List of Papers Browse by Subject Areas Author List
Abstract
Grounding radiology report descriptions to 3D CT volumes is essential for verifiable clinical interpretation, yet remains challenging due to the semantic-spatial gap between free-text narratives and volumetric anatomy. Existing report-assisted and vision-language grounding methods typically rely on phrase-level alignment or dense pixel supervision, resulting in limited lesion-wise correspondence and suboptimal localization accuracy. We propose GLeVE, a graph-guided lesion grounding framework with anatomical prior verification and octree-based autoregressive refinement. GLeVE treats each lesion description as an atomic semantic unit and encodes organ attribution, attributes, and inter-lesion relations through relation-aware graph reasoning to produce discriminative lesion-wise queries. Anatomy-aware proposal generation with region-level verification enforces one-to-one text-lesion alignment, while hierarchical octree refinement progressively improves boundary delineation. Experiments on AbdomenAtlas 3.0 demonstrate consistent gains over classical multimodal foundation models and report-supervised baselines in both segmentation accuracy and lesion-level localization.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2407_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/JSLiam94/GLeVE
Link to the Dataset(s)
https://huggingface.co/datasets/AbdomenAtlas/AbdomenAtlas3.0Mini
BibTex
@InProceedings{JiaShu_GLeVE_MICCAI2026,
author = { Jiang, Shuo AND Hong, Yuhao AND Jiang, Chunbo AND Chen, Weihong AND Chen, Huangwei AND Zhu, Shenghao AND Wu, Beining AND Liu, Mingxuan AND Zhu, Zhu AND Qin, Feiwei AND Tan, Min AND Chen, Yifei},
title = { { GLeVE: Graph-Guided Lesion Grounding with Proposal Verification in 3D CT } },
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
The paper proposes GLeVE, a framework to ground radiology reports to 3D CT at the lesion level rather than phrase or global alignment. The key idea is to treat each lesion description as a complete semantic unit and encode its attributes and relations using a graph, which is then used to generate lesion-specific queries. The method combines graph-based semantic modeling, anatomy-aware proposal generation with verification to enforce one-to-one text–lesion alignment, and an octree-based refinement to improve segmentation boundaries. It also supports weak supervision by leveraging report-derived attributes when masks are limited. Overall, the contribution is a structured approach that improves lesion-level localization and alignment between report descriptions and imaging evidence.
- 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 addresses an important and well-defined problem of lesion-level grounding, which is more clinically meaningful than phrase-level alignment. The idea of modeling each lesion as a structured semantic unit using graph reasoning is intuitive and useful. The method is well designed with a clear pipeline and shows consistent improvements over strong baselines, including in low-supervision settings. Experiments are solid and support the contributions.
- 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 overall novelty is somewhat limited, as the method mainly combines existing components (graph reasoning, proposal verification, coarse-to-fine refinement) rather than introducing a fundamentally new formulation. 2.The pipeline is quite complex and involves multiple modules (LLM parsing, graph transformer, proposal generation, verification, octree refinement), which may limit scalability and real-world deployment. 3.The region-level verification strategy is not fully justified; it is unclear why verifying candidates at the region level is optimal compared to voxel-level or hybrid strategies, and no theoretical or empirical comparison is provided. 4.The method relies on several hyperparameters (e. g. , λ weights in the loss, temperature τ), but there is limited discussion on how these are selected or how sensitive the model is to these choices. 5.The optimization strategy and training stability are not well analyzed, especially given the combination of weak and strong supervision losses, which could affect convergence behavior. 6.The approach depends on accurate report parsing and structured extraction, making it potentially sensitive to noise or variability in real clinical reports. 7.Evaluation is limited to a single dataset, and the generalization of the method to other datasets, organs, or imaging modalities is not demonstrated.
- 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
If some of the concerns raised above (particularly regarding design choices, hyperparameter sensitivity, and generalization) can be clarified or addressed during the rebuttal, it would further strengthen the paper.
- 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 paper is overall strong due to addressing an important problem of lesion-level grounding between radiology reports and 3D CT, and proposing a well-structured framework that combines semantic modeling, anatomical priors, and hierarchical refinement. The experimental results are solid and show consistent improvements over strong baselines, including under limited supervision, which highlights the practical value of the approach.
