List of Papers Browse by Subject Areas Author List
Abstract
N/A
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2223_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YuQin_SemanticAware_MICCAI2026,
author = { Yu, Qinji AND Guo, Guangyu AND Yao, Jiawen AND Zheng, Zhilin AND Lin, Tiancheng AND Xia, Yingda AND Guo, Yi AND Wang, Qifeng AND Zhou, Jian AND Zhang, Ling},
title = { { Semantic-Aware Organ-Level Esophageal Tumor Synthesis via Latent Rectified Flow } },
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
This paper proposes SOFT, a generative framework that shifts esophageal tumor synthesis from lesion-level texture inpainting to organ-level editing. Using Latent Rectified Flow conditioned on clinical attributes (T-stage, tumor location) via BioMedCLIP, SOFT simultaneously deforms the esophageal structure and synthesizes tumors. A synchronous segmentation head co-generates aligned labels to resolve the label shift caused by organ-level editing. Experiments on two large-scale datasets show improved downstream segmentation, particularly for early-stage (T1) tumors.
- 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 clinical motivation is well articulated and convincing. Figure 1 effectively demonstrates the morphological differences between normal and cancerous esophagus, and the statistical analysis of shape compactness (p < 0.01) provides quantitative evidence for why organ-level editing is necessary. The flow from clinical observation to method design is logical.
- The introduction clearly identifies the gap between solid-organ tumor synthesis and the unique challenges of hollow-organ tumors, particularly the distinction between staging by invasion depth vs. absolute tumor size. This is a valid and under-explored problem.
- The dataset scale is noteworthy, with 2,716 EC cases from Center 1, 1,028 from Center 2, plus 20,622 normal CTs as templates. External validation on Center 2 is a significant strength that demonstrates cross-center generalizability.
- The authors also validated their framework with DDPM in addition to rectified flow (SOFT vs. SOFT (LDM)), which directly supports the choice of backbone.
- The synchronous segmentation head addresses a real problem. When organ-level editing deforms structures, pre-defined masks become spatially invalid, and co-generating labels is a natural and practical solution.
- 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.Oversimplified conditioning scheme. T-stage (T1-T4) and tumor location (Upper, Middle, Lower, EGJ) yield at most 16 unique combinations. It is unclear why a full text encoder (BioMedCLIP) is needed over simpler alternatives like one-hot or learnable embeddings. This should have been ablated.
2.Masking the entire esophagus is too aggressive. All normal tissue information of the patient’s own esophagus is discarded during inference. Ironically, the introduction criticizes prior methods for only masking the tumor region, but the proposed solution goes to the opposite extreme. A more nuanced strategy preserving parts of the healthy esophagus would have been more principled, and this design choice is not sufficiently discussed.
3.No direct evaluation of generation quality or condition fidelity. All evaluations are downstream segmentation metrics. There is no verification that generated tumors actually reflect the specified T-stage and location, and no quality assessment of the re-synthesized esophagus itself. A reader study, FID, or T-stage classification on generated images would have been much more convincing.
4.No boundary-aware blending. The flow loss is computed only within the mask, and the final composite uses a binary mask (Eq. 5). There is no mechanism for smooth transitions at the boundary, which likely produces visible seams.
5.Limited evaluation diversity. Fig. 3, Fig. 4(a), and Table 1 all analyze the same downstream task from different angles. For a generation paper, this is insufficient. The number of synthetic EC cases generated is also not reported.
6.Baseline performance is suspiciously low. LeFusion and LesionDiffusion are described as state-of-the-art in Section 3.1, yet they achieve extremely poor synthetic-only results. This gap is too large to attribute solely to the lesion-level vs. organ-level distinction, and whether these baselines were fairly adapted to the esophageal setting deserves more discussion.
7.Incomplete related work. The introduction claims existing methods are limited to lesion-only texture inpainting, but some prior approaches do perform bounding-box-level inpainting. This distinction should be acknowledged.
8.Fig. 4 inconsistencies. (a) The T-stage stratification omits the synthetic-only setting, which would more directly demonstrate per-stage generation quality. (b) The inpainting region appears as a rectangular box, contradicting Section 2.4’s description of padded esophagus masks.
