List of Papers Browse by Subject Areas Author List
Abstract
Automatic voxel-level grounding of lung lesions from free-text reports in 3D chest Computed Tomography (CT) is critical for clinical interpretability.
However, this task remains highly challenging due to the intricate spatial complexity of large 3D volumes and the heterogeneity of free-text findings.
Existing end-to-end approaches often struggle to simultaneously learn the localized feature representations required for accurate 3D segmentation and the complex semantic understanding needed for text alignment, leading to suboptimal grounding performance. To overcome this fundamental limitation, we propose a novel decoupled framework that disentangles the problem into two specialized stages: (1) class-agnostic lesion segmentation and (2) text-volume reasoning.
This structural separation allows the model to first extract candidate sub-volumes by localizing potential abnormalities.
Subsequently, intensive cross-modal reasoning is performed to align these localized sub-volumes with free-text medical findings.
To resolve the spatial ambiguities inherent in local regions, the reasoning module is augmented with explicit anatomical guidance, utilizing relative spatial coordinates and lung lobe priors. Evaluated on the ReXGroundingCT benchmark, our method achieves state-of-the-art performance in instance-level metrics on the official leaderboard.
These results demonstrate that decoupling detection from reasoning is a highly effective paradigm for handling the complexity of 3D medical visual grounding. Our code is publicly available at https://github.com/khuhm/DAGG.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3331_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
Link to the Dataset(s)
N/A
BibTex
@InProceedings{UhmKwa_Decouple_MICCAI2026,
author = { Uhm, Kwang-Hyun AND Son, Inhwa AND Ko, Sung-Jea},
title = { { Decouple and Reason: Anatomically Guided Two-Stage Grounding of Lung Lesions in 3D Chest CT from Free-Text Reports } },
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 medical grounding task by proposing a two-stage framework, consisting of class-agnostic lesion segmentation and lesion-text alignment. Specifically, the first stage identifies potential lesion regions, while the second stage leverages structural priors of lesions to establish fine-grained correspondence between lesion regions and report text.
- 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 key strength of this work lies in decoupling two closely related tasks: precise lesion localization and text alignment. By adopting class-agnostic lesion estimation, the method reduces the difficulty of localizing diverse lesion types. Furthermore, by incorporating anatomical location information into class-agnostic lesion representations, it effectively narrows the semantic gap between lesion features and textual descriptions. In addition, the methodology is clearly presented and reproducible, and the experimental analysis is thorough and well-reasoned, providing solid support for the paper’s claims.
- 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 method lacks a unified and clearly defined overall loss formulation. Moreover, the experimental section does not report model complexity, such as parameter size and computational cost.
- 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 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.
(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 core idea of this work demonstrates good novelty and soundness. The method is clearly designed, and the experimental results provide sufficient evidence to validate its effectiveness.
- 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 DAGG, a decoupled two-stage framework for voxel-level grounding of free-text radiology findings in 3D chest CT. The core idea is to separate lesion localization from semantic alignment:
-
Stage 1: Class-agnostic lesion segmentation using a 3D semantic nnU-Net to generate candidate sub-volumes via connected component analysis
-
Stage 2: Cross-modal text-volume reasoning, where candidate sub-volumes are matched to free-text findings using a contrastive-style alignment loss
The key methodological novelty lies in decoupling semantic segmentation and text-prompt matching. Further, (1) the Independent Pairwise Alignment strategy cleverly handles false-positive candidates from Stage 1 and (2) the Explicit Anatomical Guidance mechanism that augments local patch features with normalized centroid coordinates and lung lobe distribution priors from TotalSegmentator for better embeddings. Evaluated on ReXGroundingCT, the method reports state-of-the-art performance at submission time.
-
- 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.
-
Evaluation is solid and consistent: Metrics (Dice, F1, HIT10%, Precision, Recall) align with the RexGroundingCT benchmark setup; no cherry-picking of metrics
-
Ablation confirms each design choice contributes incrementally and meaningfully
-
The Independent Pairwise Alignment is genuinely clever: Handling N text prompts against K image embeddings (where K can include false positives with no GT text) is a non-trivial problem, and the cosine embedding loss formulation is a practical, principled solution
-
SOTA on RexGroundingCT leaderboard at submission time. The reported results are competitive in a task where overall performance across the field is still very low (best current dice is ~0.28 on the leaderboard)
-
Code is provided (anonymous link): big plus for reproducibility in a field where this is rare
-
- 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.
