List of Papers Browse by Subject Areas Author List
Abstract
Text-conditioned generative models for volumetric medical imaging provide semantic control but lack explicit anatomical guidance, often resulting in outputs that are spatially ambiguous or anatomically inconsistent. In contrast, structure-driven methods ensure strong anatomical consistency but typically assume access to ground-truth annotations, which are unavailable when the target image is to be synthesized. We propose a retrieval-augmented approach for Text-to-CT generation that integrates semantic and anatomical information under a realistic inference setting. Given a radiology report, our method retrieves a semantically related clinical case using a 3D vision-language encoder and leverages its associated anatomical annotation as a structural proxy. This proxy is injected into a text-conditioned latent diffusion model via a ControlNet branch, providing coarse anatomical guidance while maintaining semantic flexibility. Experiments on the CT-RATE dataset show that retrieval-augmented generation improves image fidelity and clinical consistency compared to text-only baselines, while additionally enabling explicit spatial controllability, a capability inherently absent in such approaches. Further analysis highlights the importance of retrieval quality, with semantically aligned proxies yielding consistent gains across all evaluation axes. This work introduces a principled and scalable mechanism to bridge semantic conditioning and anatomical plausibility in volumetric medical image synthesis. Code is available at https://github.com/arco-group/RAGText2CT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1917_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/arco-group/RAGText2CT
Link to the Dataset(s)
N/A
BibTex
@InProceedings{MolDan_RetrievalAugmented_MICCAI2026,
author = { Molino, Daniele AND Caruso, Camillo Maria AND Soda, Paolo AND Guarrasi, Valerio},
title = { { Retrieval-Augmented Anatomical Guidance for Text-to-CT Generation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper tackles the problem of synthesising three‑dimensional CT volumes from free‑form radiology reports while ensuring anatomical plausibility. Purely text‑conditioned models capture semantics but produce anatomically inconsistent volumes because reports lack spatial information, whereas mask‑conditioned methods require segmentation labels at inference and are therefore impractical.
The authors propose a retrieval‑augmented framework: a three‑dimensional vision‑language encoder retrieves the most semantically similar case from a training corpus, uses the retrieved case’s segmentation mask as a structural proxy, and injects this proxy into a frozen report‑conditioned latent diffusion model via a ControlNet branch. This design retains the flexibility of text conditioning, approximates anatomical structure without requiring ground‑truth masks at inference, and preserves the pretrained generative backbone.
The contributions include (1) framing report‑conditioned CT generation as a retrieval‑augmented problem where anatomy is treated as a latent proxy, (2) integrating this proxy with a latent diffusion backbone through a separate ControlNet branch while keeping the backbone frozen, and (3) empirically evaluating the approach on the CT‑RATE dataset against text‑only baselines, mask‑conditioned upper bounds and retrieval‑quality ablations.
The work therefore represents an engineering combination of retrieval‑augmented generation and existing diffusion architectures.
- 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 method addresses a clinically relevant gap: synthesising CT volumes from reports without relying on segmentation masks at inference. Retrieving a semantically similar case and using its mask approximates anatomical structure while preserving the flexibility of text conditioning.
2.The approach leverages a pretrained latent diffusion model and adds a ControlNet branch to inject the proxy mask through zero‑initialised residual connections, so only the control branch is trained. This preserves the existing generative capabilities, simplifies optimisation and makes the system modular.
- Experiments report image fidelity using two‑dimensional and three‑dimensional FID, clinical consistency via a pretrained CT‑Net classifier, and spatial controllability via Dice and HD95 after segmenting generated volumes. Baselines include multi‑stage pipelines (GenerateCT, MedSyn), diffusion‑based report models (Text‑to‑CT, Report2CT), a mask‑conditioned upper bound (MAISI) and ablations on retrieval quality. An ablation comparing nearest, farthest and random retrieval demonstrates that semantically aligned proxies improve FID, classification AUC and Dice scores, whereas poor retrieval degrades performance. Statistical tests (paired Wilcoxon) are reported.
4.The paper clearly formulates the problem, retrieval mechanism and ControlNet integration with equations and describes training and inference procedures. Dataset splits, preprocessing and hyper‑parameters are provided, and code release upon acceptance is promised.