However, the overall novelty is somewhat limited as the method mainly integrates existing ideas rather than introducing a fundamentally new formulation. In addition, the pipeline is relatively complex and relies on multiple components, which may affect scalability and real-world applicability. There is also limited analysis on key design choices such as the region-level verification strategy, sensitivity to hyperparameters, and optimization behavior. Finally, evaluation is restricted to a single dataset, leaving questions about generalization.
Considering both the strengths and weaknesses, I find the paper to be slightly above the acceptance threshold.
- 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 GLeVE, which is a model/framework that grounds lesions from CT reports using combines anatomy-conditioned embeddings and prompted with query sets generated from graph transformer embeddings. The paper additionally introduces an Octree refinement module to refine the proposed segmentation candidates.
- 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 tackles a very important problem related to grounding findings from radiology reports. The ideas proposed, including Lesion Semantic Modeling and Querying, Anatomy-Prior Lesion Proposal and Verification and octree refinement of predicted masks are relatively novel and interesting
- The results are strong and reporting of results is comprehensive
- 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’s result are limited to one dataset that only contains tumor leions. Other datasets like ReXGroundingCT (https://arxiv.org/abs/2507.22030) which covers more diverse lesion types could be used. Moreover, I think mixing grounding of organs and grounding of findings in the same evaluation introduces inflated results. Grounding of organs is a much easier task, and arguably one in which grounding is not useful, because the set of possible organs is very small. Gounding findings is much harder as findings can in multiple different places on the scan, and have varying characteristics (size, shape, etc)
- The paper took me a lot of time to read mainly because of how it is written, especially the choice of mathematical symbols. I think a lot of the equation and symbols should be re-written to be more consistent across the paper and with literature. Many variables (like ϕ in equation 2 or Ψ in the QueryBank or C in R^C) are not explained/defined. Even if something seems obvious, please try to define all variables. Also, why is ψ used for MLP? These are non-standard definitions that create confusion.
- Please rate the clarity and organization of this paper
Poor
- 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 ideas presented are novel and interesting, and the paper is tackling a very important task.
- 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 #3
- Please describe the contribution of the paper
The authors present GleVe, a novel framework for grounding radiology reports that targets pixel-level lesion localization adding a one-to-one alignment with key report descriptions while also adding relevant patterns from imaging. This presents a tool with high potential for clinical practice by delivering an interpretative and quantitative methodology that has improved results compared to state-of-the art appraoches. The authors have leveraged the information provided in radiology reports with frameworks for organ segmentation using medical imaging data.
- 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 authors present an interesting combination of frameworks to leverage each for a optimized grounding radiology reporting. They use a mix of composite semantics, organ segmentation and an octree-based autoregressive refinement, which provides an increased performance compared to state-of-the-art appraoches. These methods collectively enhance model understanding and precision in complex tasks – which can be beneficial for clinical reporting. For instance, graph-structured representation learning captures the information from descriptions in reports, allowing the model to reason about interactions between observed patterns in the scan. Additionally, they interestingly leveraged geometry-aware anatomical embeddings, informed by priors, to provide domain knowledge crucial in clinical settings for optimal and localized diagnosis. The multi-candidate verification and autoregressive refinement strategies also provide this method with a built-in quality-check-like mechanism and a key-to-key alignment between reports and imaging.
- 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.
Limitations of the present work should be further discussed. For instance, graphs have been known to misunderstand structure whilst using priors can bias the results by over-assumptions on the data. Furthermore, refinement can further exarcebate the mistake-to-success ratio by polishing the mistakes dragged from previous steps instead of correcting them. It should be discussed whether the octree approach for small lesions could yield wrong results if the data is noisy or when reconstruction kernels are diverse – which is very frequent in clinical imaging. Latency is not discussed and raises the question about the computational cost of the present architecture. For clinical feasibility this could be a major concern.
- 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
Figure 1 is very hard to read due to the font size choice. Would strongly recommend increasing by at least 4 pt. Figure 2 could also benefit from a font size increase. The README.md file on the repository could benefit from deeper explanation on how to use the code and reproduce the results.
- 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 paper presents a highly relevant methodology for grounding radiology reports that could be exploited for diagnosis-aid in clinical settings specifically due to the lesion-specific modelling appraoch and one-to-one key matching. In addition, the performance is above state-of-the-art, the methodology is novel and an interesting combination of multiple existing frameworks, and the description is thorough enough for replication studies. I have minor concerns about the computational cost of the architecture, but the overall quality of the paper is good and interesting for both the technical and medical community.
- 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.
All reviewers agreed that the paper addressed a meaningful problem and the method was well designed