List of Papers Browse by Subject Areas Author List
Abstract
We study how to train visually grounded vision–language models (VLMs) for radiology without manual spatial annotations. We introduce RefRad2D, a large-scale bilingual (German/English) dataset of 1.2M CT and MR image–text pairs derived from clinical practice, with task-specific VQA and spatial grounding subsets generated automatically via LLM-based curation and automated segmentation. Trained on this data, our model RadGrounder jointly performs report generation, visual question answering, and spatial grounding via bounding-box detection or segmentation. On external VQA benchmarks (Slake, VQA-RAD), RadGrounder achieves competitive results with specialized medical VLMs. Adding our clinical data to the training mixture improves open-ended VQA over fine-tuning on the downstream datasets alone, showing the transferability of our dataset. Crucially, adding grounding supervision does not degrade language quality, enabling spatially verifiable outputs at no cost to VQA performance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4764_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/lmb-freiburg/radgrounder
Link to the Dataset(s)
N/A
BibTex
@InProceedings{SalYus_Scalable_MICCAI2026,
author = { Salcan, Yusuf AND Ging, Simon AND Schirrmeister, Robin Tibor AND Arnold, Philipp AND Kotter, Elmar AND Bozorgtabar, Behzad AND Brox, Thomas},
title = { { Scalable Training of Spatially Grounded 2D Vision–Language Models for Radiology } },
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 introduces a framework for training spatially grounded 2D vision-language models for radiology without requiring manual spatial annotations. The key contribution is the construction of RefRad2D, a large-scale bilingual (German/English) dataset comprising approximately 1.2 million CT and MR image-text pairs derived from clinical practice. The dataset includes automatically generated spatial grounding annotations obtained through a pipeline combining LLM-based report parsing and segmentation masks from TotalSegmentator. The authors have proposed RadGrounder, a multi-task VLM based on a PaliGemma-style architecture that jointly performs report generation, visual question answering (VQA), and spatial grounding. Spatial grounding is primarily formulated as a text-generation task via tokenized bounding box prediction, with an optional auxiliary segmentation head.
- 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 introduction of RefRad2D is a significant contribution. The authors also present a weakly supervised strategy that removes the need for manual annotations by leveraging pretrained models and report parsing. Additionally, the paper provides comprehensive experiments across internal and external benchmarks, along with ablations that analyze grounding strategies, training configurations, and the impact of grounding on language quality.
- 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 model architecture is largely based on existing components rather than introducing fundamentally new modeling ideas. Pix2Seq, LLaVA have already explored similar directions. The paper lacks the analysis of failure cases(misalignment between text and segmentation etc). The paper also lacks comparisons with more recent or specialized medical VLMs and grounding models.
- 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?
Dataset use LLMs to generated annotations this should be clearly stated as okay for research but a careful selection for any practice.
- 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 paper introduced a novel and large-scale RefRad2D dataset and RadGrounder, a multi-task VLM. The author has also performed ablations thoroughly. However, the main contribution of science is the dataset; releasing the dataset publicly is expected.
- 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
Review #2
- Please describe the contribution of the paper
This paper presents RefRad2D, a large-scale bilingual CT/MR slice-level dataset with automatically generated VQA and spatial grounding annotations, and RadGrounder, a multi-task radiology VLM that jointly performs report generation, VQA, and spatial grounding. The work is motivated by the need for spatially verifiable radiology outputs without relying on manual spatial annotations. The experiments show competitive external VQA performance and suggest that adding grounding supervision does not harm language quality.
- 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 problem: improving the verifiability of radiology VLMs through spatial grounding. A major strength is the scale and clinical origin of the proposed dataset, which extends grounded VLM training beyond synthetic or narrowly scoped medical datasets. The multi-task framework is well designed, the ablation studies are informative, and the finding that token-based detection grounding can be added without degrading VQA/report quality is practically useful.