-
Unfair baseline comparison for VoxTell: The authors state “For fair comparison, all baselines were fine-tuned on the training split.” However, VoxTell’s own paper reports higher numbers on the RexGroundingCT validation split than what DAGG achieves. In the paper they report that the model was also specifically fine-tuned for RexGroundingCT similar to what the authors did. Also, it is currently first on the official leaderboard. I acknowledge that this leaderboard entry was submitted after the MICCAI deadline and could not have been considered, but the discrepancy between the fine-tuned VoxTell numbers in this paper vs. those in the VoxTell paper itself is not explained.
-
Scope is severely limited but not clearly stated: The title says “3D Chest CT grounding” but the method only works for lesions/abnormalities. Normal anatomical structures (airways, lymph nodes, heart, mediastinum, etc.) are completely out of scope because Stage 1 only produces a class-agnostic lesion mask. This is a dataset limitation, yes, but the paper should be upfront about it rather than implying generality.
-
Hard dependency on Stage 1 creates a structural false-negative problem: The paper discusses how the system handles false positives from Stage 1; but what happens with false negatives? If Stage 1 fails to segment a lesion (OOD pathology, subtle findings, or anything the nnU-Net didn’t see during training), Stage 2 has zero chance of matching that text prompt to anything. Is there a fallback? Some end-to-end VLM approaches avoid this by projecting the text prompt directly into a latent space that guides the segmentor, which at least gives the model a chance. This architectural limitation is not discussed.
-
Similarly, TotalSegmentator-based anatomical guidance is limited to lung structures: If a finding is outside the lung (pleural effusion, mediastinal mass, pericardial involvement), the lobe distribution vector and potentially the coordinate normalization break down. The paper doesn’t address this.
-
Single-dataset evaluation only: The study is entirely conducted on RexGroundingCT, with no OOD testing whatsoever. Given that overall performance in this field is still extremely low (best Dice ~0.28), it’s hard to draw any strong conclusions about real-world applicability. At minimum a qualitative discussion of failure modes would help.
-
(Minor) Missing relevant references: For a paper that is centrally about image-text alignment and contrastive learning, CLIP is not cited. InfoNCE is cited but the broader image-text alignment literature could use at least one or two more references.
-
- 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
In the introduction, the sentence “advancing this from coarse bounding boxes to precise voxel-level segmentation masks is of paramount importance for developing reliable Computer-Aided Diagnosis (CAD) systems “ cites a method which doesn’t work with bounding boxes and has no stated connection to CAD systems. This citation is incorrect and a bit misleading.
- 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?
The core idea of decoupling lesion localization from text-volume reasoning is reasonable and the independent pairwise alignment is a genuine contribution. The ablation study is clean and the code is provided. However, several concerns push me below the acceptance threshold:
The VoxTell comparison looks like it may understate the actual best-known performance on this benchmark, even if the top leaderboard entry came after the deadline, the VoxTell paper’s own numbers are already higher than DAGG’s reported results. Even if the proposed DAGG system may not end up as the best performing model, a clear and fair analysis of when which model is better and what DAGG brings to the table that strong competing methods lack might strengthen the contribution post-rebuttal.
The architectural hard dependency on Stage 1 is a significant limitation that is not acknowledged, i.e. false negatives from Stage 1 are completely unrecoverable.
The scope is much narrower than the title and framing suggest (only lung lesions, no other structures), which inflates the perceived generality.
Evaluation on a single dataset with no OOD analysis in a field where absolute performance is still very low limits the strength of the conclusions.
I would reconsider upon a strong rebuttal that addresses the VoxTell comparison discrepancy and the false-negative problem. I would appreciate the authors making a clearer, exhaustive and fair comparison to competing methods and acknowledge their own limitations in order to be a meaningful submission to the MICCAI conference.
- Reviewer confidence
Very confident (4)
- [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 reviewing the rebuttal, I am borderline but lean toward acceptance. The decoupled two-stage formulation, the Independent Pairwise Alignment strategy, and the clean ablation study constitute a genuine methodological contribution to an emerging and clinically relevant problem, and the authors adequately clarified the lobe vector interpretation, the extra-pulmonary handling, and committed to revising the title/scope, adding the unified loss formulation, model complexity, and missing references.
Several concerns remain unresolved and should be reflected in the final version. The VoxTell comparison is explained but not fully resolved: VoxTell still leads the official leaderboard in global Dice and HIT@10%, so the authors should not claim overall state-of-the-art and should instead frame DAGG’s strength as instance-level metrics (Precision, F1). The architectural hard dependency on Stage 1, where false negatives are unrecoverable, is a real design limitation that the rebuttal sidesteps by citing ~83% recall rather than acknowledging it, and this needs explicit discussion in the final manuscript. The unified loss formulation is currently missing from the paper and its inclusion is necessary, not optional, for reproducibility. Single-dataset evaluation remains a limitation, though understandable given ReXGroundingCT is the only benchmark of its kind. Given the genuine novelty, reproducibility, and clinical relevance, the paper warrants acceptance conditional on the authors tempering SOTA claims, explicitly stating the false-negative limitation, and including the promised loss formulation and scope clarification.