9.Segmentation head is limited to esophagus and tumor. Predicting masks for neighboring organs would serve as a useful sanity check that organ-level editing does not violate anatomical priors.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
- Table 1 appears to violate MICCAI formatting guidelines. The smallest allowed font size is 8pt, and the table seems to use smaller fonts to fit all the information. The raw count format (e. g. , “511/552”) for each sensitivity entry could be replaced by simply reporting the test set composition separately, freeing up space to comply with font size requirements.
- Fig. 3 shows that adding more synthetic data does not keep improving performance. This suggests limited diversity in the generated samples, which reinforces the concern about the 16-combination conditioning. More fine-grained clinical attributes could help.
- It would be informative to know the total number of synthetic EC cases generated and their T-stage distribution. If the goal is to boost T1 sensitivity, how many T1 cases were synthesized vs. other stages?
- The ablation study covers attribute conditioning and organ-level padding, but the scope could have been broader. In particular, ablating the text encoder itself (e. g. , BioMedCLIP vs. one-hot or learnable embeddings) would have been valuable given the limited number of conditioning combinations.
- 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 valid and clinically relevant problem, and the framing around organ-level editing for hollow-organ tumor synthesis is compelling. The dataset scale and external validation are clear strengths. However, the method itself is largely a combination of existing components (VQ-VAE + rectified flow + BioMedCLIP + segmentation head) without a strong standalone technical novelty. The oversimplified conditioning (16 combinations with a text encoder), the aggressive masking strategy that discards all normal esophageal tissue, and, most critically, the lack of any direct evaluation of generation quality or condition fidelity significantly weaken the paper. All experiments evaluate only downstream segmentation from different angles, which is insufficient for a generation-focused paper. While the clinical packaging is well done, the methodological and experimental gaps push this below the acceptance threshold.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper tackles an important and underexplored problem in medical image synthesis: tumor generation for hollow organs, specifically esophageal cancer (EC) in CT. The authors argue that most prior 3D tumor synthesis methods are designed for solid organs and formulate synthesis as local lesion inpainting, typically controlling disease progression through lesion size. This assumption is poorly matched to esophageal cancer, where clinical stage is determined more by depth of wall invasion than by absolute lesion size, and where tumors induce substantial organ-level morphological deformation such as irregular wall thickening and mass effect.
To address this mismatch, the paper proposes SOFT (Semantic-aware Organ-level Flow for Tumor synthesis), a unified generative framework that shifts the synthesis paradigm from lesion-level texture editing to organ-level anatomy editing. The method is built on Latent Rectified Flow, which performs efficient 3D generation in a compressed latent space obtained with a pretrained VQ-VAE. The flow model is conditioned on both anatomical context and structured clinical attributes, including T-stage and tumor location. These attributes are converted into text and encoded using BioMedCLIP, then injected into the 3D U-Net through cross-attention to enable semantically controllable synthesis.
A key technical contribution is the introduction of a synchronous segmentation head jointly trained with the flow model. Because organ-level deformation changes the spatial extent and shape of the esophagus and tumor, static pre-existing labels become invalid after synthesis. The proposed segmentation head predicts updated masks for background, esophagus, and tumor directly from intermediate decoder features, thereby producing aligned image-label pairs that reflect the synthesized anatomy. This addresses the label-shift problem caused by structural editing and makes the generated samples directly useful for downstream supervised learning.
The experimental study is extensive and clinically meaningful. The authors use a large multi-center dataset with 2,716 EC CT scans from one center for development/testing, 1,028 EC scans from an external center for generalization evaluation, and 20,622 normal chest CT scans as templates for synthesis. Compared with lesion-focused baselines such as LeFusion and LesionDiffusion, SOFT produces synthetic data that are substantially more useful for downstream segmentation. When synthetic data generated by SOFT are added to real training data, the downstream model achieves better overall performance and, notably, improves external T1 tumor sensitivity from 63.33% to 71.11% (p = 0.03). This is particularly important because early-stage tumors are the most difficult to annotate and detect, yet the most valuable clinically.
In summary, the paper’s main contribution is threefold:
1.It identifies and formulates the need for organ-level, rather than lesion-level, synthesis for esophageal tumors in hollow organs. 2.It proposes a semantic-aware latent rectified flow framework for controllable synthesis conditioned on clinical attributes. 3.It introduces a joint segmentation branch to dynamically generate labels aligned with the synthesized anatomy, enabling the production of realistic and directly usable image-label pairs for downstream training.