- 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 proposed method uses known components: a three‑dimensional vision‑language encoder, nearest‑neighbour retrieval and ControlNet conditioning and applies them to CT generation. MAISI already uses ControlNet with segmentation masks; here the mask is replaced by a retrieved proxy. The contribution is therefore incremental and primarily combines existing tools rather than introducing a fundamentally new generative mechanism.
2.The method relies on retrieving a single nearest neighbour based on report embeddings. There is no quantitative analysis of how often the retrieved mask matches the target anatomy. Failure modes where the nearest neighbour is anatomically dissimilar for example, different pathology or organ size are not examined. Using multiple neighbours or alternative retrieval strategies is not explored.
- Experiments are conducted only on the CT‑RATE dataset using automated metrics. Clinical consistency is measured via CT‑Net classification on real CTs, which tests recognisability rather than whether the generated image reflects the described pathology. Moreover, spatial controllability compares generated masks to the retrieved proxy rather than ground‑truth anatomy, thus high Dice scores may reflect copying the proxy rather than generating correct anatomy. FID is known to correlate poorly with clinical plausibility. There is no external validation by radiologists, no pathology‑specific breakdown, and no tests on other body regions or report styles.
4.The paper compares to several baselines but does not discuss or evaluate alternative controllable text‑to‑CT strategies, such as models guided by pseudo‑masks, landmarks or multiple proxies. Retrieval‑based generative models outside medical imaging are not cited. A comparison against a simpler structural prior without retrieval would help isolate the contribution of the retrieval step.
- 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?
The paper addresses a real limitation of report-conditioned 3D CT generation by introducing a practical mechanism for adding anatomical guidance without requiring target masks at inference time. The central idea: retrieving a semantically related case and using its annotation as a structural proxy is reasonable and better aligned with realistic deployment than methods that assume direct access to anatomical labels at generation time. The experimental results also support the claim that structural guidance improves over text-only conditioning. At the same time, the contribution is incremental rather than fundamental, since it mainly modifies how structural conditioning is obtained within an already active line of work. In addition, the validation would be more convincing with stronger evidence on generalization, failure modes, and anatomy-faithfulness beyond adherence to the retrieved proxy.
Overall, I find the paper worthy of acceptance, but with a moderate confidence level and mainly on the basis of a useful methodological refinement rather than a major conceptual advance.
- 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.
While the rebuttal effectively clarifies the method’s implementation and design choices, there remain opportunities for strengthening the validation narrative in future work. On retrieval validation, the authors cite strong MAP metrics from prior work on their encoder, which is reassuring, though providing direct quantitative evidence of retrieval success on their own generated test cases would further support the core assumption that semantic similarity reliably yields anatomically coherent proxies. The ablation comparing nearest, random, and farthest retrieval is valuable empirical evidence, and the authors might consider extending this with documented examples of cases where semantic misalignment occurred, to help readers understand the method’s failure modes and boundaries. The evaluation design, comparing Dice scores to the retrieved proxy rather than ground truth is sensible given that reports underspecify anatomy, and the authors make a fair point that high AUC alongside high Dice suggests the method achieves both structural coherence and clinical plausibility. That said, future work might explore whether validation on out-of-distribution cases or cross-dataset experiments would demonstrate robust generalization beyond CT-RATE. On novelty, the authors reasonably position their contribution as addressing a systematic limitation of text-to-CT methods through retrieval-augmented conditioning; while the individual components are established, the framing as solving the spatial underspecification problem in reports is a valid contribution. A radiologist reader study remains an important next step to validate clinical utility, which the authors acknowledge. Overall, the work makes a solid methodological contribution with clear practical motivation, and these directions would strengthen its impact.
Review #2
- Please describe the contribution of the paper
This paper proposes a retrieval-augmented framework for report-conditioned 3D CT generation, where anatomical structure is approximated via a retrieved proxy rather than requiring ground-truth annotations at inference time. Specifically, given a radiology report, a semantically similar case is retrieved using a 3D vision-language encoder, and its associated segmentation mask is used as a structural prior injected into a latent diffusion model via ControlNet. The method aims to bridge semantic conditioning and anatomical consistency under a realistic inference setting.
- 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) Well-motivated problem: addresses the gap between semantic-only and structure-driven generation. 2) Interesting use of retrieval to approximate anatomical priors without requiring annotations at inference. 3) Technically sound integration with latent diffusion + ControlNet. 4) Multi-axis evaluation (FID, CT-Net, Dice/HD95). 5) Ablation on retrieval strategies (nearest/random/farthest) is 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.