- 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 main limitation is that the methodological novelty is moderate. The modeling components are largely built on existing architectures and grounding formulations, so the contribution is driven more by data curation and training design than by a fundamentally new VLM architecture. In addition, the current grounding targets are primarily anatomical structures derived from TotalSegmentator rather than pathology-level findings, which limits clinical relevance. The grounding evaluation also risks becoming somewhat circular, since both supervision and evaluation depend heavily on automatically generated labels. Finally, while LLMScore is interesting, its human validation remains limited and does not fully cover long-form report generation.
- 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.
(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?
I find the paper valuable because it tackles a highly relevant practical bottleneck and provides a scalable data-centric solution with encouraging empirical results. However, the contribution is stronger as a dataset-and-system paper than as a methodological advance, and the current validation is not yet fully sufficient to establish strong clinical grounding performance beyond anatomical localization.
- 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 rebuttal, I think the authors have reasonably clarified several of my concerns. In particular, they better position the work as a scalable data-and-training framework rather than a novel VLM architecture, and provide additional analyses on long-tailed anatomical classes, grounding errors, and the validation of LLMScore. My main concerns are not fully eliminated. The methodological novelty remains moderate, and the grounding supervision is still mainly based on anatomical structures generated by TotalSegmentator rather than physician-validated pathology-level findings. The lack of a standardized 2D CT/MR grounding benchmark is understandable, but it still limits the strength of direct grounding comparison. Overall, the rebuttal improves the clarity and transparency of the work. Although some limitations remain, I still find the dataset-and-system contribution practically valuable. Therefore, I maintain my original weak accept recommendation.
Review #3
- Please describe the contribution of the paper
The main contribution of the paper is the development of a scalable framework for training spatially grounded VLMs for radiology using unannotated clinical data.Rather than relying on expensive, manual expert labeling, the authors prove that high-performing medical AI can be built by mining existing clinical records through an automated pipeline.
1.The authors created a massive dataset that serves as a blueprint for scaling medical VLMs，named RefRad2D Dataset,about 1.2 million image–text pairs from a decade of clinical CT and MRI scans.
2.They bypassed manual labor by using TotalSegmentator to find anatomical structures and an LLM to link those structures to the text in radiology reports.
3.The paper introduces RadGrounder, a multi-task model based on PaliGemma 2 that bridges the gap between seeing and describing.It performs three tasks at once: generating reports, answering medical questions (VQA), and localizing findings. Unlike “black-box” models that just provide text, RadGrounder can output bounding boxes to show where it is looking.
4.The authors proposed two new ways to measure success:LLMScore: Using a powerful LLM (Gemma 3) as a judge to grade reports on clinical factuality. They proved this scores almost identically to human radiologists (r = 0.977).G-IoU (Grounding-IoU): A metric that ensures the model is not just pointing to the right box, but also identifying the correct anatomy inside that box.
- 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 major strengths of this paper lie in its pragmatic approach to data scarcity and its commitment to making medical VLM outputs verifiable through automated grounding. 1.The most significant strength is the RefRad2D pipeline, which converts unstructured clinical logs into a high-quality, grounded dataset without manual intervention. This allows the authors to leverage a decade of “raw” hospital data (1.2M pairs), proving that 2D slice-level supervision is a scalable and viable alternative to computationally expensive 3D volumetric training. 2.The paper shows that adding spatial grounding tasks (detection or segmentation) does not degrade language performance. RadGrounder maintains high VQA and report quality while adding the ability to localize findings. This demonstrates that spatial interpretability can be integrated into VLMs essentially in terms of performance trade-offs.
- The paper doesn’t just succeed on its own internal data; RadGrounder achieves state-of-the-art or competitive results on Slake and VQA-RAD. This is a strong testament to the transferability and quality of the RefRad2D dataset.
- 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.
There are several weaknesses regarding its novelty, the limitations of its grounding approach, and the evaluation depth.
1.Although the paper claims that spatial grounding capabilities can mitigate hallucinations, the experimental design lacks a quantitative analysis of the correlation between localization errors and textual hallucinations. Is the model still reliable if it localizes the correct region but provides an incorrect description (or vice versa)? The analysis of these complex failure modes is insufficient, and the authors fail to fully leverage their proposed G-IoU metric to perform a deeper diagnostic evaluation of when and why the model’s visual and textual outputs diverge.