- 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) Strong and well-justified problem motivation
A major strength of the paper is that it clearly identifies a genuine limitation of prior tumor synthesis methods. Most existing approaches are tailored to solid organs and assume that tumor synthesis can be done through local lesion texture generation or mask-conditioned inpainting. The authors convincingly explain why this is inadequate for esophageal cancer, where disease progression involves wall invasion and organ-scale deformation rather than simply growing a larger lesion. This is a meaningful conceptual contribution because it challenges a widely used assumption in medical image synthesis.
### 2) Novel shift from lesion-level synthesis to organ-level editing
The central methodological idea—moving from localized lesion inpainting to organ-level morphological editing—is the most interesting aspect of the paper. This is particularly relevant for hollow organs, where tumors alter the geometry of the organ itself. This makes the work distinct from many existing lesion synthesis methods and opens the door to more anatomically faithful augmentation strategies in settings where local texture changes alone are insufficient.
### 3) Clinically relevant improvement on early-stage tumors
The reported gain in external T1 sensitivity from 63.33% to 71.11% is particularly compelling. Early-stage disease is the regime where data scarcity is most severe and where improved detection is most clinically valuable. The fact that this gain is statistically significant (p = 0.03) further supports the practical impact of the proposed approach.
### 4) Reasonable and informative ablation study
The paper includes useful ablations that isolate the contribution of:
- Rectified Flow vs. latent DDPM,
- attribute conditioning vs. no conditioning,
- organ-level padding vs. no padding.
These experiments support the design choices and provide evidence that both semantic control and sufficient anatomical context are important for realistic synthesis.
- 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) Baseline comparison is somewhat limited relative to the paper’s claims
The paper mainly compares against LeFusion and LesionDiffusion, both of which are lesion-focused diffusion/inpainting methods. Since the core claim is that organ-level editing is necessary, it would be stronger to compare against a broader set of related methods, especially those involving text-guided medical synthesis or image-mask pair generation. For example, MedSegFactory (Mao et al., ICCV 2025) is cited by the authors and appears relevant as a text-guided image-mask generation method, yet it is not included in the experiments. The paper also cites Chen et al. (CVPR 2024) and Wu et al. (Nature Communications 2025), but does not benchmark against adapted versions or explain in detail why such comparisons would be infeasible.
### 2) The joint segmentation idea is useful but not entirely novel in the broader literature
The synchronous segmentation head is a practical contribution, but the broader idea of using generative model features for semantic segmentation is not entirely new. The paper itself cites Baranchuk et al. (ICLR 2022) on label-efficient semantic segmentation with diffusion models. Therefore, the novelty here lies more in the adaptation to organ-level tumor synthesis and label-shift handling rather than in the general concept itself. The paper could better clarify this distinction and position the contribution more precisely.
### 3) Semantic controllability is shown mainly qualitatively
The paper claims semantic control over T-stage and tumor location, and provides qualitative examples, but lacks a strong quantitative evaluation of controllability. For example, it would be helpful to measure:
- whether independent models can recover the requested stage/location from generated images,
- whether synthesized tumor positions match the specified anatomical region,
- or whether stage-dependent morphometric patterns differ as intended. Without such analysis, the controllability claims remain only partially validated.
### 4) Generalization of the method beyond esophageal cancer is not demonstrated
The authors motivate the work as relevant for hollow organs, but experiments are limited to esophageal cancer only. As a result, it remains unclear how broadly the framework extends to other hollow-organ pathologies, such as colorectal or airway lesions. This does not invalidate the contribution, but it limits the scope of the claimed generality.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I recommend Weak Accept for this paper. The work addresses a clinically relevant and technically nontrivial problem that is not well served by existing lesion-focused synthesis approaches. The central insight—that esophageal tumor synthesis for a hollow organ should be modeled as organ-level morphological editing, rather than simple lesion inpainting—is well motivated and convincing. This conceptual shift is the paper’s main strength.
Methodologically, the proposed framework is coherent and practical. The combination of Latent Rectified Flow, semantic conditioning using clinical attributes, and a joint segmentation head for aligned label generation is a solid design. In particular, the label-shift issue caused by organ deformation is real and important, and the paper offers a useful solution that increases the direct value of the generated data for downstream supervised learning.