1) Lack of validation of the core retrieval assumption: The method critically depends on retrieving a “semantically related clinical case using a 3D vision-language encoder” (Abstract, p.1), yet there is no rigorous quantitative or qualitative validation that the retrieved cases are indeed semantically or anatomically aligned with the input reports. The paper does not provide retrieval metrics (e.g., recall@k, similarity distributions) or qualitative retrieval examples beyond limited illustration (Fig. 2, p.8).
2) Unclear training and domain alignment of the vision-language encoder: On page 3, the authors state that “text conditioning is implemented through embeddings extracted from a vision-language model based on the CLIP paradigm [24], extended with a 3D image encoder.” However, it is unclear whether this encoder is sufficiently pretrained or fine-tuned on radiology-specific data(in this case CT). Without strong modality-specific alignment, the reliability of semantic retrieval from radiology reports is questionable.
3) No radiologist or expert evaluation: Despite operating in a clinically sensitive domain, the paper lacks any expert (radiologist) validation. This is particularly important given that the method assumes that retrieved reports and their associated anatomy are clinically meaningful proxies.
4) Assumption that radiology reports fully represent image content is not validated: The approach assumes that the input report provides a sufficiently complete and accurate semantic description of the underlying CT volume. However, radiology reports are often incomplete, variable, and focused on salient findings rather than full anatomical description. The paper does not analyze how this limitation affects retrieval or generation.
5) Evaluation of spatial controllability is proxy-based: Dice and HD95 are computed against retrieved masks rather than ground truth (p.7), which may overestimate anatomical correctness and does not confirm true alignment with the intended anatomy.
6) Moderate methodological novelty: The framework is primarily a combination of existing components (retrieval + latent diffusion + ControlNet), with limited fundamentally new modeling contributions.
- 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
The proposed idea of using retrieval to approximate anatomical priors is interesting and potentially impactful. However, the paper would significantly benefit from:
1) Explicit evaluation of retrieval quality (e.g., retrieval accuracy, qualitative case studies across multiple examples)
2) Clarification of the training and domain adaptation of the 3D vision-language encoder
3) Expert (radiologist) validation of generated outputs and retrieved cases
4) Analysis of failure modes when retrieval is semantically incorrect or anatomically mismatched
Addressing these points would strengthen confidence in the central claim that retrieval provides meaningful anatomical guidance.
- 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?
While the paper addresses an important problem and proposes a conceptually interesting retrieval-augmented framework for text-to-CT generation, the central assumption of the method—that semantically meaningful anatomical proxies can be reliably retrieved from a vision-language embedding space—is insufficiently validated. The paper claims that “a semantically related clinical case is retrieved using a 3D vision-language encoder” (p.1), yet provides no rigorous quantitative or qualitative evidence supporting the quality or reliability of this retrieval process.
Furthermore, it is unclear whether the underlying vision-language model is adequately pretrained or adapted to the radiology domain (p.3), raising concerns about the semantic alignment between reports and CT volumes. The absence of radiologist evaluation further limits confidence in the clinical validity of both retrieved proxies and generated images.
While the experimental results show improvements across several metrics, these gains may be partially driven by proxy-based evaluation and do not fully establish anatomical correctness or clinical realism. Given that the core contribution relies heavily on retrieval quality, the lack of thorough validation in this component is a significant limitation.
Overall, the paper presents a promising direction but requires stronger validation and analysis to support its central claims.
- 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.
Thank you to the authors for the detailed rebuttal. I appreciate the additional clarification regarding the retrieval mechanism, particularly the retrieval ablation analysis (Nearest vs Random vs Farthest) and the inclusion of retrieval statistics from the underlying encoder framework. These additions strengthen confidence that the retrieved proxies contain meaningful phenotype-level information rather than arbitrary anatomical guidance.
I also appreciate the clarification that the goal is not exact anatomical reconstruction, but structurally coherent conditioning aligned with report semantics. While concerns regarding the absence of radiologist evaluation and the indirect nature of some evaluation components remain, the rebuttal substantially improves the technical justification of the central retrieval assumption. I therefore revise my assessment toward a borderline positive recommendation.
