-aware alignment framework.” (3) Component novelty: We introduce intra-slice and cross-slice graph modules to model regional interactions and cross-slice continuity, enabling precise characterization of subtle myocardial enhancement patterns.
AC, R1W2: Limited human assessment, error analysis, and ethics Regarding Point 1, two cardiologists with over 10 years of clinical experience were invited to evaluate existing results. They independently reviewed 20 randomly selected reports using a 5-point Likert scale to assess enhancement detection accuracy and report usability. Their analyses were consistent with existing results and further supported our superiority over the baseline Med3DVLM. These analyses will be included. Regarding Point 2, Table 1 already characterizes model errors and demonstrates consistent improvements over prior state-of-the-art methods. We will further include additional qualitative error analysis on existing results. Regarding Point 3, All data were properly anonymized in accordance with institutional and data governance guidelines.
R3W1: Dataset description The development cohort includes normal, myocardial infarction, and DCM cases, while the external cohort primarily comprises DCM cases. Variations in report terminology, granularity, and imaging systems (Siemens vs. Philips) introduce substantial heterogeneity, increasing task difficulty and enabling a more rigorous evaluation of robustness and generalizability.
R3W2: Deeper experimental analysis including case studies and error analysis The manuscript already included a representative qualitative case study (Fig. 3) to support interpretability. Our method identified subtler enhancement patterns (e.g., apical inferior wall enhancement), produced fewer localization errors, better captured continuous inferior wall enhancement across adjacent slices, and reduced anatomically inconsistent segment assignments.
Existing experiments in Table 1 characterize model errors and demonstrate consistent improvements over prior state-of-the-art methods. We will include additional qualitative error analysis on existing results.
R3W3: Stronger baselines Our experiments already include representative state-of-the-art report generation methods under identical settings (Table 1). Some additional structure-aware approaches were excluded due to differing settings, limiting fair comparison. We will add citations and discussion for a more comprehensive overview.
R3W4: Generalization ability The proposed global–local structure alignment paradigm is modality-agnostic. The short-axis continuity modeling and AHA 17-segment graph are cardiac-specific anatomical priors and would need to be replaced with organ-specific structural priors when extending our method to other anatomies.
R4W1: Reproducibility We will open-source our code to provide full details for reproducibility.
R4W2: Missing segment-level metrics Our task focuses on report generation, so conventional segment-level metrics are less applicable. Our case analysis already covers lesion location and extent accuracy. We will explore finer-grained localization metrics in future work.
R4W3: Lack of interpretability visualizations See our reply to R3-W2. R4W4: Clinical applicability While our work focuses on methodological advancement, its promising results, together with the unanimous recognition of its clinical significance by all reviewers, suggest its potential for clinical applicability.
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.
Automated LGE MRI reporting is a challenging clinical task and an interesting technical problem. Weaknesses are the incremental methodological novelty, limited dataset characterization, and reliance on automatic or GPT-based metrics. In the rebuttal, the authors should strengthen the evidence for clinical reporting quality, and better position the contribution in currently existing report-generation methods.
- 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 rebuttal addressed the main concerns: the expert clinical evaluation by two cardiologists provides human-assessed evidence beyond automatic metrics, the dataset heterogeneity is clarified, and the task-specific design choices are justified. R1 and R4 did not change the rating and R3 moved to accept. Overall the paper meet the threshold of acceptance.
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.
After considering the reviews and rebuttal discussion, I recommend acceptance. The paper addresses an important and relatively underexplored problem in automated LGE MRI report generation, with clear clinical motivation and meaningful technical contributions.
The proposed GLSA-RG framework is well designed and introduces a coherent global–local structure-aware alignment strategy that explicitly models intra-slice regional interactions, cross-slice continuity, and structured myocardial segment relationships. Compared with prior global-only vision–language alignment approaches, the proposed framework provides a more clinically grounded representation for capturing subtle enhancement patterns and structural consistency in LGE MRI.
The experimental evaluation is reasonably comprehensive, including multi-center paired MRI-report data, external validation, comparisons with recent baselines, and multiple ablation studies. The reported improvements are consistent across both standard report-generation metrics and clinical evaluation metrics, supporting the effectiveness of the proposed alignment and graph modeling strategies.
While some concerns remain regarding the depth of human clinical evaluation, dataset description, and reproducibility details, the rebuttal adequately addressed the main reviewer concerns and clarified the intended scope and contribution of the work. Overall, I believe the paper provides a valuable and practically relevant contribution to 3D medical vision-language modeling and automated structured report generation, and is suitable for acceptance at MICCAI.
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.
All reviewers highlight the contributions of the paper and suggest to accept the submission.