2.While the authors define G-IoU to mathematically penalize diagnostic omissions (false negatives) by normalizing with the maximum number of entities, the paper lacks a dedicated analysis of these omissions. There is no breakdown of which anatomical structures are most frequently missed, nor is there a qualitative discussion on the clinical implications of the model’s false negatives compared to its hallucinations (false positives).
3.The reported grounding performance (G-IoU) is presented only as a macro-average across 121 anatomical classes, which significantly obscures the model’s reliability on a per-class basis. In clinical radiology, performance on small or low-contrast structures (e.g., lymph nodes, adrenal glands) is far more critical and challenging than on large organs like the liver or lungs. The lack of a per-class performance breakdown or an analysis of the long-tail distribution makes it impossible to assess whether the model achieves a medically acceptable level of precision for fine-grained anatomical localization.
4.Since the source data (CT/MRI) is natively 3D, the decision to train on independent 2D slices creates a volumetric disconnect. Discuss or evaluate the spatial coherence of the model’s grounding across adjacent slices. Does the bounding box stay consistent as one “scrolls” through the generated outputs, or does it jitter?
- 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 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.
(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 paper presents an framework for scaling vision-language models in radiology. While the RefRad2D dataset is a commendable contribution to the community’s data infrastructure, the study currently operates more as a large-scale data pipeline demonstration.
My recommendation of 3 (Weak Reject) is driven by:
1.The “Semantic-Spatial Gap” and the Illusion of Performance The authors frame spatial grounding as a breakthrough for mitigating hallucinations, yet the evaluation fails to probe the epistemic consistency of the model. By relying on macro-averages across 121 classes, the lack of per-class breakdown prevents us from seeing whether the model truly understands complex anatomy or is simply exploiting the pixel-heavy prevalence of large organs.
2.The reliance on Gemma 3 as the primary judge for a PaliGemma-based model raises questions regarding the semantic neutrality of the evaluation. While the authors demonstrate a correlation with human experts, using an “intra-family” LLM judge potentially prioritizes linguistic alignment over diagnostic precision. In radiology, a report can be syntactically perfect and stylistically similar to the ground truth yet contain a critical lateralization error (e.g., mistaking left for right). Without incorporating a more heterogeneous set of evaluators or a specialized Clinical Entity-based metric, it remains unclear whether the high LLMScore reflects genuine clinical reasoning or merely a high degree of stylistic mimicry between closely related language models. 3.While the authors define G-IoU to mathematically penalize diagnostic omissions (false negatives) by normalizing with the maximum number of entities, the paper lacks a dedicated analysis of these omissions. There is no breakdown of which anatomical structures are most frequently missed, nor is there a qualitative discussion on the clinical implications of the model’s false negatives compared to its hallucinations (false positives). 4.The transition from 3D clinical volumes to independent 2D slices without addressing spatial coherence or inter-slice consistency. A model that localizes in isolation across 2D frames, without a 3D context, risks producing a jittery and clinically unusable spatial narrative when applied to real-world volumetric scans. 5.The study utilizes labels automatically generated by TotalSegmentator as the ground truth for calculating G-IoU, which introduces a significant risk of systematic bias. TotalSegmentator is itself a deep learning model with inherent limitations. Without validation against a physician-curated “Gold Standard” for at least a subset of the test data, the reported high G-IoU scores may merely reflect the model’s success in mimicking the TotalSegmentator segmentation logic rather than capturing authentic anatomical truths.
- 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.
The rebuttal addressed several of my main concerns. In particular, the authors provided additional analyses regarding omission/error patterns, semantic–spatial disagreement, and the potential bias of the LLM-based evaluator.