Review #3
- Please describe the contribution of the paper
The paper proposes augmenting a text-conditioned 3D CT synthesis framework with a retrieval-based anatomical proxy. Given a radiology report at inference time, the method retrieves the most semantically similar case from the training corpus using a shared 3D vision-language embedding space, and uses its segmentation mask as a structural prior. This proxy is injected into a frozen latent diffusion model via a ControlNet branch. The main claim is that this enables anatomically coherent generation without requiring any annotations at inference time, bridging the gap between purely text-conditioned generation and structure-conditioned methods.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The clinical motivation is strong. Text-only CT synthesis often produces spatially inconsistent outputs, and requiring a ground-truth segmentation mask at inference is unrealistic in most deployment settings. Framing anatomy as a “retrievable proxy” is a natural and promising solution to this tension.
2.The retrieval quality ablation is clean and directly supports the paper’s central message. It cleanly isolates the contribution of semantic alignment from the structural conditioning mechanism itself.
- 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 framework is quite incremental over Text-to-CT [22]. The proposed method essentially adds a ControlNet branch to that prior model, with the mask sourced from retrieval rather than ground truth. This is fine, but the framing overstates the architectural novelty.
2.The paper positions MAISI as an upper bound that “assumes access to ground-truth annotations at inference time,” but this is not fully accurate. MAISI itself supports inference with TotalSegmentator-predicted masks rather than ground-truth masks, the same tool used here to generate training proxies.
- Comparing FID between systems with different training distributions is problematic and is not acknowledged anywhere. MAISI was trained on ~10K multi-site CT volumes from diverse datasets; this system and all text-conditioned baselines train on CT-RATE (~27K chest CTs).
- 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.
(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 tackles a real problem and the three-axis evaluation is well-designed. The method is however incremental over Text-to-CT [22]. The MAISI comparison is also unfair, as MAISI already supports TotalSegmentator-predicted masks at inference, the same tool used here, making the annotation-free claim overstated.
- 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 authors’ rebuttal and the discussion, I would like to update my recommendation from Weak Reject (3) to Weak Accept (4). The rebuttal addressed my main concerns about the MAISI comparison and the FID evaluation. In particular, the clarification that MAISI uses the target case’s own mask as an oracle condition, while the proposed method retrieves a proxy from the report, makes the comparison clearer. The authors also clarified that all methods were retrained on CT-RATE, which helps address my concern about the FID comparison. I still think the method is somewhat incremental, and the lack of radiologist evaluation and more detailed failure analysis remains a limitation. However, after the rebuttal, I think the paper is reasonably strong for a marginal acceptance.
Author Feedback
We thank all reviewers and the AC for the constructive feedback. [R1,R2,R3] Retrieval Validation: We clarify that our ablation was designed to validate the core retrieval assumption. Comparing RAG-Nearest, Random, and Farthest under identical generative conditions directly isolates retrieval quality: the monotonic, statistically significant degradation from Nearest→Random→Farthest across all evaluation axes (paired Wilcoxon p<0.05) is direct empirical evidence that semantic similarity correlates with anatomically and clinically meaningful proxies. Regarding explicit retrieval metrics: our system uses the 3D encoder of [22], the following numbers are from that paper, not new results from our experiments. [22] evaluates the encoder retrieval on CT-RATE achieving Re@5=0.041, Re@10=0.072, Re@50=0.229, Re@100=0.368 and MAP@1=0.987, MAP@5=0.848.MAP@K is the most informative metric for our task: the proxy does not need to replicate the target anatomy, but to provide a structurally coherent scaffold aligned with the report pathological content. MAP@1 confirms the nearest neighbor shares the same pathological class in nearly all cases. [R1,R2,R3] Novelty: No prior work addresses annotation-free anatomical conditioning for text-to-CT generation. We agree the individual components are not new in isolation, and that architecturally we build on [22], a deliberate choice, as it ranks as SOTA on the VLM3D Challenge [12], the MICCAI benchmark for this task. Our core contribution is identifying and addressing a systematic limitation of text-to-CT methods: radiology reports are inherently underspecified with respect to spatial structure, and this limitation propagates directly into anatomically inconsistent outputs, with consequences in real deployment scenarios (Morrison et al., “A Human-centered approach to identifying promises, risks & challenges of text-to-image generative AI in radiology”, AIES 2025). We propose retrieval-augmented anatomical proxy conditioning to counter this flaw, achieving statistically significant improvements across all three evaluation axes. [R1,R2] Spatial Controllability: Dice and HD95 are intentionally computed against the retrieved proxy, measuring the model’s ability to