The experimental evaluation is another strong point. The paper uses large-scale multi-center data, includes both internal and external testing, and evaluates synthetic data through downstream task performance rather than only visual examples. The improvement in early-stage T1 sensitivity on the external cohort is clinically meaningful and directly supports the motivation of the study.
But the paper still has some limitations. The baseline comparison is somewhat narrow given the broadness of the central claim, direct realism and controllability evaluation are limited, and some implementation details are missing for full reproducibility. In addition, while the synchronous segmentation component is useful, it is not completely novel in the broader context of generative-model-based semantic prediction.
Overall, however, I believe the paper makes a meaningful contribution: it identifies an important problem, proposes a sensible and reasonably novel solution, and provides strong downstream evidence of practical value, especially for early-stage disease. These strengths place it slightly above the acceptance threshold.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper tackles an important problem generating realistic esophageal tumors with organ-level deformations, which is a clear shift from existing lesion-inpainting methods for solid organs. The proposed SOFT framework, combining latent rectified flow with a synchronous segmentation head, is technically sound and the reported improvements (especially for early-stage T1 tumors) are promising.
- 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.
This paper tackles an important problem generating realistic esophageal tumors with organ-level deformations, which is a clear shift from existing lesion-inpainting methods for solid organs. The proposed SOFT framework, combining latent rectified flow with a synchronous segmentation head, is technically sound and the reported improvements (especially for early-stage T1 tumors) are promising.
- 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 paper focuses heavily on downstream segmentation performance, but provides no direct assessment of the realism of synthesized CT volumes or the accuracy of the generated segmentation masks. Metrics such as FID, KID, or user study by radiologists are missing. Without this, it is unclear whether the downstream gains come from realistic augmentations or from unintended artifacts. Please add quantitative image quality metrics and, if possible, a small reader study comparing real vs. synthetic cases.
2.Only two baselines (LeFusion and LesionDiffusion) are compared. Both are lesion-inpainting methods not designed for organ deformation. The paper does not compare against other conditional generative models that could be adapted, e.g., latent diffusion models with mask conditioning or recent organ-aware GANs. Also, the ablation “SOFT (LDM)” replaces rectified flow with DDPM but keeps the same conditioning and segmentation head – this is useful, but why not compare against a standard LDM baseline without the segmentation head? Please add at least one strong generative baseline that is not purely lesion-focused (e.g., a vanilla LDM with mask inpainting and attribute conditioning).
3.The generative model is trained on Center 1 data (80% of EC cases) and synthetic cases are generated from normal CT templates from a third center. Downstream segmentation models are then trained on real Center 1 data, synthetic data, or both, and tested on the remaining 20% of Center 1 plus external Center 2.This design is reasonable, but it is not clear whether the same normal templates used for synthesis might appear in the downstream training or testing splits. Also, the authors do not report whether the synthetic data causes overfitting when added to small real datasets (e.g., 10% real data in Fig. 3). Please clarify how normal templates are selected and whether any template overlaps with test cases. Also, report variance across multiple runs for the data efficiency experiments (error bars).
4.Figure 4(b) shows visual examples of T1 vs. T4, but there is no systematic evaluation of whether the model correctly interprets text attributes (e.g., “upper” vs. “middle” esophagus) or avoids spurious correlations. The paper claims “fine-grained semantic manipulation,” but no quantitative control study (e.g., swapping attributes for the same template) is provided. Please perform a controlled attribute manipulation experiment, e.g., generate from the same normal template with different T-stages and locations, and measure whether the resulting tumor depth and position match the intended attributes (using the segmentation head or manual review).
5.The paper introduces a synchronous segmentation head to handle label shift caused by organ deformation. However, it does not compare against a simple alternative: generating images with the flow model and then applying a separate, pre-trained segmentation network to obtain labels. If that simpler method works well, the claimed necessity of co-training is weakened. Please add a baseline where labels are obtained by a separate segmentation model (trained on real data) on the synthetic images, and compare the quality of those labels (e.g., Dice against ground-truth synthetic labels) with your joint head.
- 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.
(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?