While I still believe the work is stronger as a dataset/system contribution than as a fundamentally novel methodological advance, I find the overall contribution meaningful and practically valuable. The scale of RefRad2D, the weakly supervised grounding pipeline, and the empirical demonstration that grounding can be incorporated without degrading report/VQA quality together represent a useful step toward more interpretable radiology VLMs.
Author Feedback
We thank the reviewers and AC for their thoughtful feedback and address their comments below.
“limited methodological novelty” (Meta)
Indeed, RadGrounder’s backbone and coordinate-token grounding follow established designs. The novelty is not a new backbone, but a clinically grounded data-and-training framework that scales spatially grounded VLM training to routine 2D CT/MR practice. Existing grounded medical VLMs are largely confined to chest X-ray (CheXagent [4], MAIRA-2 [3]), synthetic data (VividMed [14]), or medical-paper-figure pairs (LLaVA-Med [12]).
“include comparisons with more recent medical VLM and grounding methods” (Meta)
No standardized 2D CT/MR grounding benchmark currently allows apple-to-apple grounding comparison. On Slake and VQA-RAD, we compare to the highest-scoring VLMs.
“current grounding targets are primarily anatomical structures derived from TotalSegmentator rather than pathology-level findings” (R2)
We agree this is an important limitation, as we noted in our submission. Routine radiology reports describe pathology through anatomical locations (e.g. “lesion in the right hepatic lobe”), so anatomical grounding captures the dominant spatial signal in clinical text. Lesion-level grounding would require manual annotation, infeasible at 1M+ slices.
“G-IoU scores may merely reflect the model’s success in mimicking the TotalSegmentator segmentation logic” (R3)
The external Slake/VQA-RAD evaluation is independent of TotalSegmentator. However, we agree that G-IoU is not a physician-validated gold standard.
“[G-IoU] is presented only as a macro-average across 121 anatomical classes” (R3)
G-IoU is averaged per sample (like per-image mIoU). The 88 classes with test support are heavily long-tailed: top 5 cover 50% of instances; 34 have 5 or fewer instances. We also computed per-class G-IoU: model ranking is preserved but the best score drops from 45.2 to 29.3 with CI increased from 1.7 to 4.3, reflecting tail-class noise. Class frequency and G-IoU correlate mildly with Pearson r=0.34, so more frequent classes indeed perform slightly better.
“provide deeper analysis such as […], omission (false negative) analysis, and correlation between localization errors and textual hallucinations.” (Meta)
CIDEr and F1 already measure text quality independently of grounding. G-IoU penalizes both FN and FP. We will rephrase “grounding mitigates hallucinations” to “grounding adds verifiable spatial output at no cost to VQA / report quality”. Per-entity errors: 18.7% GT missed (FN), 13.1% spurious predictions (FP). Of matched boxes: 54.7% fully correct, 28.6% wrong text only, 4.6% wrong box only, 12.1% both wrong. Box and text errors correlate weakly (r=0.29).
“validating LLMScore with broader human evaluation” (Meta) “using an intra-family LLM judge potentially prioritizes linguistic alignment” (R3)
We agree that LLMScore should not be the sole evidence for clinical correctness. It is used as one complementary report/VQA quality measure and is empirically checked against three radiologists on 200 VQA-RAD pairs of our model’s output and ground truth answer, showing Pearson r=0.977 with mean human scores. If the Gemma judge primarily rewarded stylistic alignment with PaliGemma-family outputs rather than tracking clinical content, correlation with the radiologists would not be this high.
“clarifying limitations such as 2D slice-based training on 3D data” (Meta)
Our 2D slice scope is deliberate and computationally scalable. We do not claim volumetric reasoning. Inter-slice coherence is future work.
“Releasing the dataset publicly is expected” (R1)
Raw images and reports cannot be released for institutional reasons, but we will release code, pretrained models, and the annotation pipeline.
“Dataset use LLMs […] should be clearly stated as […] careful selection for any practice.” (R1)
We agree that LLM-generated annotations are research artifacts and should not be used for direct clinical deployment.
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.