follow the provided scaffold, which is the relevant capability in our setting. Evaluating against the GT would be misleading, as the report does not describe individual anatomy. In addition, a model trivially copying the proxy would score high here but degrade on AUC (Table 2); RAG-Nearest achieves both the highest Dice and highest AUC, confirming the two axes are complementary and that structural guidance does not come at the cost of semantic quality. [R2] Encoder Domain Alignment: We apologise for the ambiguity. The encoder used for retrieval is pre-trained on CT-RATE, not a generic CLIP encoder. This will be stated explicitly in the final version. [R3] MAISI Comparison: R3 correctly notes that MAISI supports predicted masks at inference, however the key distinction is not the mask source, but whose mask is used: MAISI conditions on the target volume’s own mask, giving it direct structural access to the case being synthesized, making it an oracle upper bound. Our method requires no mask at inference; the proxy is obtained purely via retrieval from the input report. MAISI has no mechanism to obtain a structural proxy from a report alone, nor can it leverage the report for conditioning. MAISI’s oracle status will be stated more explicitly in the final version. [R3] FID comparability: All methods were retrained on CT-RATE following official repositories, ensuring uniform training distribution, shape, and preprocessing. This was necessary since FID is a distributional metric sensitive to data shape and CT-Net would otherwise operate on heterogeneous inputs. This will be further acknowledged in the final version. [R1,R2] Radiologist Evaluation: We acknowledge the absence of a reader study as a limitation and consider it a priority for future work
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 is borderline. One weak accept and two weak rejects, but all three reviewers agree the problem is important and that the retrieval-based anatomical proxy is a sensible idea.
The main issue is that the core retrieval assumption is not validated strongly enough. Multiple reviewers point out that the whole method depends on retrieving anatomically and semantically meaningful neighbours, yet the paper does not provide enough direct evidence that this actually happens. This is the most important point for rebuttal.
There are also concerns that the contribution is more incremental than the paper claims, essentially adding retrieval-based proxy conditioning to an existing text-to-CT and ControlNet setup. In addition, Reviewer 3 raises a specific concern that the positioning against MAISI may be overstated, since MAISI can also operate with predicted masks rather than true masks at inference.
The evaluation is generally solid, but several parts remain indirect. Spatial controllability is measured against the retrieved proxy rather than true anatomy, and there is no expert validation or deeper analysis of retrieval failures and pathology-specific behaviour.
The authors should focus on validating retrieval quality more directly, clarifying the training and domain alignment of the vision-language encoder, moderating or justifying the novelty claims, and addressing the fairness of the MAISI comparison. If these points are answered convincingly, the paper could still move into the accept range.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The central claim depends on the retrieved proxy being anatomically and semantically meaningful for the input report, but the submitted paper does not provide sufficient direct validation of retrieval quality. The rebuttal clarifies that some retrieval metrics come from prior work rather than the present experiments. The nearest, random, and farthest ablation is useful, but it remains indirect evidence: it shows that better embedding similarity improves generation metrics, not that the retrieved masks are anatomically appropriate for the intended target case.
The method is incremental over an existing text-to-CT model and ControlNet-style conditioning. Spatial controllability is evaluated against the retrieved proxy rather than true anatomy, so it primarily measures adherence to the proxy, not anatomical correctness. There is no expert validation, no pathology-specific failure analysis, and the MAISI comparison still requires careful qualification as an oracle or upper-bound comparison.
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.
The rebuttal successfully addresses the core concerns: the retrieval ablation and encoder clarification confirm that semantic retrieval yields anatomically meaningful proxies, satisfying Reviewers #1 and #2.Reviewer #3’s novelty and MAISI comparison concerns are acknowledged but do not undermine the practical contribution of annotation-free anatomical conditioning. The paper is a solid, well-motivated methodological refinement with clear clinical relevance, though future work should include radiologist validation and direct retrieval metrics.
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 was reviewed by three experts in the field. The initial recommendations were one Weak Accept and two Weak Rejects. After the rebuttal, reviewers updated their recommendations to Accept. Based on the authors’ rebuttal and the reviewers’ post-rebuttal feedback, I believe the authors have adequately addressed the key concerns, including retrieval validation, novelty, and comparison with MAISI. Therefore, the paper is recommended for acceptance to MICCAI 2026.The authors are encouraged to incorporate the necessary changes discussed during the rebuttal process into the final camera-ready version.