This paper tackles an important problem generating realistic esophageal tumors with organ-level deformations, which is a clear shift from existing lesion-inpainting methods for solid organs. The proposed SOFT framework, combining latent rectified flow with a synchronous segmentation head, is technically sound and the reported improvements (especially for early-stage T1 tumors) are promising.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank R2, R3, and R5 for their constructive feedback. All reviewers recognize the clinical relevance of hollow-organ tumor synthesis, the shift from lesion- to organ-level editing, the large-scale multi-center evaluation (2,716 + 1,028 EC + 20,622 normal CTs), and the gain in external T1 sensitivity (63.33% to 71.11%, p=0.03).
[Response to R2]
(W1) Conditioning. We directly follow the LesionDiffusion setting and adopt BioMedCLIP for fair comparison with prior lesion-synthesis works. One-hot/learnable embedding alternatives will be explored in our journal extension.
(W2) “Masking the entire esophagus.” We tried preserving more healthy esophagus in early experiments, but this artificially fixed the tumor location and removed the controllability we want from text conditioning. A systematic comparison will be in the journal version.
(W3) Direct generation quality / condition fidelity (shared with R3-W3, R5-W1/W4). We agree this is the most important gap. FID/KID, an independent T-stage/location classifier on synthetic cases, and a controlled attribute-swap study (same template, varied T-stage/location) will be reported in the journal extension.
(W4) Boundary blending. The network receives the surrounding background as context, which already smooths the transition; in our visualizations no obvious seam is observed, so explicit blending is not needed.
(W5/W8a) Synthetic count and diversity. All 20,622 normal chest CTs were used as templates, with synthetic EC uniformly generated across T1-T4 and four locations. A synthetic-only column in Fig. 4(a) will be added in the journal version.
(W6) Baseline fairness. LeFusion and LesionDiffusion were reproduced from their official code on our esophageal dataset. The main failure mode is that their first-stage shape generator does not produce plausible esophageal tumor shapes; the resulting masks deviate substantially from real EC morphology, so the subsequent texture inpainting cannot recover anatomically faithful tumors. This supports, rather than contradicts, our motivation.
(W7/W9) Related work and segmentation head. We will acknowledge prior bounding-box-level inpainting; extending the segmentation head to neighboring organs is left as future work, as we focus on the directly edited esophagus and tumor.
(W8b) “Rectangular box” in Fig. 4.The box is correct. Our “padding” differs from morphological dilation: for each axial slice we compute the per-slice bounding box of the esophagus mask and stack them into a 3D box mask. Dilation would leak the original esophagus shape and contradict organ-level editing; a single 3D bounding box would be excessively large. We will clarify this in the caption.
[Response to R3]
(W1) Broader baselines. We will discuss MedSegFactory (Mao et al., ICCV 2025), Chen et al. (CVPR 2024), and Wu et al. (Nat. Commun. 2025) more explicitly, including feasibility of adaptation to hollow-organ editing.
(W2) Joint segmentation novelty. We agree the broader idea traces to Baranchuk et al. (ICLR 2022); our contribution is the formulation and resolution of the label-shift problem caused by organ-level deformation. We will state this more precisely.
(W4) Generalization beyond EC. Out of scope here; extension to colorectal/airway lesions is framed as future work.
[Response to R5]
(W3) Template overlap. Normal CT templates come from a separate cohort with no patient overlap with EC train/test splits. We will state this explicitly and add error bars in Fig. 3. (W5) Pre-trained segmenter alternative. A post-hoc segmenter cannot resolve label shift from organ-level deformation, since the deformed anatomy differs from the prior label distribution it was trained on. A comparison applying a real-data-trained nnU-Net to synthetic images will be added.
We thank the reviewers again for helping us strengthen the paper.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper addresses a clinically important and methodologically novel problem, namely shifting from lesion-level inpainting to organ-level editing for hollow-organ tumor synthesis. The dataset scale, external validation, and downstream improvement on early-stage T1 tumors are all strong points. One reviewer recommends Accept (5), one Weak Accept (4), and one Weak Reject (3). The criticisms raised, including lack of direct generation quality metrics (FID, user study), limited baselines (no comparison with UDA or histogram matching), insufficient ablation of the conditioning scheme (BioMedCLIP vs. one-hot), and an overly aggressive masking strategy, are all reasonable and addressable during rebuttal. Given the overall strengths and the clear path to address the concerns, the paper should proceed to early acceptance if the authors adequately respond to the reviewers’ concerns.