The reviewers request rebuttal on the limited methodological novelty, as the architecture largely builds on existing Pix2Seq/LLaVA-style designs and grounding formulations; the authors should better position their contribution in terms of data curation, training strategy, and clinical relevance, and include comparisons with more recent medical VLM and grounding methods.
The authors should address concerns regarding the grounding design and evaluation, including the reliance on automatically generated anatomical labels, limited pathology-level grounding, potential circularity in supervision/evaluation, and provide deeper analysis such as per-class performance, omission (false negative) analysis, and correlation between localization errors and textual hallucinations.
The rebuttal should also strengthen experimental depth by discussing failure cases, validating LLMScore with broader human evaluation, and clarifying limitations such as 2D slice-based training on 3D data and its impact on spatial coherence across slices.
- 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 paper is accepted. All reviewers are positive after rebuttal and agree that the work provides a practically valuable large-scale dataset/system contribution for grounded CT/MR VLM training. Although the methodological novelty is moderate and limitations remain regarding anatomical rather than pathology-level grounding, LLM-generated annotations, and data release/reproducibility, the authors addressed key concerns with additional analyses on long-tailed classes, omission/error patterns, semantic–spatial disagreement, and LLMScore validation. Overall, RefRad2D and RadGrounder represent a useful step toward scalable and interpretable radiology VLMs.
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.
This paper presents a large-scale and carefully engineered framework for spatially grounded radiology vision-language modeling, introducing the RefRad2D dataset and the RadGrounder model for joint report generation, VQA, and anatomical grounding in CT/MR imaging. The work is motivated by an important and practical problem: enabling scalable, weakly supervised grounding in medical VLMs without requiring manual spatial annotations. The empirical scope is strong, with extensive training data, multi-task evaluation, and experiments across in-distribution and external benchmarks.
The paper’s main strength lies in its data-centric contribution and end-to-end system design at scale, demonstrating that clinically derived radiology reports can be effectively leveraged to train grounded multimodal models. The results show that incorporating grounding supervision does not degrade performance on report generation or VQA tasks, which is an important and practically relevant finding for the field. The ablation studies and benchmarking across datasets further support the robustness of the approach.
At the same time, the reviewers consistently note that the methodological novelty is moderate, as the model builds on established VLM and grounding architectures rather than introducing fundamentally new modeling ideas. The grounding supervision is also derived from automatically generated anatomical labels (e.g., TotalSegmentator outputs), which limits the degree of clinical validation and introduces some dependence on proxy labels rather than physician-annotated ground truth. Evaluation of grounding quality is therefore indirect, and stronger clinical or pathology-level validation remains an open direction.
Despite these limitations, the paper is well-motivated, clearly written, and technically solid. The rebuttal further clarified key concerns, particularly regarding grounding error analysis, omission behavior, and evaluation design choices, which strengthens confidence in the reported findings without changing the overall scope of the contribution.
Overall, I view this work as a valuable and scalable contribution to radiology VLMs, particularly in its dataset construction, weakly supervised grounding pipeline, and empirical demonstration that spatial grounding can be integrated without harming core language/VQA performance. The strengths in scale, execution, and practical relevance outweigh the limitations in novelty and evaluation depth.
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.
This paper introduces RefRad2D, a large-scale bilingual dataset of ~1.2M CT/MR slice–report pairs with automatically generated grounding annotations, and RadGrounder, a multi-task VLM that jointly performs report generation, VQA, and spatial grounding. The work is a strong data-centric contribution to making radiology VLMs more interpretable. The rebuttal successfully added missing granularity: per-class grounding performance, omission/error patterns, and human correlation for LLMScore. These additions directly answered the reviewers’ most critical requests. However, several concerns persist, as the authors acknowledged, e.g., the limitation of anatomical-level (versus pathology-level) grounding and the reliance on automatically generated labels, while noting that no standard 2D CT/MR grounding benchmark exists for direct comparison. The authors also clarify that inter-slice coherence is beyond the current scope. Despite these limitations, the paper still makes a valuable and timely contribution. Therefore, my decision is to accept.