Review #3
- Please describe the contribution of the paper
The work describes a novel framework for voxel-level grounding of free-text findings in associated 3D medical images. In terms of contribution, the authors propose a two-stage approach, in which lesions/ anomalies are first detected and then linked to text findings. After the detection phase, regions of interest are further processed, and linked independently to text findings. This is improved through anatomical guidance, which aids the localisation of the visual finding region with respect to the global lung anatomy. The method is applied on an open CT data set and achieves competitive performance wrt the currently listed alternative approaches.
- 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 work addresses a highly novel and emerging application, i.e. the visual grounding of free-text findings. The formulation as a two-stage approach is well argued, original and intuitively makes sense. The experiments performed allow to assess the value of the contribution, of each component of the contribution and with respect to limited previous work in the field.
- 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.
It is unclear how the Independent Pairwise Alignment between visual candidate lesions and textual findings is optimal with respect to alternative approaches. Specifically, the final selection of a match, is based on a single threshold criterion of the multi-modal similarity between visual and textual findings, irrespective of whether textual findings have other visual matches. While this aligns with the example in which one textual finding (several nodules) aligns with multiple visual regions, it seems the approach could also lead to false positive matching, and is sensitive to tuning of the threshold (tau). More advanced textual reasoning (e.g. are we looking for a unique visual finding or multiple regions) seems appropriate.
The paper lacks details on how training and hyperparameter tuning was performed, understandable because of the short format. The authors do list an anonymised link to source code, but some key aspects should be considered for addition to the paper. For example: the training and inference of the auxiliary segmentation network, the tuning of m and tau, the fine-tuning of the baselines.
While scientifically interesting, the need and motivation for visual grounding could be better argued. Are there direct clinical applications? Is the goal to maximally exploit radiology reports and associated images for training segmentation networks?
- 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
The chosen representation of the anatomical guidance could be clarified, specifically for the second component. Is the vector containing the fraction of volume of the candidate region that falls within each of the lobes?
- 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 authors propose a novel approach to an emerging application. The method is sound and well explained overall. The experiments demonstrate a good performance, competitive to other attempts.
- 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
Author Feedback
We sincerely thank all reviewers for acknowledging our core idea of decoupled localization/reasoning, our independent pairwise alignment, and the clear ablation study. We address the raised primary concerns below.
1.VoxTell Comparison Discrepancy & DAGG’s Advantages [R2]: To clarify the VoxTell performance discrepancy: Their paper reports only global Dice and Hit@5%, lacking other official benchmark metrics (Hit@10%, Precision, Recall, F1). Furthermore, their repository lacks the fine-tuning code and the ReXGroundingCT-specific weights. To enable comprehensive multi-metric comparison and detailed analysis, we manually fine-tuned their generic model based on limited details available in their paper, inevitably causing the discrepancy. We will clearly mark and explain both their originally reported numbers and our reproduced metrics. Crucially, while VoxTell leads the leaderboard in global area overlap (Dice, Hit@10%), DAGG ranks 1st in all instance-level metrics (Precision, Recall, F1), outperforming VoxTell significantly (e.g., ~10% higher in F1). Thus, DAGG’s approach offers a distinct advantage when precise, instance-by-instance reasoning is prioritized over total global area overlap.
2.Hard Dependency on Stage 1 & Pairwise Alignment Rationale [R2, R3]: We acknowledge the structural limitation that Stage 1 false negatives cannot be recovered, unlike end-to-end VLMs. However, Stage 1 is designed to maximize recall (achieving ~83% during development), functioning as a highly sensitive region proposal network. This shifts the challenge to accurate text-image alignment. DAGG’s 1st place ranking in instance Precision and Recall demonstrates that Stage 2 effectively resolves this challenge, making our framework highly viable. Regarding Independent Pairwise Alignment: this is essential to filter out Stage 1 false positives. Additionally, clinical findings often describe multiple instances (e.g., “several nodules”). Evaluating candidates independently allows one text to validly match multiple visual regions, reflecting clinical reality. We will expand on this multi-modal reasoning.
3.Scope Clarification & Extra-Pulmonary Anatomical Guidance [R2, R3]: We will revise the title to “3D Chest CT Lung Lesion Grounding” and clarify the scope. Specifically, the lobe vector represents the candidate’s volume fraction within each lobe. For findings outside the lung (e.g., pleural effusion), spatial guidance remains robust. The lobe vector becomes all-zeros, but the relative coordinates (normalized to the lung bounding box) remain valid. The combination of an all-zero lobe vector and external relative coordinates specifies not only that a finding is outside the lung parenchyma, but also its precise directional location (e.g., inferior-right) relative to the lungs.
4.Clinical Applicability & Single Dataset Evaluation [R2, R3]: ReXGroundingCT is currently the only available dataset for 3D CT free-text lesion grounding, making OOD testing infeasible. However, DAGG’s framework (using structure vectors/relative coordinates) has strong potential to generalize to whole-chest parsing. Given the currently low absolute performance across the field, DAGG is motivated as a “radiologist-in-the-loop” application. Visual grounding directly assists in visualizing reports for patient explanations, resident education, and facilitating communication with referring physicians, provided a radiologist verifies the outputs.
5.Additional Clarifications & Revisions [R1, R2, R3]: To address the remaining constructive feedback, the revised manuscript and code README will include: A unified and clearly defined overall loss formulation. Model complexity metrics, including parameter size and computational cost (FLOPs). Detailed training and hyperparameter tuning configurations (including the auxiliary segmentation network, fine-tuning baselines, and threshold tau selection). Expanded literature review, including CLIP and broader image-text alignment references.
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.
This paper received three good quality reviews with two accept and one reject rating. The final score will be depended on the rebuttal quality and details.
The paper proposes a two-stage 3D chest CT visual grounding method that first performs class-agnostic lesion localization and then aligns candidate lesion embeddings with free-text findings using anatomical guidance and independent pairwise alignment, addressing a novel and clinically relevant image-text grounding problem. Reviewers praised the decoupling of localization and text alignment, the clever handling of multiple candidate lesions and prompts, strong ablations, reproducibility/code availability, and competitive RexGroundingCT results, but raised concerns about single-dataset evaluation, unclear loss/training details, and whether the title overstates the scope because the method is limited to lesions rather than general anatomical structures.
- 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 rebuttal, all three reviewers agree to accept this submission. 1) The technical novelty is quite good and practical. “The paper proposes DAGG, a decoupled two-stage framework for voxel-level grounding of free-text radiology findings in 3D chest CT. The core idea is to separate lesion localization from semantic alignment:
-
Stage 1: Class-agnostic lesion segmentation using a 3D semantic nnU-Net to generate candidate sub-volumes via connected component analysis
-
Stage 2: Cross-modal text-volume reasoning, where candidate sub-volumes are matched to free-text findings using a contrastive-style alignment loss
The key methodological novelty lies in decoupling semantic segmentation and text-prompt matching. Further, (1) the Independent Pairwise Alignment strategy cleverly handles false-positive candidates from Stage 1 and (2) the Explicit Anatomical Guidance mechanism that augments local patch features with normalized centroid coordinates and lung lobe distribution priors from TotalSegmentator for better embeddings. Evaluated on ReXGroundingCT, the method reports state-of-the-art performance at submission time.”
2) The experimental evaluation is fair and still not fully complete, especially considering VoxTell.
“Several concerns remain unresolved and should be reflected in the final version. The VoxTell comparison is explained but not fully resolved: VoxTell still leads the official leaderboard in global Dice and HIT@10%, so the authors should not claim overall state-of-the-art and should instead frame DAGG’s strength as instance-level metrics (Precision, F1). The architectural hard dependency on Stage 1, where false negatives are unrecoverable, is a real design limitation that the rebuttal sidesteps by citing ~83% recall rather than acknowledging it, and this needs explicit discussion in the final manuscript. The unified loss formulation is currently missing from the paper and its inclusion is necessary, not optional, for reproducibility. Single-dataset evaluation remains a limitation, though understandable given ReXGroundingCT is the only benchmark of its kind. Given the genuine novelty, reproducibility, and clinical relevance, the paper warrants acceptance conditional on the authors tempering SOTA claims, explicitly stating the false-negative limitation, and including the promised loss formulation and scope clarification.”
-
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 rebuttal, all reviewers recommend acceptance of this work, considering the novel methodological contribution (separate lesion localization from semantic alignment) to an emerging and clinically relevant problem. Despite of some of R2’s valid concerns (unrecoverable false negatives, single dataset evaluation, and VoxTell comparison), this work should be accepted and represented to the MICCAI community.
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.
Three reviewers agreed on acceptance after rebuttal. The paper represents a methodological contribution to an emerging and clinically relevant problem of visual grounding, achieving competitive RexGroundingCT results. Although some issues remain unsolved as indicated by R2, it is an interesting paper to the MICCAI community.
Meta-review #4
- 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 of the 3 reviewers agree to accept this paper.